www.elsevier.com/locate/eswa

Expert Systems with Applications 34 (2008) 291–299

Expert Systemswith Applications

Expert system development for vibration analysis in machinecondition monitoring

Stephan Ebersbach *, Zhongxiao Peng

School of Engineering, James Cook University, Townsville, Qld 4811, Australia

Abstract

Expert systems can be adapted for machine condition monitoring data interpretation due to the ability to identify systematic reason-ing processes. As vibration analysis in condition monitoring is still generally performed by highly trained professionals, the use of expertsystems would allow a greater analysis throughput as well as enabling technicians to perform routine analysis. The development of anexpert system for vibration analysis of fixed plant is discussed, as well as laboratory and industry testing. Unique to existing develop-ments, the expert system incorporates triaxial and demodulated frequency and time domain vibration data analysis algorithms for highaccuracy fault detection. The tests confirm the potential value of the expert system for both laboratory and on-site maintenance depart-ments of large manufacturing and mineral processing plants.� 2006 Published by Elsevier Ltd.

Keywords: Expert system; Vibration analysis; Condition monitoring

1. Introduction

Vibration analysis is a commonly used machine condi-tion monitoring technique for fixed-plant rotating machin-ery, due to relatively fast data collection and interpretationwhen compared to other available off-line techniques. Sincethe data is collected as digitally sampled time domain sig-nals, the vibration analysis technique has allowed furthermanipulation using computers. The development of trans-forms, such as the fast Fourier transform (FFT) (Peng &Chu, 2004), have allowed the conversion of the timedomain data into frequency spectra with ease, as the datawas already stored in a digital format. This contrasts tooil and wear debris analysis techniques, which often relyon extensive chemical analysis (Toms, 1998) and data inter-pretation by experienced/trained analysts.

Artificially intelligent systems have been applied to alarge range of technical problems, in order to automatean otherwise tedious or complex analysis algorithm. Of

0957-4174/$ - see front matter � 2006 Published by Elsevier Ltd.

doi:10.1016/j.eswa.2006.09.029

* Corresponding author. Tel.: +61 4781 5284; fax: +61 4781 4660.E-mail address: Stephan.Ebersbach@jcu.edu.au (S. Ebersbach).

the artificially intelligent methods, expert systems are wellsuited for the vibration analysis technique, as a knownset of rules are used to diagnose various machine faults.The rules were developed by relating the physical fault con-dition to the frequencies emitted by the machine, and ana-lysing the vibration data for the unique fault signatures,which can be high amplitude peaks at a characteristic fre-quency or several frequencies, depending on the fault. Thistype of analysis is typically performed by experiencedmaintenance engineers by manually examining the vibra-tion time histories and frequency domain spectra.

The use of computers for digital signal processing (DSP)has allowed the implementation of filters and signalenhancing calculations to be performed on the vibrationdata for improved noise reduction and signature detection.This technology has enabled vibration analysis to be usedfor monitoring road vehicles, which inherently have a highnoise component in the raw vibration data. Many of theDSP algorithms have been included in the data acquisitionunits, which feature time to frequency domain conversionusing FFT, demodulated spectra acquisition, as well ascoupling with a tachometer to allow the analysis of variablespeed machinery.

Table 1Possible faults of machine components

Machinecomponent

Fault

Roller bearings Cage fault or cage loadingBall or Roller faultRace defectInadequate lubricationInstallation faultBearing loose in housingBearing turning on shaft

Journal bearings Excessive clearance (and looseness)Oil whirlOil whip

Coupling Misalignment

Pump/fan Hydraulic related pumping problem

Spur gears Input & output gear loosenessInput & output gear eccentricityMisalignmentBent shaft (input & output)Backlash or oscillating gearsBroken, cracked chipped or pitted teeth (input &output gear)Gear or pinion fault (due to manufacture ormishandling)Preferential wear

Belt Worn, loose or mismatched beltsBelt/sheave misalignmentEccentric sheavesBelt resonance

General ImbalanceBent drive shaftLooseness

292 S. Ebersbach, Z. Peng / Expert Systems with Applications 34 (2008) 291–299

Despite the use of computers for manipulation of vibra-tion data, the interpretation of the vibration spectra anddiagnosis of machine faults has generally remained thejob of highly trained experts. The difficulty of building aknowledge base from human experts, and implementingthe expert system for a broad range of possible faults is acommon drawback of artificially intelligent systems (Rich& Knight, 1991).

