Quality and Inspection of Machining Operations: Tool · monitoring, and the detection and...

John T. RothPenn State Erie,Erie, PA 16563

Dragan DjurdjanovicUniversity of Texas,

Austin, TX 78712

Xiaoping YangCummins Inc.,

Columbus, IN 47202

Laine Mears

Thomas Kurfess

Clemson University,Clemson, SC 29634

Quality and Inspection ofMachining Operations: ToolCondition MonitoringTool condition monitoring (TCM) is an important aspect of condition based maintenance(CBM) in all manufacturing processes. Recent work on TCM has generated significantsuccesses for a variety of cutting operations. In particular, lower cost and on-boardsensors in conjunction with enhanced signal processing capabilities and improved net-working has permitted significant enhancements to TCM capabilities. This paper presentsan overview of TCM for drilling, turning, milling, and grinding. The focus of this paperis on the hardware and algorithms that have demonstrated success in TCM for theseprocesses. While a variety of initial successes are reported, significantly more research ispossible to extend the capabilities of TCM for the reported cutting processes as well asfor many other manufacturing processes. Furthermore, no single unifying approach hasbeen identified for TCM. Such an approach will enable the rapid expansion of TCM intoother processes and a tighter integration of TCM into CBM for a wide variety of manu-facturing processes and production systems. !DOI: 10.1115/1.4002022"

1 IntroductionGlobal demands for improved quality, reduced downtime,

lower production costs, and overall improved systems and produc-tion line control are driving the need for improved productioncapabilities and higher performance processes. Condition basedmaintenance #CBM$ is a critical element enabling continuous im-provement of any modern manufacturing facility. CBM can beseen as an integral process of the seamless transformation of rawdata related to equipment health and performance into informationabout process and system health that is essential in decision mak-ing regarding production operations !1,2",

The standard for open-systems architecture for CBM #OSA-CBM$ defines the various stages of this data transformation as isshown in Fig. 1 !2". Information about health of any piece ofequipment is obtained from the readings of one or more sensorsmounted on that equipment. Often, situations exist where sensorreadings are augmented with historical knowledge pertaining toequipment behavior, engineering model of phenomena occurringin the equipment, and human expertise. Based on these sources ofinformation, features relevant to equipment health are extractedfrom sensor readings through various forms of sensory signal pro-cessing and feature extraction. These features form behavior mod-els of equipment in different health states #normal behavior anddifferent faulty behavior modes$. Those models may be in variousdifferent forms, including a statistical form #distributions of sen-sory signatures under normal or various faulty conditions$, dy-namic model #differential equations describing various healthstates of the equipment$, and others. Based on the models of nor-mal and current equipment behavior, equipment health assessmentcan be accomplished by quantitatively expressing the proximity ofthe currently observed system behavior to the model describing itsnormal health state. Similarly, the presence or absence of any faultcan be diagnosed through proximity of the model of the currentlyobserved equipment behavior to the behavior model correspond-ing to a specific fault. Finally, the temporal dynamics of signaturesextracted from sensor readings can be captured and extrapolatedto predict their behavior in the future and thus predict likelihoods

of various behavior modes for the equipment. Figure 2 illustratesthe concepts of quantitative health assessment and diagnosis inCBM based on simple statistical models of various behavioralmodes while Fig. 3 illustrates the concept performance predictionin CBM.

Based on the quantitative information about current and/or pre-dicted equipment health, maintenance and operational decisionsthat are optimal from the system level point of view can be made.In a manufacturing system that entails maintenance and/or pro-duction decisions that typically deliver some combination ofmaximum productivity, target quality levels, and minimum costs,given the equipment condition, work-in-progress states, costs ofmaintenance operations, availability of maintenance resources,and other criteria !3,4". In practice this target decision point isdefined by maximum profit or return on investment. Various as-pects of this “data to information to decision” transformation havereceived significant attention, especially in the case of sophisti-cated, expensive and safety critical systems, such as manufactur-ing equipment, computer networks, automotive and aircraft en-gines, etc. A thorough survey of latest activities and achievementsin CBM can be found in Ref. !1".

This paper is a review paper and presents the latest researchachievements in CBM applications for tool condition monitoring#TCM$ for cutting processes in manufacturing. TCM is a key el-ement of CBM for cutting processes and is in its infancy regard-ing it application to these processes. Thus, the advancement andutilization of TCM presents research and implementation chal-lenges. Next generation technologies such as wireless networksystems, integrated low-cost sensors that can be embedded di-rectly into tooling, and more powerful system controllers capableof performing significant analyses on streaming data to estimatetool conditions are facilitating more advanced application of TCMin the modern production units. However, as with all advance-ments there are some hurdles that must be overcome when imple-menting TCM. In particular, there are two aspects of TCM thatmake it difficult to implement. First is that a wide variety ofcutting operations are employed #e.g., drilling, turning, grinding,etc.$. Second, to be most effective the measurement of the toolcondition must be done in situ where extremely harsh and varyingconditions may exist. Such monitoring is typically done usingindirect measurements and estimation of various tool parameterssuch as sharpness, material and chip parameters and contact inter-face conditions.

Contributed by the Manufacturing Engineering Division of ASME for publicationin the JOURNAL OF MANUFACTURING SCIENCE AND ENGINEERING. Manuscript receivedJanuary 6, 2009; final manuscript received June 3, 2010; published online August 3,2010. Assoc. Editor: Suhas Joshi.

Journal of Manufacturing Science and Engineering AUGUST 2010, Vol. 132 / 041015-1Copyright © 2010 by ASME

Downloaded 04 Aug 2010 to 146.6.84.35. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm

The use of CBM for general machining process monitoring wasdiscussed in 2004 in Ref. !5" with sections dedicated to monitor-ing of surface texture and integrity, dimensional accuracy, toolcondition and chatter. A more focused review can be found in Ref.!6", where research achievements in CBM applications in TCMalone are surveyed. This area was further surveyed by Rehorn etal. !7", who examined TCM from a process-specific perspective.Nevertheless, due to the rapid pace of advances in new sensingand computing technologies, and increasingly stringent require-ments for improved quality and reduced equipment downtimes,one can observe a surge of recent activity in the area of TCM

research. Furthermore, new hardware and algorithmic develop-ments render some of the old approaches based on limited sensingand computational capabilities obsolete. Therefore, a thoroughsurvey focusing on the latest developments in the TCM is needed.

With the goal of presenting a focused, state-of-the-art review ofthe most significant research in TCM published in the last 5 years,this paper discusses TCM for the turning, drilling, milling, andgrinding applications. Conceptually, this paper focuses on indirectsensing methods applied to the signal processing and feature ex-traction, and health assessment and health prediction layers ofCBM. The reason for this is that even though the most accuratemeasurements of the tool condition can be done in situ, high costs,intrusiveness on the normal manufacturing process, as well asharsh and unpredictable conditions under which direct tool condi-tion sensors most operate often render these direct approaches toTCM unfeasible. Indirect measurements and estimation of varioustool parameters such as sharpness, contact interface conditionsand material and chip parameters are hence often more amenableto applications in full scale manufacturing. Some discussion isprovided on advances in data presentation and decision-making#when to change tool and when not$; however, these topics are notthe focus of this paper. The purpose of the ensuing text is toprovide a clear snapshot of current capabilities, and serve as afoundation for further studies into TCM. This paper is organizedby each of the four manufacturing processes reviewed. For con-sistency, within each process section three areas are discussed:sensing and hardware, signal processing and feature extractionmethods, and tool condition health assessment methods.

2 Tool Condition Monitoring for TurningIndirect sensing-based tool condition monitoring seems to have

obtained the most attention in turning operations. The reason forthis is that it is less complex than other processes with definedcutting edges #drilling and milling$ because only one cutting edgeis engaged with the material, and the depth of cut is usually con-stant #at least in the case of machining of cylindrical features$.

A good sense of the immense amount of earlier research in thearea of indirect sensing-based tool condition monitoring in turningcan be obtained from Ref. !8", where a comprehensive review of138 publications dealing with the use of artificial neural networks#ANNs$ for on-line and indirect tool condition monitoring in turn-ing is given. The generic model adapted by the authors for char-acterization and comparison of approaches is given in Fig. 4.

This paper focuses on the use of ANNs for continuous wearmonitoring, and the detection and differentiation of tool breakageand collision from tool wear. It also offers a thorough synthesis ofa decade of generic research regarding the adequate sensing, sig-nal processing, and extraction of features out of sensor signalswith a focus on the ANN paradigm at the process model anddecision-making levels. Most recent surveys on the use of other

Fig. 1 Concept of CBM as transformation of sensing data intoinformation about equipment condition and further into main-tenance and operational decisions

Fig. 2 Performance assessment and diagnosis through over-lapping of signature distributions

Fig. 3 Concept of feature-based performance prediction withprediction confidence intervals

Fig. 4 Generic sensor fusion architecture described in Ref. †7‡for ANN application to tool condition monitoring

041015-2 / Vol. 132, AUGUST 2010 Transactions of the ASME


methods for creation of process model and decision-making, suchas expert systems, fuzzy logic, and statistical pattern recognitioncan be found in Refs. !9,10". This section presents the most recentadvances in the area of indirect sensing-based tool conditionmonitoring in turning.

2.1 Advances in Sensing and Hardware. Acoustic emission#AE$ sensors have significant applications in TCM. AE includes aclass of phenomena in which elastic waves are generated by therapid release of energy from local sources within the material.These waves propagate through the structural elements of the ma-chine and workpiece generating significant information content inthe MHz frequency band. Increased attention to AE based toolcondition monitoring in turning #as well as in other areas$ wasspurred by recent advances in computational technology, permit-ting the processing of these high-bandwidth signals enabling in-creased use of AE signals. In Refs. !11,12", the authors utilizedbasic AE signal features and pursued increased sensitivity of AEbased tool condition monitoring to wear in turning through the useadvanced pattern classification methods. In Ref. !13", the authorsstudy the statistical properties of AE signals through the analysisof the AE amplitude and root-mean square #RMS$ data. It wassuccessfully demonstrated that aging features could be seen inexperimental histograms of amplitude and RMS features. Sensi-tivity to tool breakage of AE based tool monitoring in turning wasimproved in Refs. !14,15" through the use of nonstationary signalanalysis applied to the AE signals while a similar advanced signalprocessing approach was used in Ref. !16" to improve the sensi-tivity of AE based tool monitoring to tool wear in turning. Thisapproach helps to address the difficulty with AE approaches, par-ticularly, the method sensitivity to nonhomogeneity in the struc-ture through which the wave propagates and the resultant signaldisturbance.

Besides the use of AE sensing, many recent publications dem-onstrate that significant benefits can be obtained through the useof multiple sensors. In Refs. !17,18", AE and force sensing arecombined while feed-direction motor current and sound signalsare jointly considered in Ref. !19". The multisensor fusion in theaforementioned research was facilitated through the use of elabo-rate and sophisticated signal processing and pattern recognitionmethods, which are further described Secs. 2.2 and 2.3.

Rather than using high-bandwidth sensing or sophisticated mul-tisensor fusion, the research in Ref. !20" pursued a higher signal tonoise ratio for tool condition monitoring through data acquisitionaccomplished very near the actual cutting tool tip. Based on thetool tip position sensing and piezoelectric actuation signals inte-grated into a newly developed boring tool, the well-known distur-bance observation method #from traditional control theory$ is usedto provide an on-line estimate of the cutting forces. This tooldesign gives greatly increased flexibility during the boring processwithout decreasing its accuracy since increased compliance in thecutting tool is compensated through the piezoelectric tool tip ac-tuator and the tool tip deflection measurement sensor. These sig-nals are used to indirectly observe the cutting force, bypassing theneed for direct force measurements, which are both costly andmay incur changes in cutting tool dynamics due to the presence ofa force/torque sensor. The estimated force patterns were used tosuccessfully detect tool breakage and misalignment of the work-piece. The estimated forces might also be used for tool wearmonitoring, even though this application is not considered in Ref.!20".

Among recent hardware improvements in tool wear monitoringin turning, it is also worth mentioning the work in Ref. !21",where robustness improvements important for actual industrialimplementation of a tool condition monitoring system for turningare pursued through hardware integration of simple force sensingdata acquisition with elementary feature extraction and patternrecognition methods onto a single chip. This highly specialized,robust and rugged system was tested in actual industrial environ-ment and demonstrated excellent results in terms of on-line detec-

tion of tool breakage but the researchers noted that it was not assensitive to turning tool wear. This is not surprising since theintegration constraints necessitated the use of rudimentary pro-cessing pattern recognition methods that have been shown in thepast to cause sensitivity issues.

2.2 Advances in Signal Processing and Feature ExtractionMethods. Analysis of recent work in TCM in turning indicates astrong shift toward the use of advanced, nonstationary signalanalysis techniques that jointly analyze the distribution of signalenergies in various time intervals and frequencies ranges #in thissection, the term “nonstationary” signals is used for signals whosefrequency content varies over time$. This is enabled by advancesin computational technology and the development of new signalprocessing algorithms, and facilitates a more accurate representa-tion of time-varying, nonlinear, and stochastic dependencies oftool wear on various signals !19,22,23".

