Event Classification from Sensor Data using Spectral Analysis in RoboticFinishing Processes

Bobby K. Pappachan1, Tegoeh Tjahjowidodo2 and Tomi Wijaya1

1Rolls-Royce @ NTU Corporate Lab c/o, Nanyang Technological University, 65 Nanyang Avenue, Singapore, Singapore2Nanyang Technological University, Singapore, Singapore

{kbobby, twijaya, ttegoeh}@ntu.edu.sg

Keywords: Machining, Finishing, Passes, Welchs Estimate.

Abstract: Process monitoring using indirect methods leverages on the usage of sensors. Using sensors to acquire vitalprocess related information also presents itself with the problem of big data management and analysis. Due touncertainty in the frequency of events occurring, a higher sampling rate is often used in real-time monitoringapplications to increase the chances of capturing and understanding all possible events related to the process.Advanced signal processing methods helps to further decipher meaningful information from the acquireddata. In this research work, power spectrum density (PSD) of sensor data acquired at sampling rates between40 kHz-51.2 kHz was calculated and the co-relation between PSD and completed number of cycles/passes ispresented. Here, the progress in number of cycles/passes is the event this research work intends to classifyand the algorithm used to compute PSD is Welchs estimate method. A comparison between Welchs estimatemethod and statistical methods is also discussed. A clear co-relation was observed using Welchs estimate toclassify the number of cyceles/passes.

1 INTRODUCTION

In machining processes, ensuring the quality of a fin-ished product is crucial and with advances in man-ufacturing technology, a need exists to integrate in-process monitoring technology into the productionenvironment, so as to avoid manufacturing inducedanomalies (G. Byrne and Teti, 1999; D. Dornfeldand Vijayaraghavan, 2009). Advanced process mon-itoring technology coupled with an intelligent deci-sion making support system can reduce the time takento otherwise perform rework on finished componentswith defects. This will save costs and also reducethe dependency on skilled operators. A report re-leased by the Federal Aviation Authority (FAA) inpartnership with the Aerospace Industries Association(AIA) Rotor Manufacturing (RoMan) Project Team inthe year 2006, stresses the importance of incorporat-ing advance process monitoring and control technol-ogy in manufacturing processes especially for criticalaerospace components (Team, 2006). Process mon-itoring is generally classified into direct and indirectmethods. In direct method, the quantity of the outputvariable is measured or monitored directly whereasin indirect method, the output variable is deducedthrough monitoring the quantity of process variables

such as vibration, speed (P. Stavropoulos and Chrys-solouris, 2013). While direct method is known forits accuracy, indirect method is widely accepted sincethey are more realistic to be implemented in an indus-trial environment as the cost incurred is comparativelyless than direct method. Indirect process monitor-ing is performed by capturing these process variableswith the means of sensor systems for e.g., accelerom-eter, dyanamometer, temperature sensor etc. Standarddata acquistion (DAQ) systems are then used to ac-quire, sample and log the data. The data is furtheranalysed to identify any significant and/or persistingtrend/patterns. Subsequently, the analysed data canbe used to deduce the required output variable. Thisanalysis can also be performed real-time making indi-rect method more efficient than direct methods. A im-portant step involved in the data analysis is to identifythe signal signature. This paper focuses on identify-ing a relevant signature that gives direct informationof the progress of machining process. In the sectionsfollowing, an overview of sensor based monitoringand signal processing methods in machining applica-tions is presented. Sections 3 and 4 covers the exper-iment setup under which this research work was per-formed, the results received and inferences deducedfrom the results.

80Pappachan, B., Tjahjowidodo, T. and WIjaya, T.Event Classification from Sensor Data using Spectral Analysis in Robotic Finishing Processes.In Proceedings of the 6th International Conference on Sensor Networks (SENSORNETS 2017), pages 80-86ISBN: 421065/17Copyright c© 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

2 PROCESS MONITORING INMACHINING

Machining is the term used for manufacturing pro-cesses that involves varying range of material re-moval rate (MRR). Machining performed with hardtool/cutter has higher material removal rate comparedwith finishing processes wherein compliant abrasivetools/brushes/belts are used. In a machining processmonitoring system as shown in figure 1, the cuttingregion comprises of several process variables, such asvibrations, cutting forces, acoustic emission, temper-ature, surface finish. The various factors that influ-ence these process variables include the state of thecutter/tool, coolant flow, chip packing and other ma-terial removal process conditions.

Figure 1: Building blocks of machining process monitor-ing(Indirect).