Although artificially intelligent systems have been devel-oped for vibration analysis, they have been developed for aparticular machine component, including rolling elementbearings (Li, Chow, Tipsuwan, & Hung, 2000), and trans-formers (Islam, Wu, & Ledwich, 2000; Wang, Liu, & Grif-fin, 1998). One expert system was recently developed toanalyse frequency domain vibration analysis data for gen-eral machine condition monitoring, using decision tableand decision tree techniques (Yang, Lib, & Chiow Tanc,2005). This development demonstrated that expert systemscan be useful for analysing vibration data and can success-fully diagnose faults of rotating machinery. While thisexpert system is useful in application, expert analysts typi-cally employ triaxial as opposed to single axis frequencydomain analysis, as well as demodulated frequency spectrawhich allows better fault frequency detection in the selectedregion. Time domain analysis is also still commonly usedby analysts to detect faults such as imbalance and geartooth cracks. An expert system that truly incorporatesthe techniques used by maintenance engineers for highaccuracy fault detection of vibration data has not yet beendeveloped.

The expert system development discussed in this paperfocuses on establishing a knowledge base, peak detectionalgorithm and user interface to analyse triaxial frequencydomain, demodulated frequency domain, and time domainvibration data. The objective was to develop an expert sys-tem to analyse vibration data with similar accuracy as anexpert maintenance engineer in an automated softwarepackage allowing high analysis throughput, and hence suit-able for commercial condition monitoring laboratories oron-site use. The ultimate goal is to develop a first artificiallyintelligent system for fault diagnosis and machine condi-tion monitoring using integrated analysis of vibration, oiland wear debris analysis technique.

2. Expert system design

The design of an expert system to analyse vibration con-dition monitoring data was considered the first step in thedevelopment of the integrated artificially intelligent systemusing vibration, oil and wear debris analysis technique. Thedevelopment objectives were to interpret vibration data offixed plant using proven techniques, provide an easy to useinterface for stand-alone operation, and output results in away that can be used for further processing by the compre-hensive analysis expert system, currently in development.

The expert system was developed to be used for highthroughput condition monitoring laboratories of fixed

plant, common in mineral processing and manufacturingindustries. Due to the requirement to operate in a commer-cial environment, the efficient use of human resources is ofprime importance. This was achieved by using proven ana-lysis techniques, allowing operators familiar with manualfault detection to use the expert system with minimal train-ing. Many new vibration analysis techniques operate byblack box methodologies, and do not provide the operatorwith transparent fault detection.

2.1. Knowledge base design

The knowledge base was constructed to detect the faultsassociated with mechanical systems typically used in fixedplant, including roller and journal bearings, spur gears, beltdrives, couplings, and centrifugal pumps. Fault charts foreach component type were constructed using triaxial fre-quency spectra, demodulated spectra and time domaintechniques from handbooks and literature (Taylor, 2003;Neale, 1995), outlining the detection algorithm. The faultsassociated with each component type are summarised inTable 1.

The developed flow charts were discussed with threeindependent experts, working in the vibration analysis con-

S. Ebersbach, Z. Peng / Expert Systems with Applications 34 (2008) 291–299 293

dition monitoring industry. The flow charts were compiledinto one set of reasoning algorithms, and implemented inMicrosoft visual basic software code. Microsoft VisualBasic was selected for implementation due to the require-ments of the user interface, allowing ease of use includingon-line help screens. The implemented flow chart of detect-ing a roller bearing race defect is shown in Fig. 1, includingfault severity assessment.