Wang et al. !22" extracted dynamic characteristics of tool wearfrom Daubechy’s wavelet coefficients of the vibration signals. Inhis work, signal energies at various scales were used as a featureset for a hidden Markov model #HMM$ that evaluated the likeli-hood that the observed signals came from either a worn or a sharptool. The tool was assigned to the worn or sharp tool class basedon which situation had a higher likelihood according to the HMM.A similar signal processing approach was also used in Ref. !24"where flank wear in turning was assessed using sound signals. InRefs. !15,14", time-scale analysis of AE signals based on Haarwavelets was coupled with simple statistical process control#SPC$ charts of wavelet coefficients to enable the detection ofsignificant wear or tool breakage in turning. Gao and Xu !17" alsoused wavelet analysis to process AE and three orthogonal forcesignals generated during turning to extract a number of featuresfrom various frequency bands of those signals. They subsequentlyutilized feature-by-feature correlation to select features that arethe most correlated with the actual tool wear in order to accom-plish multisensory TCM through the use of ANNs. Scheffer andHeyns !25" used multisensory data into ANNs to develop a wearmonitoring system. More recently !16", Daubechy’s wavelet coef-ficients of raw AE signals emitted during turning were analyzedand compared between a sharp and a worn tool. Through qualita-tive observations, it was noted that the wavelet resolution coeffi-cient norm, rather than modulus maxima, was a more stable fea-ture distinguishing between a sharp and a warn tool.

Liu et al. !26" departed from the prevalent trend of utilizingwavelets to decompose nonstationary signal energy over time andfrequency #or rather over time and scale domains$, and utilizeCohen’s class time-frequency distributions for signal energy de-composition !27". Spindle load signals emitted during 2 weeks ofactual industrial boring operations in automotive industry wereanalyzed using higher order moments of reduced interference dis-tributions #RIDs$. This signal processing technique inspired byachievements in quantum physics is relatively unknown in themanufacturing community because regular computers only re-cently became capable of producing these distributions because oftheir immense complexity. Nevertheless, at the expense of signifi-cant computational load, RIDs possess a number of mathematicalproperties that make them increasingly interesting, especially withthe further advances in computational technology !28". Figures 5and 6 illustrate the benefits of the use of RIDs and the clarity withwhich RIDs represent nonstationary signals #figures taken fromRef. !29"$. The well-known #and routinely used$ frequency-domain signal representation used in Fig. 5 is not able to differ-entiate between the two signals since their frequency contents areidentical #they are different only in when each of the frequenciesappeared in the signal$. In Fig. 6, the two frequency hoppingsignals are easily differentiated by the corresponding binomialRID distributions that clearly show what frequencies existed inthe signal and when they existed in the signal.

Salgado and Alonso !19,23" departed from the main trend ofwavelets utilization and proposed the use of singular spectrum

Journal of Manufacturing Science and Engineering AUGUST 2010, Vol. 132 / 041015-3


analysis #SSA$ for tool wear detection in turning. Unlike Cohen’sclass of time-frequency distributions, SSA deals with nonstation-ary signals in a relatively simple manner. First the data are pro-cessed using a sliding window technique, and the windowed dataare used to construct a Hankel matrix. Singular value decomposi-tion #SVD$ of this matrix robustly isolates the noise from thesignal components carrying the energy in various frequencybands, allowing a more precise analysis of the signal’s frequencycontent. In Ref. !23", the SSA is applied to longitudinal and trans-verse vibration signals, observing that the basic statistical proper-ties #mean and variance$ of SSA-extracted “noise” components ofboth vibration signals demonstrated a direct dependency on theflank tool wear. This is consistent with observations from earlierresearchers that the high-frequency range of vibrations #whichlives in the “noise” terms extracted through the SSA$ is moreaffected by tool wear. Nevertheless, SSA based identification ofthe high-frequency signal components enabled this dependency tobe more pronounced, improving subsequent training of an ANNfor estimation of flank wear based on those features. Similar rea-soning is used in Ref. !19", where high-frequency component ofsound signals were identified using SSA after which SSA-extracted features of sound signals were merged with cutting forcelevels estimated from feed motor currents to accomplish multisen-sor estimation of tool wear through an ANN.

One should note that better availability and feasibility of new,powerful signal processing methods does not completely usurp theuse of more traditional methods for stationary signal analysis,such as the use of basic statistical properties #mean, standard de-viation, RMS, kurtosis, etc.$, time-series analysis, and/orfrequency-domain analysis #enabled through the use of the Fast

Fourier Transform$. Robustness and ease of implementation ofthese more traditional solutions has significant appeal, especiallyfor industrial implementation. In Ref. !30", the resultant force ofthe tangential, feed, and radial force components is utilized fortool breakage detection in turning. In Refs. !21,31", raw readingsof the feed and cutting direction forces are used for the estimationof tool wear. Cutting forces were also employed in Ref. !32" whenthe mean cutting force was used to estimate turning tool wearlevels. This work was significantly improved in Ref. !18", wheretool wear levels in turning were estimated using the median andvariance of forces in the feed direction as well as the mean andminimum values of the RMS time series of AE signals. Morerecently, Du and Yeung !33" utilized the mean values of thespindle motor and feed motor currents #coupled with the off-linecalculated mean material removal rates$ to monitor the progress oftool wear in boring while Srinivasa-Rao et al. !34" utilized rawtemperature readings from a K-type thermocouple embedded atthe bottom of the turning tool insert for on-line prediction of dif-fusion flank wear. Prediction results closely match observed flankwear diffusion data as shown in Fig. 7.

Despite the fact that this significant body of research!18,21,30–34" avoids the utilization of nontraditional signal pro-cessing methods, it does employ advanced computational con-cepts by coupling traditional techniques with highly complex andsophisticated health assessment methods such as HMMs, ANNs,fuzzy logic, and others #these advanced pattern recognition meth-ods and function approximation methods are discussed in moredetails in Sec. 2.3$. A relative departure from the aforementionedtrend is the work in Ref. !35", where the authors use the autore-gressive moving average #ARMA$ modeling for dynamic analysisof the deterministic and stochastic components of the turningforce signals in the feed direction. For each wear level, theGreen’s Function #impulse response$ coefficients of the corre-sponding ARMA models are extracted and tracked as the cuttingtool wears out. It is observed that ARMA models of higher au-toregressive orders are adequate for higher wear levels, thus indi-cating the occurrence of novel dynamic modes with the progressof tool wear. One could say that rather than using sophisticatedpattern recognition and function approximation methods, the au-thors of Ref. !35" effectively augment their basic stationary signalprocessing and feature extraction technique with the knowledge ofthe physics of the problem and achieve good results.

2.3 Advances in Tool Condition Health Assessment andPrediction Methods. Recently, advances in pattern recognitionhave enabled methods for identification of various states of toolwear in turning based on features extracted from various signals.Furthermore, sophisticated function approximation techniques forexpressing quantitative turning tool wear characteristics in termsof signal characteristics has also been developed. These ap-proaches enable researchers to pursue TCM using dynamic signalfeatures rather than static or “snap shot” characteristics !36". Forexample, rather than indicating tool breakage when SPC basedforce threshold is sensed, such an indication state is not set untilthis threshold has existed for two revolutions of the part. Such an

Fig. 5 Application of Fourier analysis on two frequency hop-ping signals

Fig. 6 Binomial joint TFD of the two frequency hopping sig-nals identical to those analyzed in Fig. 5

Fig. 7 Neural network prediction of turning flank wear versusindependent experiment †30‡



approach has yielded significantly lower false indications !30".A more sophisticated, yet theoretically tractable, and intuitive

method able to capture the dynamics of a sequence of signaturesis the use of HMMs for the modeling of features extracted out ofsignals. HMMs consist of a finite set of “hidden” states that aretraversed according to certain state transition probabilities. Statetransition probabilities pi,j capture the temporal dynamics of theprocess by modeling probabilities of the process in state i movingto the hidden state j in the next time step. #Markovianity stemsfrom the fact that these transition probabilities for the next stateare dependent only on where the process is at present$. HMMstates are not directly observable #they are “hidden”$ and informa-tion about them is conveyed through observations that are sto-chastically related to each state through an observation probabilitydensity function. This makes HMM ideal for modeling CBM.

Since HMMs are a natural fit to CBM, it has been successfullyemployed in a number of research efforts. In Ref. !22", twoHMMs are fit to two respective sequences of wavelet packet en-ergies #a sharp tool packet and one from a worn tool$. These twoHMMs are used to classify subsequent observations as sharp orworn, based on the likelihood of that particular sequence emanat-ing from the sharp tool or warn tool HMM. This technique isextended to different wear levels in Ref. !36". It is important tonote that each level of wear must be initially characterized by anHMM. The authors also make use of the fact that wear progressesonly in one direction over time #e.g., the tool does can only wearand not sharpen$. The state transition is unidirectional in nature. Astandard Baum–Welch training algorithm is used to identify theHMM parameters #state transition probabilities and output prob-ability density parameters$.

One potential weakness of the use of HMMs for TCM is itsstrong dependency on initial data to set-up the HMM. Such initialdata may require significant efforts to develop. In Refs. !33,37",incorporation of expert knowledge into a Markovian model frame-work is accomplished through a merger of fuzzy logic and HMMsin the fuzzy-transition probabilistic #FTP$ framework. Stochastic-ity of signal features in different stages of system degradation#tool wear in the case of TCM$ are represented as Fuzzy setswhile the concept of transition probabilities for modeling of tem-poral dynamics of signal features is inherited from the more tra-ditional HMMs. This new method for monitoring of progressivefaults is demonstrated in Ref. !33" for crack propagation monitor-ing and boring tool condition monitoring based on mean values ofthe main spindle and feed motor currents. It should be noted thatHMMs, as with any model/empirical based approach is typicallyquite good at interpolating within its training set but unexpectedand incongruent results can occur if the approach is used to ex-trapolate to states that have not been considered in the initial dataset.

Another significant trend in the last 5 years is the increased useof artificial intelligence #AI$ methods, such as ANN, fuzzy logic,genetic and evolutionary computation, and support vector ma-chines #SVMs$. An ANN is a function approximation and infor-mation processing paradigm based on the heuristics of intercon-nected simple#r$ computational units that function similarly toneurons in a human brain !34". Salgado and Alonso !23" utilize athree-layer #one hidden$ multilevel perceptron #MLP$ ANN to es-timate flank wear of a turning tool based on the statistical proper-ties of features extracted from SSA of longitudinal and transversevibrations. The number of hidden neurons was established al-though trial and error, minimizing the RMS of ANN estimationerrors evaluated on a testing set of data. Another three-layer MLPANN was used in Ref. !34", estimating diffusion flank wear of theturning tool insert based on process conditions #cutting speed,feed, depth of cut, material properties of the tool and workpiece$and temperatures obtained from the bottom of the turning insertusing a k-type thermocouple. Results demonstrated a significantimprovement over the sole use of Fick’s law. Compared with

physics only methods !34", the need to retrain the ANN for eachcombination of tool and workpiece materials represents one weak-ness of the ANN-based approach to tool wear estimation.

More recently, Salgado and Alfonso !19" augmented their workpresented in Ref. !23" by using a least-squares version of a sup-port vector machine #LS SVM$ to accomplish multisensory fusionof feed-direction forces from the ac motor currents with the SSA-extracted features from the sound signals. Support Vector Ma-chines are a relatively novel machine learning tool particularlywell-suited for learning with small sample sizes. Unlike ANNs,SVMs have a solid statistical grounding and a high capacity forgeneralization !23". The least-squares version of SVMs used inRef. !23" led to a solution based on only systems of linear equa-tions #rather than a system of nonlinear equations that needs to besolved in the case of a general SVM$, yielding a solution with agreater analytical tractability. Testing and validation of the meth-odology were performed on a set of cutting conditions that werenot presented during the training process. The authors demon-strated that even though LS SVM and ANN #MLP ANN, similarto the one used in Ref. !19"$ had similar accuracy for large train-ing data sets, the LS SVM was able to provide highly accurateestimates of tool wear even when the number of training sampleswas reduced to the point where the ANN accuracy wascompromised.

Jelmeniak and Bombinski !18" utilized an ANN #feed-forwardback-propagation #FFBP$ MLP$ to augment a univariate approachfor turning tool condition monitoring introduced in Ref. !32". Themethod introduced in Ref. !32" is essentially a univariate methodexpressing the remaining useful tool life based on a single, vary-ing signal feature #in this paper, authors demonstrate the use ofmean cutting forces$, which is related to the remaining tool lifeusing a simple mathematical relation with time-instances whenoperator notices unacceptable tool wear #either based on productdimensionality variations, or based on jeopardized surface integ-rity$. In Ref. !18", these univariate estimates are fed into an ANN,facilitating multivariate feature considerations in a hierarchicalmanner. First level estimations are accomplished using individualfeatures through formulae proposed in Ref. !32", followed byANN enabled merger of individual tool life estimations at the nextlevel$. In Ref. !18", the authors used the median and variance ofthe feed force signal as features, as well as the median and vari-ance of the RMS of AE signals. The novel concept in this ap-proach is that remaining tool life is estimated, rather than toolwear. The approach does not estimate wear but predicts whenunacceptable products will be machined, effectively incorporatingobservations and knowledge of machine tool operators into thehierarchical TCM system. Furthermore, as this approach does notestimate tool wear, the need for direct measurements of tool wearare not necessary.

A more traditional approach for incorporation of expert knowl-edge is through the use of fuzzy logic. Balazinski et al. !31" useda fuzzy decision support system representing expert knowledge inthe form of a set of “if-then” rules, connecting fuzzy-inputs#simple signal features extracted from the cutting and feed-direction forces$ and fuzzy-outputs #levels of tool wear$. The needfor expert knowledge in defining fuzzy rules is significantly di-minished by the work in Ref. !21", where the paradigm of geneticalgorithms is utilized to facilitate the autonomous generation offuzzy rules. The survival of the fittest heuristic is used to generatenew rules out of those yielding good predictions of wear whilediminishing the influence of #or removing$ the rules that resultedin poor wear prediction. The basic application of fuzzy-logic inturning TCM was enhanced in Ref. !31" by casting the problem oftool wear estimation as a set of computational fuzzy rules withGaussian membership function, making the evaluation schememathematically equivalent to Radial Basis Function ANNs. In thiscase, the operator only needs to specify the number of fuzzy rules#i.e., the number of nodes of the ANN—obtained through trial anderror$ after which model parameters are optimized based on the



training set. In Ref. !17", fuzzy logic was used to define a two-stage approach where B-spline ANNs are trained to estimate toolinsert wear based on individual sensor readings, followed by afuzzy-model sensor fusion of ANN outputs. The tool wear estima-tion accuracy of individual B-spline ANNs was significantly im-proved when a fuzzy model refined the tool wear estimations fromindividual ANNs. Experiments with turning under 64 differentcutting conditions with different levels of tool insert wear yieldedremarkable accuracy of wear estimation through ANN/fuzzy-model fusion simple features of three orthogonal force readingsand AE RMS signals.