By using appropriate physical sensors, the vari-able that needs to be measured can be continuallymonitored and variations can be logged (R. Teti andDornfeld, 2010). The data acquired is processed withthe aim to identify patterns, trends or abnormal pro-cess conditions. Further analysis is performed on theacquired data with the help of machine learning algo-rithms such as neural networks and fuzzy logic. Upondetection of any process related information or pro-cess faults, the information is communicated eitherto the operator or fed directly to robot controller totake relevant corrective/adaptive actions (R. Teti andDornfeld, 2010; C. Bisu and Cahuc, 2013). A ma-jority of past research works on process monitoring isperformed on processes involving hard tool, for e.g.milling, turning. In most cases the focus of the work isinclined towards tool condition monitoring (G. Byrneand Teti, 1999). Focus of this research work is toidentify information that has some co-relation withthe completed number of passes in a robot assistedfinishing process. This was achieved by analyzing thecorresponding magnitude levels of frequency domain

signal from different passes and belonging to a fixedfrequency band. Further details on the experiment arementioned in section 3.

2.1 Sensing System and SignalProcessing

Some potential measurable process phenomena in arobotic machining environment are shown in figure2. Power and current flow of the spindle deliversthe required cutting force. Hence monitoring thepower intake and current flow in motors that drive thespindle can be used to understand the MRR (G. W.Fritz Klocke, 2008; Pritschow and Kramer, 2005).However, in robot assisted finishing processes, mon-itoring and implementing spindle drive control is im-practical due to complex architecture compared withtraditional milling or turning machines or numericalcontrol (NC) machines. As shown in figure 2, mon-itoring the measurable phenomena which are closerto the machining area is a better alternative to under-stand and analyse the nature of the process. This in-clude acoustic emission (AE), force/torque exerted bythe tool on the workpiece, vibration and spindle mo-tion displacement.

Figure 2: Measurable phenomena in machining environ-ment.

Signal signatures often consist of embedded infor-mation which can be co-related to process variableitself. Signal processing plays a pivotal role in per-forming this task by extracting the relevant signaturesand also to identify trends/patterns. Wide range ofsignal processing techniques exists and choosing anideal technique relies heavily on the type of applica-tion. Acoustic emission (AE) sensor is known for itssusceptibility towards high frequency signals (above20 kHz) and clearly seems to be a favorite choice inmost machining process monitoring applications. Us-ing AE sensor also reduces the requirement to per-form further signature extraction as AE signal has

Event Classification from Sensor Data using Spectral Analysis in Robotic Finishing Processes

81

Table 1: Process monitoring classified based on monitoringvariable. (R. Teti and Dornfeld, 2010).

relatively higher signal to noise (S/N) ratio and im-proved frequency response for high frequencies (Leeand Dornfeld, 1998). However, in applications thatinvolve low material removal rate (MRR), magnitudeof elastic waves produced by tool-workpiece interac-tion is much lower compared to processes like turn-ing, milling, drilling etc. In such cases, other sensingsystems must be relied upon to give meaningful in-formation regarding the process (e.g. Vibration sen-sor). Table 1 gives a summary of past literature ontypes of sensing system and signal processing used,classified based on monitoring aspect. An exhaustivereview on process monitoring including the types ofsensors used and signal processing methods is givenin (R. Teti and Dornfeld, 2010). In our experimentswe have employed a tri-axial accelerometer to capturevibration data.

2.2 Frequency Domain Analysis

Signal processing techniques can be broadly classi-fied as time domain and frequency domain. Sev-eral researches have used both techniques in appli-cations involving tool wear/breakage detection andindirect surface integrity detection. For instance, in(T. I. El-Wardany and Elbestawi, 1996), kurtosis(timedomain) and frequency domain analysis is used suc-cessfully to understand tool properties in a drillingprocess. In time domain analysis, statistical meth-ods are used to distinguish persisting patterns/trends.This includes skewness, kurtosis, co-relation coeffi-cient etc.In frequency domain analysis, captured sig-nal is analyzed in the frequency domain and changesto individual frequency components are often indica-tive of the changes in process variables.Frequency domain analysis also helps to visualize theeffect of noise filtering and various other window-ing and filtering techniques. As the sampling fre-

quencies in this experiment falls in the range of 40-51.2 kHz, performing time domain analysis proved tobe challenging due to the size of data captured andhence frequency domain analysis was effective to un-derstand process characteristics. Signal power alsocontains pertinent information regarding the source ofsignal generation. Conventionally, fast fourier trans-form (FFT) can be used to determine power spec-trum. In stochastic processes, performing FFT willnot be useful to reduce the noise embedded in the sig-nal, hence some averaging needs to be performed toincrease the S/N ratio. Welch’s power spectrum esti-mate essentially calculates power spectrum using FFTcoupled with averaging. This helps to minimise thesignal power caused by random variations. FFT donot account in for discontinuities between successiveperiods as the data captured is assumed to be of a sin-gle period of a periodically repeating waveform andthis phenomenon is referred to as spectal leakage. Ap-plying Welch’s estimate method also helps to reducespectral leakage and reduces the effect caused by un-desired frequencies.Welch’s method to compute PSD is performed by di-viding the time series data into segments that are suc-cessive and averaging the periodograms of each seg-ments or frames. Consider xm(n) to be the input signalwhere m = 0,1, ..,K−1,K = total number of framesand n = 0,1, ..,M− 1, periodogram of m th frame isgiven by,