The expert system incorporates 75 rules, implemented inIf loop type statements, in order to diagnose 54 differentmachine component faults. The pseudo code used to imple-ment the flow chart of Fig. 1 is outlined below:

If (Amplitude of BPFO or BPFI is above Alarm Thresh-old) then

Fig.

If (Amplitude of 2 BPFO or BPFI is above AlarmThreshold) then

If (Amplitude of BPFO or BPFI greater than 2BPFO) then

Race Defect – Moderate-High Defect Severity

Else IfRace Defect – Low-Moderate Defect Severity

End IfEnd If

Else If (Sidebands present on BPFO or BPFI peaks)then

Race Defect

End If

In addition to fault diagnosis, a confidence factor is cal-culated for every detected fault, which is useful for decidingthe major fault in cases where a number of faults have beendetected for one component. The confidence factor is adimensionless variable reflecting how closely the detected

1. Knowledgebase flowchart for bearing race defect diagnosis.

peak of a fault frequency resembles the theoretical fault fre-quency peak, in terms of both frequency and amplitude.The confidence factor is the sum of a frequency and anamplitude factor. The frequency factor is calculated toequal 0 at the frequency deviation limit set by the operator(the allowable variation in frequency of a peak to accountfor measurement inaccuracies), or 0.5 if the peak frequencyequals that of the theoretical fault peak. Peaks in betweenthis frequency range are rated using a linear fuzzy logicoperation. The amplitude factor is calculated similarly,rated 0 at the fault threshold and 0.5 at the severe alarmamplitude threshold, with all peaks in this range beingdetermined using a linear fuzzy logic operation. This prin-ciple is demonstrated in Fig. 2, where the confidence factordisplayed in the expert system results menu is the sum ofthe frequency and amplitude confidence factors. This allo-cates equal analysis significance to frequency andamplitude.

Data and analysis integrity checks have also beenincluded in the expert system, to reduce the likelihood ofanalysis errors occurring. In the event that the loadedvibration data spectra file is of insufficient range to detecthigher frequency fault peaks, an error message is outputto the operator advising to load a data file with higher fre-quency range.

Fig. 2. Principle of confidence factor calculation using linear fuzzy logic.(a) Frequency confidence factor for frequencies where the target frequencyis 1000 Hz, and the allowable frequency deviation is 5 Hz; (b) Amplitudeconfidence factor for amplitudes between Alarm and Severe Alarmthresholds, set at 1.2 and 1.6 mm/s2.

294 S. Ebersbach, Z. Peng / Expert Systems with Applications 34 (2008) 291–299

3. Analysis operation

The analysis operation of the vibration analysis expertsystem (VES) utilises the developed knowledge base forfault identification, as well as a peak detection algorithmwhich is concerned with searching the vibration data filefor a particular frequency, or frequency or time domainpattern. The peak detection algorithm was designed todetect peaks by using amplitude threshold values, and nor-malised peak amplitude based on the largest peak in thespectrum. The threshold detection mode was developedfor general analysis use, and is commonly used by mainte-nance engineers in industry. Amplitude thresholds, alsoreferred to as alarm limits, are typically unique for a certainmachine and failure mode, and while generally being deter-mined by experience, can be estimated using standardssuch as ISO 10816.

The normalised amplitude analysis mode was designedto allow VES to analyse spectra based on the relativeamplitude of fault frequencies, rather than on absoluteamplitude. This analysis mode relates each peak amplitudeby its amplitude ratio relative to the largest peak in thespectra, each peak is therefore normalised into a percent-age of the largest peak in the spectra. This principle is dem-onstrated in Fig. 3. The benefit of this mode is in analysingspectra from machines, which have not had amplitudealarm thresholds set, due to lack of historical and/or man-ufacturer data. Data analysis using this type of peak detec-tion allows fast automated fault detection of a machine,and laboratory testing has proven this technique usefulwhile not being intended for routine machine conditionmonitoring.