The recently introduced concept of support vector machineswas modified and utilized for TCM in turning in Refs. !11,12".SVM modification in Ref. !12" enabled multiclassification of turn-ing signals to discern various levels of tool wear with explicitconsiderations of costs associated with misclassifications incorpo-rated into the SVM parameter adjustments. The modified SVMwas subsequently utilized in Ref. !11" to devise a methodology fortraining data selection based on pruning the redundant trainingitems using the generalization error surface of the SVM. The gen-eralization error surface was generated through the k-fold crossvalidation procedure, where the training set is randomly split intok mutually exclusive subsets #“folds”$ of approximately equal sizeafter which k-1 folds are used for training and the kth fold is thenused for evaluation of generalization. This procedure is repeatedto make each of the folds the evaluation fold and generalizationerrors are then averaged.

Recently, the dynamic and AI-based methods for tool healthassessment were tested and compared. Such studies are particu-larly important in light of increased availability and feasibility ofuse of such methods in TCM. Balazinski et al. !31" comparedMLP-based, basic fuzzy-rule based, and ANN-based fuzzy infer-ence system #essentially a Radial Basis Function-RBF ANN$ ap-proaches for turning tool wear estimation using force readings incutting and feed directions. The authors observed a comparableaccuracy in estimating tool wear in all three cases and concludepractical convenience rather than accuracy should be the maincriterion for selecting an appropriate AI health assessment methodfor TCM. Construction of the fuzzy rules to establish a fuzzy logicbased TCM system necessitates existence of accurate and reliableexpert knowledge, limiting this method significantly. For the MLPand RBF based health assessment tools, the main obstacle is theneed to determine internal structure of the ANN #number ofnodes, number of hidden layers and connections among thenodes$. However, the RBF approach demonstrated significantlyshorter training time than required by the MLP approach. Schefferet al. !36" compared an ANN approach to an HMM approach forcondition assessment in TCM in turning. For both the ANN andthe HMM, the authors utilize an identical data set consisting offorce sensor readings #in three orthogonal directions$ collectedunder several cutting conditions. An elaborate ANN scheme isutilized to establish a connection between the tool wear levels VB,and the four features found to be the most correlated with the toolwear levels #using a simple feature-by-feature correlation$. Fourrelatively simple static networks #SNs$ were used to model thedependency of the four force signal features on the tool wearwhile a dynamic network #DN$ was used to capture dynamics ofthe wear growth. This addresses the fact that tool wear seldomfollows the same geometry and growth rate. Structures of bothSNs and the DN were selected ad hoc #feed-forward networkswith three layers$. In terms of accuracy of tool wear estimation,both ANN and HMM-based approaches achieved comparable andsatisfactory results with the ANN providing a better fit to theactually measured tool wear data according to the !2 statisticaltest. It was demonstrated that ANNs inherently produce continu-ous results and are more suitable for continuous estimation prob-lems #e.g., tool wear estimation$. Conversely, HMMs are tradi-tionally aimed at estimating stratified levels of the unknownvariable #or states of that variable$. It should be noted that HMMs

can be adapted to continuous estimation problems !36". A signifi-cant weakness of ANNs is that their architecture must be adjustedvia trial and error. An advantage of HMMs is that they are awell-known and well-understood with readily available numericalsoftware tools enabling easy implementation of HMM-based con-dition monitoring solutions. However, they do require a signifi-cant amount of training data.

Temporal performance predictions for a real industrial metal-cutting process are reported in Refs. !26,38". In both papers, time-frequency moments of binomial time-frequency RID distributionswere used to describe performance of a boring process in an au-tomotive plant. In Ref. !38", an Elman recurrent neural network#ERNN$ was constructed to predict the behavior of those featureswhile in Ref. !26", a novel prediction algorithm was introduced,based on quantitatively expressed similarities between the cur-rently observed cutting process behavior and the library of behav-ioral signatures collected in the past. In both papers, uncertainty ofprediction was assessed, which is necessary for evaluation of theremaining useful life and prediction of probabilities of unaccept-able behavior over time, as illustrated in Fig. 3. In addition, meansquared prediction errors were used to evaluate the quality ofprediction and the newly introduced methods were compared onthe same set of data with more traditional prediction methods onthe same data set #very unique and also very useful$. Figure 8#from Ref. !26"$ shows comparison between the ERNN basedprediction introduced in Ref. !38", and prediction based on the“match matrix” method introduced in Ref. !26". With a relativelylonger preview horizon, the method introduced in Ref. !26" per-forms the best.

Most recently, research in TCM applied to turning did not em-ploy advanced heath assessment methods !14,15,26" and optedinstead for more traditional Gaussian statistics based approaches#SPC in Refs. !14,15" and Mahalanobis distances in Ref. !26"$.Nevertheless, it is the use of advanced wavelets and time-frequency signal analysis in those papers that brought about amore analytically tractable behavior of signal features #approxi-mate Gaussian nature$ and enabled the use of less elaborate toolhealth assessment methods.

3 Tool Condition Monitoring for DrillingDrilling is one of the most frequently encountered machining

operations, accounting for over 30% of all cutting operations inindustry !39" and over 40% of cutting operations in the aerospaceindustry !40". The existence of multiple cutting edges with vari-able cutting speeds along those edges, increased interaction be-tween the metal chips and the cutting tool during chip evacuation,

Fig. 8 Comparison of prediction errors for ERNN and matchmatrix based prediction



and significantly altered heat transfer characteristics comparedwith the turning process make the task of tool condition monitor-ing in drilling significantly more challenging. Monitoring toolwear in a drilling process is the subject of study in the reviewwork !41". Significant new developments in sensing and hard-ware, in signal processing and feature extraction, and in toolhealth assessment methods for drilling tool condition monitoringhave been made in the last several years and are reviewed in thissection.

3.1 Advances in Sensing and Hardware. Recent research indrilling tool condition monitoring has introduced increased use ofAE sensing !42,43", as well as concurrent use of multiple sensors!43–45". Similar to turning operations, high-bandwidth AE signalscarry information about microscopic damage created by the cut-ting mechanism and changes in the tool condition. This is depictedin the high-frequency emission of stored elastic energy travelingthrough both the workpiece and machinetool in the shape of elas-tic stress waves !42". Recent developments in computing technol-ogy enabled the real-time acquisition and processing of thesehighly transient signals permitting their use in drilling tool condi-tion monitoring. In Ref. !42", raw AE signals were used to detectthe critical point in drill life when increased wear begins to de-velop and tool change needs to take place. More recently, theRMS values of AE signals were coupled with thrust force andtorque measurements to enable multisensory condition monitoringof a drill bit in drilling of small, deep holes #aspect ratio of 10 orhigher$ !43". Evolution of the total cycle AE signal over the life ofa drill is given in Fig. 9; imminent failure is easily identifiable.

Besides !43", multisensory approaches to tool condition moni-toring in drilling were also reported in Refs. !44–46". In Ref. !46",a simple simultaneous thresholding method is proposed for merg-ing thrust force and cutting torque signals. In Ref. !44", the thrustforce and cutting torque sensor readings were merged with cuttingcondition parameters #spindle speed, feed rate, and drill diameter$through an ANN to estimate flank wear of a drill bit. It was ob-served that significant improvements in accuracy of drill bit wearestimation could be obtained if optically measured chip thicknesswas also used for flank wear estimation. Inclusion of this param-eter necessitates off-line measurements of chip thickness, some-what limiting the resulting tool condition monitoring. Choi devel-oped a technique using machine vision for microdrilling !47". InRef. !45", a simple fusion of spindle motor current and voltagemeasurements through estimation of the input impedance of thespindle drive yielded excellent results in terms of the recognitionof tool breakage in drilling.

In addition to the aforementioned advances, the use of moretraditional TCM sensor signals, such as vibration signals !48",thrust force !46", torque !46", power !49,50", spindle motor cur-rents !39,51,52" and vibrations !48", received significant attentionin the recent years. Signatures obtained from the clamping fixtureof the workpiece were used in Ref. !48" to recognize five differentwear conditions on a twist drill #chisel wear, crater wear, flankwear, edge fracture and outer corner wear$. In Ref. !46", HMMmodels of raw thrust force and cutting torque readings were sepa-rately evaluated in terms of their sensitivity to tool wear with aconclusion that the thrust force demonstrated a higher sensitivity

to tool wear. Furthermore, in the same publication, two computa-tionally simple methods were proposed for drill bit condition as-sessment based on cutting torque sensing alone #in addition to theearlier mentioned simple, multisensory approach based on simul-taneous thresholding of the thrust force and torque signals$. InRef. !53" bar graph analysis of HMMs is used to monitor drillcondition in situ. Anomalies in spindle motor power were used inRef. !54" to attempt to control tool wear and in Ref. !50" fordetection of tool breakage while Al-Sulaiman et al. !49" demon-strated a high correlation between drill bit flank wear and thedifferential electrical power #the increase in spindle power over itsidle level during cutting$. Spindle motor current was used in Ref.!51" to introduce a measure sensitive to drill flank wear while inRef. !52", spindle motor current features were used to explicitlyestimate the drill flank wear. Finally, in Ref. !39", an early warn-ing tool-replacement decision scheme was proposed based on thefeatures extracted from the spindle motor current.

Less costly and technically less intrusive sensing solutions re-ported in Refs. !39,46,48–52,55" retained their appeal in spite of alow signal to noise ratio due to the increased ability to morereadily use the advanced signal processing, feature extraction, andhealth assessment methods to extract and recognize minute fea-tures within these signals that are related to the tool wear andbreakage. In the ensuing text, recent advances in signal processingand feature extraction as well as in health assessment in TCM indrilling will be reviewed.

3.2 Advances in Signal Processing and Feature ExtractionMethods. One of the most successful and heavily investigatedtechniques used in drilling TCM is wavelet analysis, which lendsitself to drilling as it can address the strong nonstationarities insignals emitted by the process !42,48,51,52". Abu-Mafhouz !48"utilized harmonic wavelets with boxcar spectrum !56" to filtervibrations signals and extract 16 consecutive averaged waveletcoefficients from each signal segment of the 4096 data samples.These time-frequency !27" based features were combined withtime-domain features #mean value, variance, skewness, and kur-tosis extracted from the vibration signal time series$ andfrequency-domain features #eight highest local maxima of the vi-brations spectrum !57"$ extracted from the same segment of the4096 vibration signal samples. This combined feature vector con-taining the total of 28 features was used to detect and recognizefive different drill bit wear conditions #chisel wear, crater wear,flank wear, edge fracture, and outer corner wear$. Velayudham etal. !42" defined a crest factor of AE signal wavelet coefficients asthe ratio of the range and mean value of the coefficients of thewavelet packet with highest energy, which they could relate todifferent stages of drill bit flank wear in TCM for drilling ofglass-phenolic composite materials. Material inhomogenitiespresent in such materials further complicate drill bit TCM andincrease the nonstationary signal characteristics emitted during thedrilling process. During the normal wear stage, crest factor in-creased due to the increasing energy work. At the onset of severewear, the crest factor decreased, probably due to the increased lowfrequency emissions caused by rubbing between the workpieceand the severely worn tool #i.e., more energy becomes emittedoutside the frequency range picked up by the AE sensor$. Such adrop in the crest factor is proposed to be an indication of the needfor tool change. Unfortunately, results of wearing out only onedrill bit #under constant cutting conditions$ were reported anddemonstration of the same concept with another drill bit and underdifferent cutting conditions would be interesting. Franco-Gasca etal. !51" utilized the well-known advantageous properties of or-thogonality and compact support of Daubehy’s wavelet transformsto filter the spindle current signals to identify quasi-periodicpulses in the spindle current. It was observed that the dissimilari-ties between pulses #evaluated through an asymmetry measurecalculated as the point to point variance between successivepulses$ increased as the drill bit wear progressed. Just like in Ref.!42", wear out of only one drill bit was reported in this paper.

Fig. 9 Development of AE process signal over drill life †38‡



Wavelet packets were also used in Ref. !52", where a formal sen-sitivity analysis demonstrated a higher sensitivity of waveletpacket features to drill bit flank wear as well as lower sensitivityto cutting conditions than those observed for the case of moretraditional time-domain features. More recently, Choi et al. !39"combined mean and variance of the feed motor current waveletcoefficients encompassing the spindle rotating frequency with thenormalized average and standard deviation of the time-series ofthe feed motor current #i.e., combined time-scale and time-domainfeatures$ to enable early warning about impending drill failure.