Pxm,M(k) = 1/M |N−1

∑n=0

xm(n)e−2πnk/N |2 (1)

then Welch’s power spectrum estimate is computedas,

SWx (wk) = 1/K

K−1

∑m=0

Pxm,M(k) (2)

Upon analyzing the power readings of certain fre-quency components, it was noted that the changesobserved was corresponding to the completed num-ber of passes. Analysing frequency component ofthe vibration signal is imperative to finshing processesas the fundamental frequency and its harmonics con-tain coherent information which can be atributed tospindle behavior and also the finshing of the compo-nent. Shop floor operators require systems that areless sophisticated and adopting a frequency domainanalysis method gives that flexbility as opposed toother machine learning algorithms or statistical meth-ods. Welch’s estimate method is preferred as an eas-ier method to implement in such cases as it gives amore visual means of interpretation. However, ma-chine learning algorithms provide more stability inapplications involving predictive maintenance.

SENSORNETS 2017 - 6th International Conference on Sensor Networks

82

3 EXPERIMENT SETUP

The experiment setup (Fig. 3) comprises of an ABBIRB 6660 machining robot and PDS colombo spin-dle. The representative work coupon used for ma-chining is a boss hole of combustor casing and theobjective is to remove the burrs until a chamfer is de-veloped. As mentioned in section 2, vibration signa-tures were measured using a tri-axial accelerometer,Kistler 8763B (IEPE). The RPM of spindle was keptconstant at 10000 RPM and feed rate at 30mm/s. Thedata acquisition devices used was NI cDAQ-9184 andNI 9234 IEPE.

Figure 3: Trial and experiment setup.

3.1 Data Analysis and Results

Data was captured at a sampling rate of 40kHz and forcomputational ease, pre-processed to 1000 samplesper each iteration of Welch’s estimate calculation. Atotal of 12 sets of experiment was conducted with 8being used for offline analysis and co-relation and an-other 4 for validation. After each cycle of machin-ing, the chamfer length was manually measured usinglaser measuring device. In offline data analysis, co-relation between the measured values and variationsin estimated power spectrum is analysed. The co-relation between the power spectrum and number ofcycles was subsequently validated in real-time. Fig-ure 4 shows the process flow of how experiments wereconducted.

Different data analysis techniques were usedon acquired data sets. For instance, kurtosis andskewness values for respective passes were calculatedbut it failed to show any consistent trend or patternwith increasing number of passes. Table 2 shows cal-culated kurtosis and skewness values for each pass. Itcan be noted from the numbers that a pattern or trend

Figure 4: Process flow diagram.

Table 2: Kurtosis and Skewness Values.

is not obvious. This can also be also understood fromthe figures 6 and 7.

Results obtained after performing Welch’s powerspectrum estimate is shown in figure 5. Figures 8, 9and 10 shows the Welch’s power spectrum estimate at1.297Khz. As shown in figures 8, 9 and 10, the powervalues of vibration signal decreases with respect tothe increasing number of passes. The trend here whencompared with kurtosis and skewness values is moreobvious to the naked eye. The decrease in signal mag-nitude is indicative of the strength of the signal. It isalso understood that the decrease in strength of sig-nal is caused due to the smoothening of edges of bosshole with increasing number of passes(Table 3 showsthe increase in chamfer radius with different pass).

From the vibration signatures, it can be concludedthat a co-relation exists with the different number ofpasses and the PSD at 1.297 kHz. This signal is be-lieved to be the 8th harmonic of the fundamental fre-quency generated due to the spindle RPM. The spin-dle RPM is controlled by a variable frequency drive(VFD) controller and the frequency is fixed at 165 Hz.This is however with the exception of Y axis measure-

Table 3: Increase in Chamfer Radius.

Event Classification from Sensor Data using Spectral Analysis in Robotic Finishing Processes

83

Figure 5: Welch’s Power Spectrum Estimate for each Pass/Cycles.