More traditional, time-domain features have been successfullyemployed for drilling TCM. Heineman et al. !43" utilized physicalreasoning to analyze RMS values of AE signals as well as rawthrust force and torque sensor readings and relate them to wearprogress in drilling of small holes. The authors propose segment-ing the torque curve into three time segments such that, when thetool is sharp, the area under the first and the last segment wasroughly equal to that of the middle segment. Moderate tool wearwas observed to increase the energy of the final segment due toincreased chip clogging. Severe tool wear was characterized bystrong microwelding and serious chip clogging, which in turncaused the middle segment to have a significant increase in energycompared with the sum of the other two. Al-Sulaiman et al. !49"observed a time-series of the differential spindle power #differencebetween total electrical power dissipated by the spindle motor andthe spindle running power dissipated when cutting is not takingplace$ and observed a high correlation with the optically measureddrill flank wear. A 33 design of experiments #DOEs$ with drilldiameter, feed, and speed rates as experimental variables was usedto assess the sensitivity of differential power to the drill flank wearand formally demonstrate the increase in sensitivity comparedwith the more traditionally utilized raw electrical power of thespindle. Average thrust force and torque readings were used inRef. !46" #as inputs for a phase plane feature map$ and in Ref.!44" #as input features for an ANN$ while raw time series ofspindle motor input impedance #obtained from the time series ofspindle motor current and voltage$ was used for robust detectionof drill bit breakage. More sophisticated time-domain feature ex-traction can be found in Ref. !50", where principal componentanalysis was used to extract principal components of the spindlemotor power signals for detection of severe drill bit wear orbreakage #detected as significant abnormalities in the signal$. An-other interesting time-domain feature extraction can be found inRef. !46", where time-series of thrust force and torque sensorreadings were separately characterized through the use of hiddenHMM with the posterior probability of each time series occurringgiven a HMM describing the sharp tool cutting behavior, servingas the indicator of drill bit wear #lower posterior probabilitieswould indicate a more severe wear stage$. In the same paper, theauthors also define a simple feature extraction scheme using cut-ting torque signals to extract the transient time during which thecutting lips enter the workpiece. Progress of corner wear extendsthe length of the cutting lips and thus slightly increases the tran-sient time. In addition, a mechanistic approach was proposed inRef. !46" where parameters of an empirical model #connecting thedrill geometry and workpiece parameters with the cutting torque$are fit to the cutting torque data !58" and their drift over succes-sive holes drilled by a given drill bit is related to the drill bit wear.While the time-domain features utilized in Refs. !43,44,46,49,50"are not as representative of variation patterns in the time andfrequency as wavelet coefficients are, tool wear characteristicsfrom these features were extracted using sophisticated conditionassessment methods. These methods included dynamic patternanalysis through HMM modeling !46", use of artificial intelli-gence methods !44,50", and use of physics principles and empiri-cal knowledge !46,49". This made the computationally less inten-sive time-domain feature extraction methods viable in spite of the

lower signal to noise ratio characterizing the time-domain fea-tures. Condition assessment methods utilized for TCM in drillingare discussed in the next subsection.

3.3 Advances in Tool Condition Health Assessment andPrediction Methods. There are several different approaches inusing ANNs in drill TCM. In Ref. !48", a three-layer FFBP net-work was used to recognize five different drill wear conditions#chisel wear, crate wear, flank wear, edge fracture and outer cornerwear$ and separate them from the new drill bit condition. In thecontext of this classification problem, it was observed that decou-pling of network architecture and moving away from a fully con-nected three-layer architecture resulted in accelerated convergenceduring learning without jeopardizing the classification accuracy.These observations were made only on a single set of experimentsconducted under identical cutting conditions and one could benefitgreatly if similar simplifications of ANN architecture could bemade in general.

In Refs. !44,52", ANNs were used to explicitly express flankwear of a drill bit based on multiple signal features. In Ref. !44",a two hidden-layer back-propagation ANN was used to estimatedrill bit flank wear based on a set of on-line and off-line features.The on-line features were mean values of cutting torque and thrustforce. The off-line features denoted the hole diameter, spindle feedand spindle speed$. A total of 52 drilling tests were conducted inmild steel with high speed steel drill bits over a wide spectrum ofdifferent cutting conditions. Thirty-nine tests were used for train-ing and 13 for testing. It was shown that the inherent ANN limi-tations in generalization of training results without the use ofphysics principles necessitates additional training when TCM isneeded extrapolate results outside of the training data parameterset. More recently, a more elaborate feature set #wavelet coeffi-cients extracted from spindle currents$ was used in Ref. !52" topredict flank wear. Nevertheless, a similar limitation inherent toANNs can be observed with respect to this work too.

Besides classification and wear estimation, ANNs were used toaccomplish tool breakage detection in drilling processes. In Ref.!50", a back-propagation ANN was used to identify abnormal cut-ting conditions #tool breakage or missing tool$ using the principalcomponents of the spindle power signal. Results were robustlyverified using industrial data, including 3 months’ worth of dataobtained under conditions for which the ANN was not trained. Inthe training set, normal drilling feature vectors were associatedwith the ANN output of one while abnormal #broken/missing tool$conditions were associated with the ANN output of 0. The methodperformed perfectly. This can be explained by the fact that thephysics of power signals changes dramatically when the tool isbroken or missing, regardless of the cutting condition. Similarheuristics were used in Ref. !45" to detect drill breakage in micro-drilling using minimal value, mean and standard deviation ofspindle input impedance that were calculated using the spindlevoltage and current signals. Since the ANN was trained to recog-nize normal cutting conditions as 0, and abnormal as 1, an in-crease in the ANN outputs was clearly visible when drill breakagewas approaching and could be used as an early warning for toolreplacement before the tool breakage occurs. Formal setting ofANN output thresholds that can be used to identify a broken tool,or raise an early warning alarm was discussed in Ref. !39". Theydefined a period of time and wavelet domain features, which werefed into a three-layer ANN with output of the ANN being a drillstate index #DSI$ depicting the state of the drill and ranging from0 #normal drill$ to 1 #broken drill$. During training, training fea-ture vectors corresponding to a normal drill were associated withthe DSI of 0.1 while those corresponding to a broken drill wereassociated with 0.9. Six different cutting conditions for drilling inthe workpiece material ANSI 1045 were used for training. Testswith six other cutting conditions with another workpiece material,and on a different machine tool were conducted to verify theproposed methodology. Through these numerous experiments, the



authors empirically notice that in all experiments, the DSI ap-proached 0.9 several seconds before the drill actually failed, thus,effectively enabling early warning before drill breakage.

Several recent publications use physics based !42,43", or statis-tical reasoning !46,51" to accomplish tool health assessment. InRef. !42", an empirical crest factor was defined as the ratio of therange and the mean value of the AE wavelet packet coefficientwith the highest energy. A change in the trend of the crest factorfrom growing in the early stages of wear into descending in thelate stages of wear was proposed as the criterion for tool replace-ment. The physical reasoning for this behavior of the crest factorwas explained by the increased friction-induced emissions in thelower spectrum bands that occur in the late stages of wear and thatcannot be seen in the AE sensor readings. In Ref. !43", the torquecurve was partitioned into three disjoint segments, such that thearea under the first and last #third$ segment was equal to that ofthe middle #second segment$. In this paper, the authors considertool wear monitoring in drilling of small holes, and propose as-sessment of drill bit wear through tracking of the ratio of areaunderneath the second segment and the sum of areas underneaththe first and third segments. With moderate levels of tool wear,chip clogging significantly increased the energy in the last #third$segment, thus causing the ratio to drop below 1. In the later stagesof wear, increased friction and more severe chip clogging duringdrilling caused and increase in middle segment energy, driving theratio to grow above 1. Such physics based on cutting behavior wasobserved in several combinations of drill bit/workpiece materialcombinations.

Franco-Gasca et al. !51" utilized statistical reasoning to trackdrill bit wear. They proposed an asymmetric measure to assessinconsistencies in the wavelet coefficients extracted from thespindle motor current signals. During normal tool operation, con-sistency in the cutting process is higher, resulting in higher asym-metric measures between features extracted from consecutive toolrotations. Nevertheless, asymmetric measure proposed in Ref.!51" is a purely static measure of statistical behavior of signalfeatures. A more dynamic approach was adopted in Ref. !46",where a HMM-based tool condition assessment scheme was pro-posed and tested on a data set obtained under constant cuttingconditions. Signals obtained from a sharp tool are used to con-struct HMMs of torque and thrust force readings corresponding tothe sharp drill bit operation. Based on those HMMs, emissionprobability of newly arrived torque and thrust force time-serieswere assessed. As the tool wore out, emission probabilitiesdropped, enabling indirect tracking of the tool wear. Experimentsreported in Ref. !46" demonstrated that thrust signals were betterindicators of drill bit wear than the torque signals since the uncer-tainty of observations of thrust force dropped with wear moresignificantly than what could be observed with the torque signals.

In summary, one can observe that even when research did notresort to advanced AI and statistical methods, such as the useANNs or HMMs, increased signal to noise ratio for successfuldrill bit condition assessment was achieved either through the useof advanced signal processing and feature extraction methods orthrough the use of physics based reasoning.

4 Tool Condition Monitoring in MillingWith changing tool engagement conditions in normal cutting,

the milling process presents unique challenges to tool conditionmonitoring. The recent advances in tool condition monitoring inmilling operations are reviewed in terms of advances in sensingand hardware, in terms of advances in signal processing and fea-ture extraction methods, and with respect to progress made in toolcondition health assessment and prediction methods.

4.1 Advances in Sensing and Hardware. In milling, just likein other areas reviewed in this paper, recent research reportedimprovements in TCM through improvements in sensing andhardware. An example of such work can be found in Refs. !9,33",

where the smart machining system combines open architecturecontrol, sensor data, and process models. Alongside a geometricsimulation of the cutting process and tool path, this allows foron-line updates to the cutting power model, improving the accu-racy of the cutting power prediction. These improvements mayallow for tool condition monitoring by examining changes in themodel. In order to extract cutting force signals from driver cur-rent, two-axis band-pass filter was used to suppress noises due toball-screw, control current, and commutation. A digital signal pro-cessing unit, implemented via a field programmable gate array#FPGA$, computes cutting force signal. Data acquisition and PCinterface was also implemented on the FPGA to give a system ona chip solution !59". The signal filtering result is shown in Fig. 10,where identification of a broken insert is apparent.

Increasing complexity of sensing and hardware as describedabove introduces additional costs and potential reliability/robustness problems into the TCM system. This becomes a barrierfor implementing such a system in industrial environment. To re-duce cost and improve robustness, Amer et al. !60" took a funda-mentally different approach, based on a reduced sensing schemeand a three-tier architecture. The first tier was implemented onprogrammable interface controller #PIC$ microcontrollers linkedtogether using a controller area network bus. These microcontrol-lers were used for acquiring the spindle speed and spindle loadsignals. The second tier controlled the first tier activities and wasdesigned to provide extra processing power if needed, and com-municate the data/health information to the third tier, the centraldatabase server. The central database processed and stored infor-mation for final decision making and could be accessed over theinternet. The system had a fault tolerant feature and was plug andplay. Zhang and Chen !61" also studied a low-cost approach byusing a microcontroller-based data acquisition system examiningvibration with a hardware cost of less than $200. Kang et al. !62"studied the measurement of cutting forces by attaching piezo loadcell to the feed system, which resulted in a 5% error comparedwith measurements with a dynamometer. Dini and Tognazzi !63"investigated a cost-effective approach by integrating low-cost ro-tating dynamometer directly into a tool holder body for tool con-dition monitoring in end milling with encouraging results.

4.2 Advances in Signal Processing and Feature ExtractionMethods. The methods used recently for signal processing andfeature extraction in TCM in milling can be characterized as time-domain based, frequency-domain based and time-frequency/time-scale based methods. Time-domain methods for signal processingand feature extraction utilize the time-domain waveform of thesignal to extract from it the features significant for description ofthe cutting performance. Time-domain based methods for signalprocessing and feature extraction in milling TCM include AR

Fig. 10 Drive current of spindle after analog filtering for OKcutting and broken insert cases †54‡



modeling !64,65", time-domain averaging #TDA$ !66", and im-proved TDA !67". In addition, basic statistical properties#maximum/minimum, mean, standard deviation, RMS, kurtosis,etc.$ were also used in recent literature !5,63,66,68,69".

Frequency-domain based methods are also frequently encoun-tered in TCM and general CBM since the frequency-domain de-scription of the signals often carries important information aboutthe underlying system dynamics. From this set of methods, Fou-rier transform-based methods were used in milling TCM in Refs.!61,63,65,70". In addition, the discrete cosine transform #a linearFourier-type transform$ was used for milling TCM in Ref. !65".Advances in computational technologies and the need for im-provement of the signal to noise ratio through advanced featureextraction led to an increase in the use of time-frequency andtime-scale signal analysis !27" based feature extraction methods.In terms of time-frequency/time-scale analysis, the use of wavelettransformation has obtained significant attention !59,66,71,72".Zhu et al. !71" transformed measured and simulated cutting forcesignals by wavelet to extract feature vectors for subsequent analy-sis. Data compression from 256 points to 8 was the objective ofDaubechies wavelet transform in Ref. !59".

Exploring beyond well-known methods were Peng !73" andAmer et al. !60". Peng !73" recognized the weaknesses of Fouriertransformation of cutting force signals, as they are sensitive tocutting conditions and require long time series to generate robustresults. As for wavelet transform, the concern was an inappropri-ate selection of a mother wavelet. To address these issues, a newlyemerging technique, empirical model decomposition #EMD$,based time-frequency analysis was used. The method was capableof analyzing nonstationary signals, such as those generated in acutting process. A key concept of EMD was instantaneous fre-quency #IF$, which was defined as the rate of change in the phaseangle at time t of the analytic signal. Another key concept was theHilbert #amplitude$ spectrum, which was an effective time-frequency distribution of the amplitude for the associated timeseries. Unfortunately, not all IF data are meaningful. Fortunately,EMD is an adaptive method to decompose an arbitrary time seriesinto a set of basic functions called intrinsic mode functions #IMFs$on which the Hilbert transform can be easily applied. This enablescalculation of a meaningful IF. There were still threshold andother judgment involved without a clearly defined methodology togenerate the IMF, nonetheless, in tool condition monitoring, therewas no threshold involved, which was interesting as thresholdsetting has been a major challenge.

Amer et al. !60" presented a sweeping filters technique, whichexecutes frequency analysis of the acquired monitoring signals,utilizing a band-pass filter capable of changing its characteristicsin real-time. It sweeps the entire frequency range of interest gen-erating total profile of the signal in terms of relative frequencypower. Filter band frequency was determined by the applicationand number of samples per cycles was determined by the errortolerance. By detecting the profile change, the technique was ca-pable of providing early warning of tool failure.

Separating signals due to cutting period from those due to non-cutting period allows the signal to noise ratio to be improved, byallowing for the known noise to be removed !66,74". utilize theinterrupt cutting nature of a milling process to assist in separatingthe noise from the desirable signal. Another physical phenomenonimproved the final results was a simple fact that wear was mono-tonically nondecreasing !66,74". Finally, cutting force, measuredor derived from another source, was the most widely used signal!59,62,65,66,68,71–78", followed by power consumption!5,69,79,80". Vibration !61,70", AE !5,69,74", acceleration,spindle signals, and current were also employed.

4.3 Advances in Tool Condition Health Assessment andPrediction Methods. One of the main unique traits of the recentdevelopments in milling TCM is the appearance of successfulmethods for effective prediction of tool failure. Cutting tool fail-ures can cause significant damage to workpiece and machine

tools. As a result, the ability to predict a cutting tool failure is ofgreat value. Roth !64" and Suprock and Roth !65" reported resultsof predicting impending tool failures due to several novel meth-ods. The focus of Ref. !64" was to identify an index predicting animpending tool failure independent of sensor orientations and cut-ting directions. The signals from triaxial accelerometer were pro-cessed by a multivariate autoregressive model. Since the eigenval-ues of the spectral matrix remained constant, the results wereindependent of cutting direction and sensor orientation. A modelrelating eigenvalue to length of cut in feed-direction was fittedusing least-squares, based on which a future state of the eigen-value was predicted. Once the mill reached the future state, thedifference between the actual value and forecasted value was cal-culated. This difference was used as the index. Repeating thisprocess gave numerous results of the index. An upper-bound 99%probability of the index was calculated using normal distribution.If the new index was above the limit, it represented either ananomaly or impending tool failure. Exceeding the limit twice in-dicated an impending failure.

Suprock and Roth !65" compared ten methods’ capability ofpredicting impending tool failure in frequency domain and timedomain: frequency tracking using the AR model, frequency track-ing using the Fourier transform, frequency tracking using discretecosine transform #DCT$, mean tracking of autoregressive spectra,mean tracking using Fourier transform, mean tracking using DCT,primary component analysis using the AR model and DCT, meananalysis using the AR model and the DCT, primary componentanalysis using the correlation function, and mean analysis usingthe correlation function. Logic similar to that of Ref. !64" wasfollowed here: the selected index formed a time series as the cut-ting tool progressed and a 99% confidence limit was drawn, whenthe index passed the limit, an impending failure was indicated.Only four out of the ten methods were identified as valid methods:frequency tracking using AR model, mean tracking using Fouriertransform, mean tracking using DCT, mean analysis using ARmodel and DCT. One major mode of failure among the other sixmethods was false failure prediction.

Besides tool failure prediction, more traditional topics of toolfailure detection and tool wear monitoring also received signifi-cant attention in the area of milling tool health assessment. Detec-tion of a tool failure after it happened can prevent further damageto workpiece and machine tool. The topic attracted significantattention !5,59,60,67,71–73,75,79". The work of Zhu et al. !71"took on one of the most challenging tasks for tool condition moni-toring, free-form surface machining with complex geometry. Theydiagnose the nature #flute chipping, breakage and spindle/cutteraxes runout$ and magnitude and faults via genetic algorithm. Toaddress the constantly changing tool engagement condition in afree-form surface machining, a process model was built to predictinstantaneous cutter engagement along the tool path.

Amer et al. !60" and Ritou et al. !75" utilized symmetric natureof a milling cutter with multiple flutes. When one tooth broke,asymmetric behavior was noted. The sweeping filter technique!60" linked frequency spectrum of machine tool signals to cuttingtool health. When a tooth broke, the frequency would change. Inaddition, tooth rotation energy estimation technique focusing onaverage energy per tooth rotation period and its variation to deter-mine tool condition was also used. The two techniques workingtogether minimized false alarms. Relative radial eccentricity ofcutters was used to detect cutter breakage and chipping !75". Theeccentricity was due to uneven distribution of load among teeth. Atwo-flute cutter was used and the assumption was that only onetooth was engaging in cutting at any given time. A constantneeded be identified for each tooth experimentally and link be-tween radial eccentricity and cutting force needed be established.To avoid false alarms due to transient cuts, only zones in whichTCM could perform reliably were monitored. About 90% of themachining time was monitored under the conditions of testing theauthors conducted. The research targeted small-batch or one-off



manufacturing.Another topic of significant interest was tool wear monitoring

!66,68,70,74,77–80". Commonly, tool wear was monitored as acontinuous variable !66,68,74,78,80". To achieve that objective, amodel linking tool wear to a monitoring index needed be estab-lished, for instance, cutting power was the monitoring index forRef. !80", the result of fusion of multiple inputs, and cutting forceswere used in Ref. !66". Kuljanic and Sortino !78" identified twotool wear indicators: normalized cutting force and torque-forcedistance, both based on cutting force signal analysis. Scatter plotsindicated correlations between these indicators and tool wear #seeFig. 11$.

Moreover, cutting parameters did not have a significant impacton these indicators. Since the torque-force distance did not requirea value from a sharp tool, it was selected as the basis of a toolwear estimation method. Having experimentally established therelationship to this indicator, tool wear could be monitored bymeasuring cutting forces and torque during milling operations andcalculating the indicator.

Fish et al. !70" explored the multilevel classification of millingtool wear to estimate the probability of a tool being worn. Theyanalyzed data from experiments done by Boeing using HMM andgeneral linear model #GLM$ and generated good results. The datacame from two different size tools with different cutting condi-tions. They reported that bias in the posterior probability out ofthe HMM resulted in an overconfident and unusable confidenceestimate and overcame the problem with a second stage classifier#GLM$.

ANNs were used to detect tool breakage !72", predict flankwear !68,74", estimate relative consumption of tool life !69", andpredict cutting force !76". A major challenge of applying an ANNis the proper training of the network. While sufficient training isnecessary for ANNs, too much training, or overfitting, leads topoor predicting capability of the trained network. Ghosh et al. !74"addressed the issue by randomly and independently generatingthree sets of data for training, testing, and validation purposes.They periodically checked the model with testing data in betweentraining set-based learning iterations. If both training error and testerror continued to drop, the learning process would continue. Oth-erwise, training would stop. Dutta et al. !68" addressed the issueof reducing training time by modifying the learning rule, enablingthe change and assignment of different learning rates acrossiterations.

The combining of different information into one index is a topicof significant interest in tool condition monitoring. This “sensorfusion” was achieved via ANN in Refs. !68,74". The input signalsof Ref. !74" included cutting forces, spindle vibration, spindlecurrent, and sound pressure level. What to combine was a criticalquestion, which was addressed by computing cross-correlationchart for feature selection for ANN inputs. Sensor signals in Ref.!68" included force and vibration. In addition, milling process

parameters and material properties were also used as the networkinputs. The critical question of what information to fuse was an-swered by comparing the final results due to different fusion strat-egies. Boutros and Liang !5", on the other hand, fused comple-mentary indices due to the same sensor to improve the reliabilityof the solution, reduce data processing efforts, and simplifythreshold setting. They performed the fusion via fuzzy logic andSugeno style inference engine, requiring significant amount ofwork to prepare and update detection rules.

Advanced mechanistic models for tool life and cutting forceshave also been employed in TCM for milling. Zhu et al. !71" usedmechanistic force model to determine threshold curve off-line.Yao et al. !69" tapped into such knowledge and built a virtualmanufacturing cell to calculate cutting forces and motor power inend milling for complex surfaces. These values fed into an ANNmodel to estimate relative consumption index of tool life. Tanselet al. !77" used an analytical model of cutting forces and geneticalgorithm to estimate cutting forces, which were used for toolcondition monitoring. Many researchers used such knowledge invarious ways and to various degrees !59,66,74,75,79,80". On theother hand, there are researchers relying only on advanced toolcondition health assessment methods for tool condition monitor-ing, for example, !70".

One of the major challenges for transferring current tool condi-tion monitoring knowledge in research to industrial applications isthe robustness of the models. Ritou et al. !75" retrieved threeprocess-based indicators for tool condition monitoring and con-cluded that they were not reliable during their experiments usingindustrial conditions. Researchers addressed the robustness issuesto various degrees !60,64,65,67,70,71,73,75,80". The approachdemonstrating that the models worked across a range of cuttingconditions carried useful information !60,64,65,67,73". The limi-tations to this approach are limited by the conditions used duringthe experiments. Zhu et al. !71" normalized fault variables andwavelet coefficients to alleviate the effects of possible processvariations on fault diagnosis method. They also did sensitivitystudies for cutting conditions #+ /!20%$ and stock size. Amer atal. !60" used two techniques concurrently to provide some robust-ness against cutting conditions changes: sweeping filters and toothrotation energy estimation. Increasing the depth of cut increasedthe spindle load and average tooth rotation energy but the varia-tions in this energy were transient unless there was a broken tooth.The sweeping filter could guard against false alarm as well be-cause the depth of cut change did not change the frequency spec-tra of the signal. Ritou et al. !75" proposed an indicator indepen-dent of cutting conditions. Shao et al. !80" used a cuttingcondition dependent threshold.

Peng !73" detected a broken insert by a pattern shift of distri-bution of IMFs energies. Many researchers, used threshold-basedapproach where threshold setting/selection was both critical andchallenging. For instance, Xu et al. !79" noted that the power ratioincrease was not consistent for different cutting conditions, leav-ing threshold setting an open challenge. Some authors used avalue for the conducted experiments, for instance !71" without adetailed methodology. Li !67" presented a formula to calculate thefloat threshold. There was one constant in the threshold computa-tion, which was to be selected based on the consequences of fail-ing to detect flute breakage. A value was given for this paperwithout detailed justification. Roth !64" and Suprock and Roth!65" presented a well defined probability-based methodology tocalculate the threshold. It would be interesting to see how wellthat methodology achieved a balance between the model sensitiv-ity and false alarm rate under extended tests, in an industrialsetting.

In industrial applications, a certain level of process variations isexpected and source signals can be contaminated in various wayson a shop floor. For threshold-based approaches without a welldefined robust methodology for threshold setting, it is difficult, ifnot impossible, to apply the models to real applications. These are

Fig. 11 Scatter plot of TFD indicator versus cut conditions andtool wear VB †74‡



challenges that must be addressed to expand industrial applica-tions of tool condition monitoring knowledge. Closer university-industry cooperation may be the key.

5 Tool Condition Monitoring for GrindingGrinding is by far the most important abrasive process because

it plays a prominent role in generating the final surface quality ofmachined parts. Monitoring grinding processes is particularlychallenging because of the large and unknown number of cuttingedges, as well as variable and stochastic cutting geometry. Boththe number of cutting edges and cutting edge geometries varyspatially across the grinding wheel, as well as temporally duringthe grinding process. It is therefore not a surprise that grindingprocess monitoring has been a research topic for several decadesnow, as documented in comprehensive review papers !81,82". Nu-merous research advances have been made in the areas of sensingand hardware, signal processing and feature extraction and toolcondition health assessment approaches in grinding, as will bediscussed in the remainder of this section.

5.1 Advances in Sensing and Hardware. In the recent years,TCM in grinding has seen increased use of high-frequency AEsensors, as well as significant efforts aimed at integration of in situsensors close to the actual cutting process. AE waves propagatethrough structural elements of the machine and workpiece, thusreliably carrying information in the Megahertz frequency domainand giving high dynamic potentials for grinding process monitor-ing !83". Hence, for high precision machining process monitoringaimed at uncovering conditions that affect the surface roughnessand subsurface damage phenomena on the workpiece, which iscrucial for grinding processes, AE shows the highest signal tonoise ratio to the most critical process conditions !84". Theinformation-rich high dynamic content of AE signals was in thesame time an impediment for more widespread use of AE forgrinding #or any machining process monitoring$ because theamount of data generated by an AE sensor during a grinding pro-cess imposes an enormous computational load on the monitoringsystem, even by modern standards. Hence, almost all the grindingprocess monitoring work reported thus far utilizes RMS values ofAE averaged within some moving window, significantly reducingthe amount of data to be processed. AE has been successfully usedin detection of spark and contact in grinding and wheel dimen-sional characterization !84". Advances in computational technol-ogy recently enabled a more frequent and effective use of raw AEsignals. In Ref. !85", six different features extracted from the timeand frequency-domain representations of raw AE signals obtainedfrom a single sensor mounted on the workpiece holder were ex-amined in terms of their sensitivity to thermal damage on thegrinding wheel, formally confirming that the raw AE signal dem-onstrates higher sensitivity to thermal damage of the grindingwheel than the traditional RMS values of the AE signal. Raw AEsignals were also used by Liao et al. !86–88" for the purpose ofclassifying the wheel state into either sharp or dull. Lee et al. !89"discuss the use of AE as a monitoring technique at the precisionscale for a variety of precision manufacturing processes includinggrinding, chemical-mechanical planarization, and ultraprecisiondiamond turning.