Figure 6: Kurtosis Vs Number of Passes.

ment and is caused due to the orientation and place-ment of the sensor. Figure 11 shows the correlationbetween magnitudes of the captured signal with in-creasing number of passes. From the figure it can beseen that, magnitude of vibration signal in Y axis isrelatively low as compared to X and Z axis measure-ment which further confirms the assumption on sensorlocation aspect.

Figure 7: Skewness Vs Number of Passes.

4 CONCLUSION

The focus of this experiment was to explore the re-lationship between captured vibration signal and theprogress of the actual finishing process. The exper-iment result establishes a linear co-relation betweenvibration signals with completed number of passes.One major limitation which the authors noted duringthis research work is that, the results obtained weredependent on the training experiments and hence ap-

SENSORNETS 2017 - 6th International Conference on Sensor Networks

84

Figure 8: Vibration signature in X direction.

Figure 9: Vibration signature in Y direction.

plying the classification technique to other finishingprocesses may not yield the same expected outcome.This will be the focus of future research directionsof this work; to validate and set up a similar clas-sification technique across other finishing processeslike polishing. Spectral analysis proved to be a vi-able solution for performing this task and is seen asa promising technique to be implemented in real-time applications involving high sampling frequen-cies. The advantage seen here is that, analyzing aparticular frequency component relieves the need ofbulk data processing as opposed to statistical meth-ods wherein packets of data needs to be computed tounderstand the co-relation between different statisti-cal attributes with the number of passes completed.Besides, the welch spectrum estimate showed signifi-cant co-relation with the completed number of passesas opposed to time-domain features like kurtosis andskewness.

This research work was conducted primarily tounderstand possible co-relation between sensor signal

Figure 10: Vibration signature in Z direction.

Figure 11: Signal Magnitude Vs Number of Passes.

features and the progress of the finishing process. Theco-relation observed will be integrated to the roboticfinishing software environment used for tool path pro-gramming and will serve as a visual aid to shop flooroperators enabling them to monitor the progress of thefinishing process. In the next phase of the project, thesignature identified as a classifier will subsequentlybe used as an input parameter for machine learningalgorithms. Additionally, a control strategy could bedeployed with a feedback loop in the robot controlsystem to dynamically adjust the process variables tocompensate for any unexpected behavior.

ACKNOWLEDGEMENTS

The authors would like to thank Mr. Arthur Wee andMr. Muhammad Izzat Roslan from Rolls-Royce PlcSingapore for the support received in preparation ofthis paper and the associated research work.

Event Classification from Sensor Data using Spectral Analysis in Robotic Finishing Processes

85

REFERENCES

C. Bisu, L. Olteanu, R. L. P. D. and Cahuc, O. (2013). Ex-perimental approach on torsor dynamic analysis formilling process monitoring and diagnosis. In Proce-dia CIRP, vol. 12, pp. 73-78.

D. Dornfeld, P. Wright, M. H. and Vijayaraghavan, A.(2009). Enabling manufacturing research through in-teroperability. In ed.

G. Byrne, D. Dornfeld, I. I. G. K. W. K. and Teti, R. (1999).Tool condition monitoring (tcm) the status of researchand industrial application, cirp annals - manufacturingtechnology. In vol. 44, pp. 541-567, // 1995.

G. W. Fritz Klocke, D. V. (2008). Design approach for adap-tive axial force control in gun drilling. In MechanicalSystems and Control vol. 11,. .

Lee, Y. and Dornfeld, D. A. (1998). Application of openarchitecture control system in precision machining. In31st CIRP International Seminar on ManufacturingSystems, CIRP, Berkeley, CA, pp. 436441. .

P. Stavropoulos, D. Chantzis, C. D. A. P. and Chryssolouris,G. (2013). Monitoring and control of manufacturingprocesses: A review. In Procedia CIRP, vol. 8, pp.421-425.

Pritschow, G. and Kramer, C. (2005). Open system archi-tecture for drives. In CIRP Annals - ManufacturingTechnology, vol. 54, pp. 375-378. .

R. Teti, K. Jemielniak, G. O. and Dornfeld, D. (2010). Ad-vanced monitoring of machining operations. In CIRPAnnals - Manufacturing Technology, vol. 59, pp. 717-739.

T. I. El-Wardany, D. G. and Elbestawi, M. A. (1996). Toolcondition monitoring in drilling using vibration sig-nature analysis. In International Journal of MachineTools and Manufacture, vol. 36, pp. 687-711. .

Team, R. P. (2006). Guidelines to minimize manufacturinginduced anomalies in critical rotating parts. In ed.

SENSORNETS 2017 - 6th International Conference on Sensor Networks

86