In Ref. !90" RMS averaged AE readings were coupled withspindle power readings #obtained from spindle motor currents$,resulting in a grinding monitoring method based on sensor fusion.The method was based on heuristics of using the spindle power tocompensate for sensitivity of AE signals to external factors #suchas sensor assembly, position, workpiece geometry, etc.$ whilehigher dynamic content of AE readings augmented the slow re-sponse characteristics of the power signals.

Improving signal to noise ratio for grinding process monitoringthrough the use of sensors near the grinding zone #i.e., into thegrinding wheel$ was also explored. Furutani et al. !91", proposeda method for in-process measurements for changes in an alumina

grinding wheel topography in cylindrical grinding using a pres-sure sensor placed with a small gap near the grinding wheel. Themeasurement principle is illustrated in Fig. 12, where additionalgrinding fluid is introduced to the gap, and hydrodynamic pressuremonitored.

As the grinding fluid is dragged into the gap between the sensorand the grinding wheel, the hydrodynamic pressure that corre-sponds to the gap length and the wheel topography is measuredon-line. High frequency components of the hydrodynamic pres-sure spectra are found to be related to the wheel loading anddulling. The method presented in Ref. !91" was successfully dem-onstrated only under fixed grinding conditions. In Ref. !92", apiezoelectric sensor was integrated into the grinding wheel, en-abling sensing of forces in grinding as well as in dressing pro-cesses. This relatively direct measurement of the cutting forcesperformed as closely as possible to the cutting area, makes grind-ing and dressing process monitoring more robust to workpiece#material, shape, etc.$ or machining conditions #cooling lubricantsupply, machine set-up parameters, etc.$. In Ref. !93", the conceptfrom Ref. !92" was augmented through integration of a thin filmthermocouple along with miniature force sensors into segmentedgrinding wheels. The concept was implemented in an externalcylindrical grinding operation of bearing rings in the finishing lineof a bearing manufacturer, demonstrating reliability and robust-ness of the new concept.

More traditional sensing has also been used as the basis forTCM in grinding. Hosokawa et al. !94" used sound signals ob-tained from a microphone positioned near the tool-workpiece con-tact to discriminate between 5 different grinding wheel conditionsdenoting progressive states of grinding wheel wear. Grindingforce readings were used in Refs. !95–98", where appeal of rela-tively cheap and nonintrusive force measurements outweighed therelatively low signal to noise ratio of force readings in grindingTCM. Warkentin and Bauer !96" used a dynamometer mounted onthe worktable carrying the workpiece and studied the grindingwheel wear influence on grinding forces. Kwak and Ha !97" useda similar force signal collection system to determine appropriatedressing time based on changes in the force signal readings.Couey et al. !98", proposed a new grinding force sensor based oncapacitance probes integrated into an aerostatic spindle, calibratedto measure grinding forces from changes in the gap between therotor and the stator of the spindle motor. These forces were thenqualitatively demonstrated to be useful for detecting workpiececontact, process monitoring with small depths of cut, detecting

Fig. 12 Measurement principle of †87‡



workpiece defects, and evaluating abrasive wheel wear. In a num-ber of instances, the signal to noise ratio was augmented by theuse of empirical models relating grinding wheel wear with grind-ing forces !95,96", advanced signal processing !97" #wavelet de-noising$, and AI pattern recognition methods !94" #ANN$. Moredetails about the advances in signal processing/feature extractionand tool condition assessment methods in grinding will be dis-cussed in the ensuing text.

5.2 Advances in Signal Processing and Feature ExtractionMethods. Unlike other areas of TCM in machining processes,signal processing and feature extraction based on stationary time-domain and frequency-domain statistics of sensor signals are stilldominant in TCM research in grinding. This can be explained bythe fact that grinding is more complex and less understood thanother machining processes, where advanced, nonstationary signalprocessing techniques have found more applications.

Advances in physics based modeling of the cutting process en-able the use of relatively cheaper #but slower$ force and powersensors, in spite of the poor signal to noise ratio. Such a conceptwas reported in Ref. !95", where detailed model-based simulationsof form grinding processes developed in Refs. !99,100" are usedto establish bounds on the spindle power signals that are charac-teristic of normal processes. Raw time-series of grinding forcereadings were also used in Ref. !96", where grinding wheel con-dition was assessed through tracking of parameters of an empiri-cal model connecting grinding forces with the grinding processparameters and wheel wear. The model used in Ref. !94" was afusion of empirical models from Refs. !101,102", enabling track-ing of wheel wear for both small and large depths of cut. A moreelaborate time-domain feature extraction was reported in Ref.!88", where AR model parameters were used as features for rec-ognition of a sharp or a dull grinding wheel under two differentgrinding conditions. Time-domain features were also used in Ref.!90", where fusion of AE and grinding power readings was ac-complished through an empirical quantity of fast abrasive process,combining AE and spindle power readings.

Frequency-domain features were used in Ref. !94", where theauthors used signal energy in ad hoc selected ranges of thefrequency-domain signal representation #obtained using FFT$ ofgrinding sound to recognize the sharp or worn state of the grind-ing wheel. The authors of Ref. !94" also demonstrated theirmethod under two different grinding conditions. Frequency-domain signal description was also used in Ref. !91" to extract thehigh-frequency components of the hydrodynamic pressure spec-trum, which were found to be related to the wheel loading anddulling. In Ref. !85", six different features extracted from the timeand frequency-domain representations of raw AE signals obtainedfrom a single sensor mounted on the workpiece holder were ex-amined in terms of their sensitivity to thermal damage on thegrinding wheel. A series of tests under various grinding depths ofcut indicated that the constant false alarm statistic Ref. !103", ratioof power, kurtosis, and autocorrelation of the AE signal demon-strate higher sensitivity to thermal damage of the grinding wheelthan the traditional RMS values of the AE signal.

Recent years have also brought about several publications re-porting the use of nonstationary signal analysis for extraction offeatures for TCM in grinding. In Ref. !97", discrete wavelet coef-ficients of force signals obtained during grinding were analyzedjointly with the ground surface roughness parameters, indicatingthat a simple analysis of time-domain evolution of the high scalediscrete wavelet parameters carry information about the timewhen the grinding wheel needed dressing. Nevertheless, such asimple approach was demonstrated only under fixed grinding con-ditions and one may need an elaborate classification technique#such as those discussed in the next subsection$ to deal withanalysis of wavelet coefficients that can cope with significantlyvarying grinding conditions. In Ref. !86", discrete wavelet trans-form was applied to segments of raw AE sensor readings to ex-tract features that could discriminate between a sharp and a worn

tool, leading to a conclusion that higher material removal ratesresult in a higher discriminatory power of signal features thangrinding with lower material removal rates. Similar signal decom-position was used in Ref. !87" with feature vectors being waveletpacket energies of AE signals obtained using various wavelet ba-sis functions. The main difference in the research reported in Refs.!86,87" is in the way signal features were clusterized and “inter-preted” in terms of whether the signal arrived from a sharp tool ora dull tool.

5.3 Advances in Tool Condition Health Assessment andPrediction Methods. Several recent papers report advances in theuse of advanced AI and pattern recognition methods for tool con-dition assessment in grinding !86–88,94". Unlike other areas ofcutting tool condition assessment, advances in grinding tool healthassessment methods seem to have been focused solely on dis-crimination and classification of different grinding wheel condi-tions. This can be understood since the complex and stochasticnature and geometry of the grinding process makes it difficult toclearly define wear measures that can subsequently be estimatedusing sensor signals #in the case of turning, milling, and drilling,one can define, measure and estimate various types of cutting toolwear$.

In Ref. !94", the authors used a three-layer ANN to discriminatebetween several different surface conditions of the grindingwheel. The ANN had input layer of nodes corresponding to eachfeature used for grinding wheel discrimination #sound pressurelevels at different frequencies$, one hidden layer #where the num-ber of nodes was optimized through the training process$, and oneoutput layer of nodes #where the number of nodes was equal to thenumber of conditions that needed to be differentiated$. The neuralnetwork structure is shown in Fig. 13.

Experiments were conducted with a conventional vitrified-bonded alumina grinding wheel and with a resinoid-bonded cubicboron nitride #CBN$ wheel. In the case of conventional vitrified-bonded alumina grinding wheel, five different levels of wheelcondition were introduced through the dressing process and cor-responding signals were used for training of the ANN. Additionalsignals were generated from those five wheel conditions, as wellas from three wheel conditions in between the conditions used fortraining. The problem was further complicated by attempting todiscriminate the various wheel states at different grinding speedsand detailed classification results are reported. In the worst case#one of the states not observed during training$, a 70% correctclassification rate is reported #seven out of ten signals were cor-rectly classified$ while overall classification accuracy was 85%#only twelve signals out of the total of 80 were misclassified$. In

Fig. 13 Neural network architecture for learning system ofgrinding SPLs †90‡



the case of a CBN wheel, three different wheel conditions wereanalyzed #the same three conditions used for training and testingsignals$ and a 100% classification rate is reported.

A series of grinding wheel condition monitoring papers emanat-ing from the collaboration between Louisiana State University andOakridge National Laboratory report interesting results in terms ofadvanced condition assessment methods applied to grinding toolcondition monitoring !86–88". In Ref. !86", the authors used anHMM-based unsupervised clustering method, rather than super-vised classification algorithms, such as the ANN used in Ref. !94".In such a case, one does not have to provide the performanceassessment method with previously identified conditions of thetool and thus human involvement in the training process is re-duced. HMM modeling is used to construct a so-called distancematrix, consisting of pairwise distances between feature se-quences from different signals. Each entry #i , j$ in the matrix isevaluated based on likelihoods that sequence i is generated by theHMM fitted to the sequence j, and that sequence j is generated bythe HMM fitted to sequence i. Five different clustering algorithmsare tested under two different material removal rates for two dif-ferent raisin-bonded diamond wheels. Based on clustering accu-racy, it was observed that discriminatory capabilities increasedwith a high material removal rate #100% clustering accuracy forhigher material removal rate and 75% clustering accuracy for thelower material removal rate$. The very same setup was used inRef. !87", where an adaptive genetic clustering algorithm wasused for clustering purposes. It was observed that the methodachieved 97% clustering accuracy for the data set consisting ofonly high material removal rate relevant signals, 86.7% for thedata set consisting of only low material rate relevant signals, and76.7% clustering accuracy for the data set consisting of a mix ofsignals relevant to both high and low material removal rates. Thisexperiment of autonomously developing a tool condition classifierout of a data set consisting of signals relevant to different cuttingconditions is quite unique and is relevant to creating a system thatcan learn from real operational data, rather than only from dataobtained in a controlled laboratory environment. Most recently, inRef. !88", the same setup was used for grinding wheel conditionmonitoring using “boosted” minimum distance classifiers #MDC$to distinguish between a sharp and a worn grinding wheel, undera set of different cutting conditions. The AdaBoost !104" andA-Boost !105" methods of MDC boosting enabled focusing of theclassifier training procedure on the items that were being difficultto classify while avoiding overfitting through elimination of out-lier training items that could potentially be items with false labelsand that could thus confuse the MDC classifier. The authors reportclassification rates of 89% for high material removal rate data,85% on the low material removal rate data and 76.9% on the dataset consisting of both high and low material removal rates. It iseasy to see that these results were better compared with Ref. !87",which can be attributed to both the use of boosted classifiers, aswell as to the use of multiple sensor readings in Ref. !88".

6 ConclusionsTool condition monitoring is emerging as a viable and useful

tool for CBM, process diagnostics and process control. Lowercost and more accurate sensors facilitate more data indicative ofthe tool condition while increasingly available and distributed ad-vanced computing capabilities enable the application of morecomplex data processing algorithms. Several trends can be ob-served in the latest developments in TCM.

• In all four processes reviewed in this paper, signal process-ing and feature extraction algorithms used for TCM in thelast several years increasingly rely on advanced, nonstation-ary signal based methods, such as various forms of wave-lets, time-frequency and time-scale analysis. Until recently,these methods often could not be implemented on commer-cially available computers while today their benefits of re-

vealing temporally varying patterns of signal energies emit-ted in various frequency bands enable TCM with highersignal to noise ratio.

• Another pervasive trend spurred by increased availability ofcomputing power is more widespread use of computationalintelligence methods for modeling, assessment and diagno-sis of the tool condition. Most notably, one can see a surgein the use of ANNs of various architectures, as well as theuse of kernel based methods, such as SVMs. The use ofthese methods facilitates modeling, assessment and diagno-sis of tool condition with minimal historical or physicalknowledge of the process, which is highly important in flex-ible and reconfigurable manufacturing environments, whereone tool undergoes various cutting regimes #varying feeds,speeds, and cutting geometries$.

• Besides the use of advanced signal processing, feature ex-traction and computational intelligence methods, the latestapproaches to TCM across all four processes reviewed inthis paper achieve success through fusion of multiple sensorreadings, using often very heterogeneous sensors #vibra-tions, sound, acoustic emission and others$. Such sensor fu-sion, accomplished either within advanced physical modelsof the cutting process, or using computational intelligencemethods, enables extraction of additional information com-pared with the information one could obtain by consideringeach sensor reading individually.

• Finally, the last several years have bought about the intro-duction into TCM of several truly predictive methods, suchas hidden Markov models, or recurrent neural networks.These generic predictive methods are able to probabilisti-cally estimate the remaining useful life of the tool. Furtheradvances of predictive capabilities will be crucial for devel-opment of proactive maintenance scheduling techniquesleading to cost-effective tool-replacement practices.

Based on the research surveyed in this paper, the authors feelthat in spite of tremendous recent advances, the area of TCM stillfaces several challenges that need to be overcome.

• Modern manufacturing is increasingly characterized bymass customization, shorter product life-cycles, high prod-uct mixes and consequent need for frequent reconfigurationof the manufacturing processes. In such an environment, thesame machine tool and same cutting tools are used for ma-chining of different features and geometries, thus leading tolimited availability of historical and/or physical knowledgeabout the process. In order to address these situations, thereis an imperative to develop generic #perhaps even “processindependent”$ TCM methods that can rapidly adapt tochanges in the process and merge the useful informationfrom prior use of the tool with the new information availablethrough most recent use of the tool. In light of this need,further advances in signal processing, feature extraction,tool condition modeling, and diagnosis #using both compu-tational intelligence and whatever physical informationabout the process is available$ will be needed.

• Advancement of truly predictive TCM methods is anotherarea that could see significant attention in the future. Real-izing ability to accurately and robustly evaluate probabilitiesof successful cuts in the future, using minimal availabletraining data #unlike most of the existing predictive methodsthat need large amounts of training data$ is an open area inTCM.

• Applicability in real life processes seems to be an everpresent need in the area of TCM. We therefore foresee thatstudies enabling migration of TCM methods onto factoryfloors will continue to receive a lot attention in the future. Inorder to enable one to evaluate applicability of certain meth-ods to actual factory floor processes, future studies shouldpay special attention to evaluating accuracy and robustness



of resulting TCM processes. These properties could beevaluated through testing with data that were not used fortraining of the TCM process #preferably, through testingbased on data obtained using a different cutting tool or ma-chine tool$, as well as through testing based on data ob-tained under different cutting conditions, compared withthose used for training,

• Finally, as the TCM methods advance, so do the manufac-turing processes and equipment. Manufacturing and societalneeds constantly drive the need for more accurate, faster,and often more complicated metal-cutting processes. There-fore, the quest for more robust and reliable TCM methodsyielding high signal to noise ratio in terms of detecting,describing and predicting the cutting tool condition remainsconstantly open.

In summary, we expect that the surge of interest in the area ofTCM driven by new sensing and computational capabilities, aswell as by the increasingly strict demands on productivity andquality in machining will continue in the years to come. TCM willcontinue to be an integral part of intelligent manufacturing, wherehigh quality products are made rapidly, and with minimal wasteenergy and raw materials.

References!1" Jardine, A. K. S., Lin, D., and Banjevic, D., 2006, “A Review on Machinery

Diagnostics and Prognostics Implementing Condition-Based Maintenance,”Mech. Syst. Signal Process., 20#7$, pp. 1483–1510.

!2" Thurston, M., and Lebold, M., 2001, “Standards Development for Condition-Based Maintenance Systems,” Fifty-Fifth Meeting of the Society for Machin-ery Failure Prevention Technology.

!3" Yang, Z., 2005, “Dynamic Maintenance Scheduling Using Online InformationAbout System Condition,” Ph.D. thesis, University of Michigan, Ann Arbor,MI.

!4" Yang, Z., Djurdjanovic, D., and Jun, N., 2007, “Maintenance Scheduling for aManufacturing System of Machines With Adjustable Throughput,” IIE Trans.,39#12$, pp. 1111–1125.

!5" Boutros, T., and Liang, M., 2007, “Mechanical Fault Detection Using FuzzyIndex Fusion,” Int. J. Mach. Tools Manuf., 47#11$, pp. 1702–1714.

!6" Byrne, G., Dornfeld, D., Inasaki, I., Ketteler, G., König, W., and Teti, R., 1995,“Tool Condition Monitoring #TCM$—The Status of Research and IndustrialApplication,” CIRP Ann., 44#2$, pp. 541–567.

!7" Rehorn, A. G., Jiang, J., and Orban, P. E., 2005, “State-of-the-Art Methods andResults in Tool Condition Monitoring: A Review,” Int. J. Adv. Manuf. Tech-nol., 26#7–8$, pp. 693–710.

!8" Sick, B., 2002, “On-Line and Indirect Tool Wear Monitoring in turning WithArtificial Neural Networks: A Review of More Than a Decade of Research,”Mech. Syst. Signal Process., 16, pp. 487–546.

!9" Scheffer, C., and Heyns, P. S., 2001, “Wear Monitoring in Turning OperationsUsing Vibration and Strain Measurements,” Mech. Syst. Signal Process., 15,pp. 1185–1202.

!10" Ghasempoor, A., Moore, T. N., and Jeswiet, J., 2000, “Tool Wear Prediction inTurning,” Proceedings of the 13th International Congress on Condition Moni-toring and Diagnostic Engineering Management, p. 8.

!11" Sun, J., Hong, G. S., Wong, Y. S., Rahman, M., and Wang, Z. G., 2006,“Effective Training Data Selection In Tool Condition Monitoring System,” Int.J. Mach. Tools Manuf., 46#2$, pp. 218–224.

!12" Sun, J., Rahman, M., Wong, Y. S., and Hong, G. S., 2004, “Multiclassificationof Tool Wear With Support Vector Machine by Manufacturing Loss Consider-ation,” Int. J. Mach. Tools Manuf., 44#11$, pp. 1179–1187.

!13" Farrelly, F. A., Petri, A., Pitolli, L., Pontuate, G., Tagliani, A., and Novi In-verardi, P. L., 2004, “Statistical Properties of Acoustic Emission Signals FromMetal Cutting Processes,” J. Acoust. Soc. Am., 116#2$, pp. 981–986.

!14" Li, X., and Du, R., 2004, “Monitoring Machining Processes Based on DiscreteWavelet Transform and Statistical Process Control,” Int. J. Wavelets, Multi-resolut. Inf. Process., 2#3$, pp. 299–311.

!15" Li, X., and Yao, X., 2005, “Multi-Scale Statistical Process Monitoring in Ma-chining,” IEEE Trans. Ind. Electron., 52#3$, pp. 924–927.

!16" Chen, X., and Li, B., 2007, “Acoustic Emission Method for Tool ConditionMonitoring Based on Wavelet Analysis,” Int. J. Adv. Manuf. Technol., 33, pp.968–976.

!17" Gao, H., and Xu, M., 2005, “Intelligent Tool Condition Monitoring System forTurning Operations,” Lect. Notes Comput. Sci., 3498, pp. 883–889.

!18" Jemielniak, K., and Bombinski, S., 2006, “Hierarchical Strategies in Tool WearMonitoring,” Proc. Inst. Mech. Eng., Part B, 220#3$, pp. 375–381.

!19" Salgado, D. R., and Alonso, F. J., 2007, “An Approach Based on Current andSound Signals for In-Process Tool Wear Monitoring,” Int. J. Mach. ToolsManuf., 47#14$, pp. 2140–2152.

!20" Min, B. K., O’Neal, G., Koren, Y., and Pasek, Z., 2002, “Cutting ProcessDiagnostics Utilising a Smart Cutting Tool,” Mech. Syst. Signal Process.,

16#2–3$, pp. 475–486.!21" Achiche, S., Balazinski, M., Baron, L., and Jemielniak, K., 2002, “Tool Wear

Monitoring Using Genetically-Generated Fuzzy Knowledge Bases,” Eng. Ap-plic. Artif. Intell., 15#3–4$, pp. 303–314.

!22" Wang, L., Mostafa, G. M., and Kannatey-Asibu, E., Jr., 2002, “Hidden MarkovModel-Based Tool Wear Monitoring in Turning,” ASME J. Manuf. Sci. Eng.,124#3$, pp. 651–658.

!23" Salgado, D. R., and Alonso, F. J., 2006, “Tool Wear Detection in TurningOperations Using Singular Spectrum Analysis,” J. Mater. Process. Technol.,171#3$, pp. 451–458.

!24" Lu, M.-C., and Kannatey-Asibu, E., Jr., 2002, “Analysis of Sound Signal Gen-eration Due to Flank Wear in Turning,” ASME J. Manuf. Sci. Eng., 124#4$,pp. 799–808.

!25" Scheffer, C., and Heyns, P. S., 2004, “An Industrial Tool Wear MonitoringSystem for Interrupted Turning,” Mech. Syst. Signal Process., 18#5$, pp.1219–1242.

!26" Liu, J., Djurdjanovic, D., Ni, J., Casoetto, N., and Lee, J., 2007, “SimilarityBased Method for Manufacturing Process Performance Prediction and Diag-nosis,” Comput. Ind., 58#6$, pp. 558–566.

!27" Cohen, L., 2000, Time-Frequency and Time-Scale Analysis, Prentice-Hall,Englewood Cliffs, NJ.

!28" Jeong, J., and Williams, W. J., 1992, “Kernel Design for Reduced InterferenceDistributions,” IEEE Trans. Signal Process., 40#2$, pp. 402–412.

!29" Djurdjanovic, D., Lee, J., and Ni, J., 2003, “Watchdog Agent—An Infotronics-Based Prognostics Approach for Product Performance Degradation Assess-ment and Prediction,” Adv. Eng. Inf., 17#3–4$, pp. 109–125.

!30" Cakir, M. C., and Isik, Y., 2005, “Detecting Tool Breakage in Turning AISI1050 Steel Using Coated and Uncoated Cutting Tools,” J. Mater. Process.Technol., 159#2$, pp. 191–198.

!31" Balazinski, M., Czogata, E., Jemielniak, K., and Leski, J., 2002, “Tool Condi-tion Monitoring Using Artificial Intelligence Methods,” Eng. Applic. Artif.Intell., 15#1$, pp. 73–80.

!32" Jemielniak, K., 2006, “Tool Wear Monitoring Based on a Non-MonotonicSignal Feature,” Proc. Inst. Mech. Eng., Part B, 220#2$, pp. 163–170.

!33" Du, R., and Yeung, K., 2006, “Fuzzy Transition Probability: A New Methodfor Monitoring Progressive Faults. Part 2: Application Examples,” Eng. Ap-plic. Artif. Intell., 19#2$, pp. 145–155.

!34" Srinivasa-Rao, C., Nageswara-Rao, D., and Someswara-Rao, R., 2006, “On-line Prediction of Diffusion Wear on the Flank Through Tool Tip Temperaturein Turning Using Artificial Neural Networks,” Proc. Inst. Mech. Eng., Part B,220#12$, pp. 2069–2076.

!35" Oraby, S. E., Al-Modhuf, A. F., and Hayhurst, D. R., 2005, “A DiagnosticApproach for Turning Tool Based on the Dynamic Force Signals,” ASME J.Manuf. Sci. Eng., 127#3$, pp. 463–475.

!36" Scheffer, C., Engelbrecht, H., and Heyns, P., 2005, “A Comparative Evaluationof Neural Networks and Hidden Markov Models for Monitoring Turning ToolWear,” Neural Comput. Appl., 14#4$, pp. 325–336.

!37" Du, R., and Yeung, K., 2004, “Fuzzy Transition Probability: A New Methodfor Monitoring Progressive Faults. Part 1: The Theory,” Eng. Applic. Artif.Intell., 17#5$, pp. 457–467.

!38" Yu, G., Qui, H., Djurdjanovic, D., and Lee, J., 2006, “Feature Signature Pre-diction of a Boring Process Using Neural Network Modeling With ConfidenceBounds,” Int. J. Adv. Manuf. Technol., 30#7–8$, pp. 614–621.

!39" Choi, Y. J., Park, M. S., and Chu, C. N., 2008, “Prediction of Drill FailureUsing Features Extraction in Time and Frequency Domains of Feed MotorCurrent,” Int. J. Mach. Tools Manuf., 48#1$, pp. 29–39.

!40" Subramanian, K., 1977, “Sensing of Drill Wear and Prediction of Drill Life,”ASME J. Eng. Ind., 99, pp. 295–301.

!41" Jantunen, E., 2002, “A Summary of Methods Applied to Tool Condition Moni-toring in Drilling,” Int. J. Mach. Tools Manuf., 42#9$, pp. 997–1010.

!42" Velayudham, A., Krishnamurthy, R., and Soundarapandian, T., 2005, “Acous-tic Emission Based Drill Condition Monitoring During Drilling of Glass/Phenolic Polymeric Composite Using Wavelet Packet Transform,” Mater. Sci.Eng., A, 412#1–2$, pp. 141–145.

!43" Heinemann, R., Hinduja, S., and Barrow, G., 2007, “Use of Process Signals forTool Wear Progression Sensing in Drilling Small Deep Holes,” Int. J. Adv.Manuf. Technol., 33#3–4$, pp. 243–250.

!44" Panda, S. S., Singh, A. K., Chakraborty, D., and Pal, S. K., 2006, “Drill WearMonitoring Using Back Propagation Neural Network,” J. Mater. Process.Technol., 172#2$, pp. 283–290.

!45" Fu, L., Ling, S.-F., and Tseng, C.-H., 2007, “On-Line Breakage Monitoring ofSmall Drills With Input Impedance of Driving Motor,” Mech. Syst. SignalProcess., 21#1$, pp. 457–465.

!46" Ertunc, H. M., and Oysu, C., 2004, “Drill Wear Monitoring Using CuttingForce Signals,” Mechatronics, 14#5$, pp. 533–548.

!47" Choi, Y. J., and Chung, S. C., 2006, “Monitoring of Micro-Drill Wear by Usingthe Machine Vision System,” Trans. NAMRI/SME, 34, pp. 134–150.

!48" Abu-Mahfouz, I., 2003, “Drilling Wear Detection and Classification UsingVibration Signals and Artificial Neural Network,” Int. J. Mach. Tools Manuf.,43#7$, pp. 707–720.

!49" Al-Sulaiman, F. A., Baseer, M. A., and Sheikh, A. K., 2005, “Use of ElectricalPower for Online Monitoring of Tool Condition,” J. Mater. Process. Technol.,166#3$, pp. 364–371.

!50" Brophy, B., Kelly, K., and Byrne, G., 2002, “AI-Based Condition Monitoringof the Drilling Process,” J. Mater. Process. Technol., 124#3$, pp. 305–310.

!51" Franco-Gasca, L. A., Herrera-Ruiz, G., Peniche-Vera, R., Romero-Troncoso,R. D. J., and Leal-Tafolla, W., 2006, “Sensorless Tool Failure Monitoring



System for Drilling Machines,” Int. J. Mach. Tools Manuf., 46#3–4$, pp. 381–386.

!52" Patra, K., Pal, S. K., and Bhattacharyya, K., 2007, “Application of WaveletPacket Analysis in Drill Wear Monitoring,” Mach. Sci. Technol., 11#3$, pp.413–432.

!53" Ertunc, H. M., Loparo, K. A., and Ocak, H., 2001, “Tool Wear ConditionMonitoring in Drilling Operations Using Hidden Markov Models #HMMs$,”Int. J. Mach. Tools Manuf., 41, pp. 1363–1384.

!54" Oh, Y. T., Kwon, W. T., and Chu, C. N., 2004, “Drilling Torque Control UsingSpindle Motor Current and its Effect on Tool Wear,” Int. J. Adv. Manuf.Technol., 24, pp. 327–334.

!55" Al-Sulaiman, F., Sheikh, A., and Baseer, M., 2004, “Empirical Models ofMechanical and Electrical Drilling Power of Mild Steel,” Proc. Inst. Mech.Eng., Part B, 218#9$, pp. 1181–1189.

!56" Newland, D. E., 1999, “Ridge and Phase Identification in the FrequencyAnalysis of Transient Signals by Harmonic Wavelets,” ASME J. Vibr. Acoust.,121#2$, pp. 149–155.

!57" Romberg, T. M., Cassar, A. G., and Harris, R. W., 1984, “A Comparison ofTraditional Fourier and Maximum Entropy Spectral Methods for VibrationAnalysis,” ASME J. Vib., Acoust., Stress, Reliab. Des., 106#1$, pp. 36–39.

!58" Chandrasekharan, V., Kapoor, S. G., and DeVor, R. E., 1995, “A MechanisticApproach to Predicting the Cutting Forces in Drilling: With Application toFiber-Reinforced Composite Materials,” ASME J. Eng. Ind., 117#4$, pp. 559–570.

!59" René de Jesús, R.-T., Herrera-Ruiz, G., Terol-Villalobos, I., and Jáuregui-Correa, J., 2003, “Driver Current Analysis for Sensorless Tool Breakage Moni-toring of CNC Milling Machines,” Int. J. Mach. Tools Manuf., 43#15$, pp.1529–1534.

!60" Amer, W., Grosvenor, R. I., and Prickett, P. W., 2006, “Sweeping Filters andTooth Rotation Energy Estimation #TREE$ Techniques for Machine Tool Con-dition Monitoring,” Int. J. Mach. Tools Manuf., 46#9$, pp. 1045–1052.

!61" Zhang, J., and Chen, J., 2007, “Tool Condition Monitoring in an End-MillingOperation Based on the Vibration Signal Collected Through a Microcontroller-Based Data Acquisition System,” Int. J. Adv. Manuf. Technol., 39#1–2$, pp.445–448.

!62" Kang, E.-G., Park, S.-J., and Lee, S.-J., 2005, “Development of In Situ Systemto Monitor the Machining Process Using a Piezo Load Cell,” Int. J. Adv.Manuf. Technol., 25#7–8$, pp. 647–651.

!63" Dini, G., and Tognazzi, F., 2007, “Tool Condition Monitoring in End MillingUsing a Torque-Based Sensorized Toolholder,” Proc. Inst. Mech. Eng., Part B,221#1$, pp. 11–23.

!64" Roth, J., 2006, “Using the Eigenvalues of Multivariate Spectral Matrices toAchieve Cutting Direction and Sensor Orientation Independence,” ASME J.Manuf. Sci. Eng., 128#1$, pp. 350–354.

!65" Suprock, C., and Roth, J., 2007, “Methods for On-Line Directionally Indepen-dent Failure Prediction of End Milling Cutting Tools,” Mach. Sci. Technol.,11, pp. 1–43.

!66" Bhattacharyya, P., Sengupta, D., and Mukhopadhyay, S., 2007, “CuttingForce-Based Real-Time Estimation of Tool Wear in Face Milling Using aCombination of Signal Processing Techniques,” Mech. Syst. Signal Process.,21#6$, pp. 2665–2683.

!67" Xiaoli, L., 2001, “Detection of Tool Flute Breakage In End Milling UsingFeed-Motor Current Signatures,” IEEE/ASME Trans. Mechatron., 6#4$, pp.491–498.

!68" Dutta, R., Paul, S., and Chattopadhyay, A., 2006, “The Efficacy of BackPropagation Neural Network With Delta Bar Delta Learning in Predicting theWear of Carbide Inserts in Face Milling,” Int. J. Adv. Manuf. Technol., 31#5–6$, pp. 434–442.

!69" Yao, Y., Liu, C., Yuan, Z., and Lu, Y., 2006, “Robustness Improvement of ToolLife Estimation Assisted by a Virtual Manufacturing Cell,” J. Mater. Process.Technol., 172#3$, pp. 445–450.

!70" Fish, R. K., Ostendorf, M., Bernard, G. D., and Castanon, D. A., 2003, “Mul-tilevel Classification of Milling Tool Wear With Confidence Estimation,” IEEETrans. Pattern Anal. Mach. Intell., 25#1$, pp. 75–85.

!71" Zhu, R., DeVor, R. E., and Kapoor, S. G., 2003, “A Model-Based Monitoringand Fault Diagnosis Methodology for Free-Form Surface Machining Process,”ASME J. Manuf. Sci. Eng., 125#3$, pp. 397–404.

!72" Alaniz-Lumbreras, P. D., Gómez-Loenzo, R. A., de Jesús Romero-Troncoso,R., del Rocío Peniche-Vera, R., Jáuregui-Correa, J. C., and Herrera-Ruiz, G.,2006, “Sensorless Detection of Tool Breakage in Milling,” Mach. Sci. Tech-nol., 10, pp. 263–274.

!73" Peng, Y., 2006, “Empirical Model Decomposition Based Time-FrequencyAnalysis for the Effective Detection of Tool Breakage,” ASME J. Manuf. Sci.Eng., 128#1$, pp. 154–166.

!74" Ghosh, N., Ravi, Y. B., Patra, A., Mukhopadhyay, S., Paul, S., Mohanty, A. R.,and Chattopadhyay, A. B., 2007, “Estimation of Tool Wear During CNC Mill-ing Using Neural Network-Based Sensor Fusion,” Mech. Syst. Signal Process.,21#1$, pp. 466–479.

!75" Ritou, M., Garnier, S., Furet, B., and Hascoet, J.-Y., 2006, “A New VersatileIn-Process Monitoring System for Milling,” Int. J. Mach. Tools Manuf.,46#15$, pp. 2026–2035.

!76" Zuperl, U., and Cus, F., 2004, “Tool Cutting Force Modeling in Ball-End

Milling Using Multilevel Perceptron,” J. Mater. Process. Technol., 153–154,pp. 268–275.

!77" Tansel, I. N., Bao, W. Y., Reen, N. S., and Kropas-Hughes, C. V., 2005,“Genetic Tool Monitor #GTM$ for Micro-End-Milling Operations,” Int. J.Mach. Tools Manuf., 45#3$, pp. 293–299.

!78" Kuljanic, E., and Sortino, M., 2005, “TWEM, a Method Based on CuttingForces—Monitoring Tool Wear In Face Milling,” Int. J. Mach. Tools Manuf.,45#1$, pp. 29–34.

!79" Xu, M., Schuyler, C. K., Fussel, B. K., and Jerard, R. B., 2006, “ExperimentalEvaluation of a Smart Machining System for Feedrate Selection and ToolCondition Monitoring,” Trans. NAMRI/SME, 34, pp. 151–158.

!80" Shao, H., Wang, H. L., and Zhao, X. M., 2004, “A Cutting Power Model forTool Wear Monitoring in Milling,” Int. J. Mach. Tools Manuf., 44#14$, pp.1503–1509.

!81" Tönshoff, H. K., Friemuth, T., and Becker, J. C., 2002, “Process Monitoring inGrinding,” CIRP Ann., 51#2$, pp. 551–571.

!82" Karpuschewski, B., and Inasaki, I., 2006, “Monitoring Systems for GrindingProcesses,” Condition Monitoring and Control for Intelligent Manufacturing,Springer, London, p. 83.

!83" Inasaki, I., and Karpuschewski, B., 2001, “Abrasive Processes,” Chapter inSensors in Manufacturing, Wiley, New York, p. 236.

!84" Dornfeld, D. A., Lee, Y., and Chang, A., 2003, “Monitoring of UltraprecisionMachining Processes,” Int. J. Adv. Manuf. Technol., 21#8$, pp. 571–578.

!85" Aguiar, P. R., Serni, P. J. A., Bianchi, E. C., and Dotto, F. R. L., 2004, “In-Process Grinding Monitoring by Acoustic Emission,” IEEE International Con-ference on Acoustics, Speech, and Signal Processing.

!86" Liao, T. W., Hua, G., Qu, J., and Blau, P. J., 2006, “Grinding Wheel ConditionMonitoring With Markov Model-Based Clustering Methods,” Mach. Sci. Tech-nol., 10, pp. 511–538.

!87" Warren Liao, T., Ting, C.-F., Qu, J., and Blau, P. J., 2007, “A Wavelet-BasedMethodology for Grinding Wheel Condition Monitoring,” Int. J. Mach. ToolsManuf., 47#3–4$, pp. 580–592.

!88" Liao, T. W., Tang, F., Qu, J., and Blau, P. J., 2008, “Grinding Wheel ConditionMonitoring With Boosted Minimum Distance Classifiers,” Mech. Syst. SignalProcess., 22#1$, pp. 217–232.

!89" Lee, D. E., Hwang, I., Valente, C. M. O., Oliveira, J. F. G., and Dornfeld, D.A., 2006, “Precision Manufacturing Process Monitoring With Acoustic Emis-sion,” Int. J. Mach. Tools Manuf., 46#2$, pp. 176–188.

!90" Oliveira, J. F. G., and Valente, C. M. O., 2004, “Fast Grinding Process ControlWith AE Modulated Power Signals,” CIRP Ann., 53#1$, pp. 267–270.

!91" Furutani, K., Ohguro, N., Hieu, N. T., and Nakamura, T., 2002, “In-ProcessMeasurement of Topography Change of Grinding Wheel by Using Hydrody-namic Pressure,” Int. J. Mach. Tools Manuf., 42#13$, pp. 1447–1459.

!92" Meyer, L., Heinzel, C., and Brinksmeier, E., 2004, “Monitoring of GrindingProcesses Using a Sensor Equipped Grinding Wheel,” Annals of the WGP.,11#1$, pp. 41–44.

!93" Brinksmeier, E., Heinzel, C., and Meyer, L., 2005, “Development and Appli-cation of a Wheel Based Process Monitoring System in Grinding,” CIRP Ann.,54#1$, pp. 301–304.

!94" Hosokawa, A., Mashimo, K., Yamada, K., and Ueda, T., 2004, “Evaluation ofGrinding Wheel Surface by Means of Grinding Sound Discrimination,” JSMEInt. J., 47#1$, pp. 51–57.

!95" Guo, C., Campomanes, M., Mcintosh, C., Becze, C., and Malkin, S., 2004,“Model-Based Monitoring and Control of Continuous Dress Creep-Feed FormGrinding,” CIRP Ann., 53#1$, pp. 263–266.

!96" Warkentin, A., and Bauer, R., 2003, “Analysis of Wheel Wear Using ForceData in Surface Grinding,” Transaction of the Canadian Society for Mechani-cal Engineering, 27#3$, pp. 193–204.

!97" Kwak, J.-S., and Ha, M.-K., 2004, “Detection of Dressing Time Using theGrinding Force Signal Based on the Discrete Wavelet Decomposition,” Int. J.Adv. Manuf. Technol., 23#1–2$, pp. 87–92.

!98" Couey, J. A., Marsh, E. R., Knapp, B. R., and Vallance, R. R., 2005, “Moni-toring Force in Precision Cylindrical Grinding,” Precis. Eng., 29#3$, pp. 307–314.

!99" Chiu, N., 1993, “Computer Simulation for Form Grinding Process,” Ph.D.dissertation, University of Massachusetts.

!100" Chiu, N., and Malkin, S., 1994, “Computer Simulation for Creep-Feed FormGrinding,” Trans. NAMRI/SME, 22, pp. 119–126.

!101" Malkin, S., 1989, Grinding Technology: Theory and Applications of Machin-ing With Abrasives, Wiley, New York.

!102" Snoeys, R., and Peters, J., 1974, “The Significance of Chip Thickness inGrinding,” CIRP Ann., 23, pp. 227–237.

!103" Nuttall, A. H., 1994, “Detection Performance of Power-Law Processors forRandom Signals of Unknown Location, Structure, Extent, and Strength,”Naval Undersea Warfare Center Division, Naval Undersea Warfare CenterTechnical Report No. 10751.

!104" Freund, Y., and Schapire, R. E., 1997, “A Decision-Theoretic Generalizationof On-Line Learning and an Application to Boosting,” J. Comput. Syst. Sci.,55#1$, pp. 119–139.

!105" Kim, Y., 2003, “Averaged Boosting: A Noise-Robust Ensemble Method,”Advances in Knowledge Discovery and Data Mining, Springer, New York, p.565.



Quality and Inspection of Machining Operations: Tool · monitoring, and the detection and...

Documents

Transcript of Quality and Inspection of Machining Operations: Tool · monitoring, and the detection and...