The analysis results obtained from the peak detectionand knowledge base can be used to assess the faults presentin a machine, as well as some fault severity assessment.However, these results do not allow the operator to assessthe probability of a detected fault actually occurring. Forthis reason, a quantitative confidence factor has been

Fig. 3. Normalised amplitude peak detection. The low and high normalised al80%, respectively.

included in the analysis algorithm, based on the frequencyand amplitude of the fault frequency compared to the the-oretical frequency and alarm amplitude. The confidencefactor is calculated between 0 and 1, and is especially usefulfor deciding the main fault in circumstances when a num-ber of faults were detected in a component.

4. Interface design

The interface design included the design and develop-ment of data input and analysis interfaces. The analysisinterface was structured so that the analysis processrequires the operator to only select the relevant data filesand pre-defined machine specifications file. This featurealso aids in the vibration analysis expert system to be usedas an expert module in the comprehensive analysis expertsystem, as the only information required by VES for theanalysis is the file name and paths of the relevant data files.When operating as an expert module, VES code will beworking in the background of the master user interface,and the analysis results written to a dedicated text file forfurther processing by the comprehensive analysis expertsystem. A text based results file is also compiled duringthe analysis process which can be used directly as a cus-tomer analysis report when using VES for stand-alonevibration data interpretation.

The machine specifications setup menu has beendesigned to allow an operator to input all machine specificinformation that is required for vibration data analysis.This includes the type and number of components suchas roller bearings, spur gears, and belt reductions, as wellas whether the machine contains a pump and/or coupling,as shown in Fig. 4. Roller bearings require fault frequenciesto be entered in an additional menu, while gear and beltreductions require the gear or pulley diameters to be spec-ified. A provision for entering interference frequencies ofneighbouring equipment has also been included, as this isa common situation in plants where large machines with

arm threshold values are operator defined, and in this case set at 50% and

Fig. 4. Machine specifications setup menu.

S. Ebersbach, Z. Peng / Expert Systems with Applications 34 (2008) 291–299 295

high amplitude vibration are located in close proximity.Once the form is completed, the saved text file is selectedduring the analysis of the machine, and thus allows theanalysis code to diagnose all component faults withoutoperator input.

5. System testing

The VES software was thoroughly tested using vibrationdata obtained from a laboratory single reduction spur geartest rig, as well as a spur gearbox connected to a grain

296 S. Ebersbach, Z. Peng / Expert Systems with Applications 34 (2008) 291–299

auger. The following test conditions were analysed: bentoutput shaft, overload, and contamination. Three tri-axialvibration spectra of 400 Hz, 1000 Hz, and 4000 Hz wereanalysed, as well as a 400 ms time domain file. The sam-pling rates for the vibration spectra were 3200 lines, whilethe time domain file was 4096 lines. Two sets of spectrawere obtained, one on the input shaft of the gearbox,and another at the output shaft.

The alarm amplitude limits were determined from spec-tra taken when the gearbox was overhauled and in goodcondition, and increasing the amplitudes by approximately30%. The gearbox condition was confirmed using oil andwear particle analysis techniques. Each spectrum was thenanalysed using the analysis by normalised amplitude, andamplitude alarm threshold options. For the peak detectionby normalised amplitude, peaks smaller in amplitude than5% of the largest peak in the spectra were disregarded. Thissetting corresponds to high sensitivity fault detection, aseven small peaks are recognised.

The VES detected the bent output shaft condition in alloutput shaft spectra, for both analysis modes. The 1000 Hzhorizontal spectra is shown in Fig. 5. The high amplitudepeak at low frequency causes the remaining spectra toappear quite small. This high amplitude spike at the lowfrequency region of the spectra is believed to be due tooperating limit of the accelerometer used to obtain thedata.

The overloaded operating regime of the laboratory gear-box resulted in the gears showing signs of mild gear loose-ness and backlash, as well as surface fatigue pitting,scuffing and misalignment. The gear looseness and back-lash were detected, as was a low severity input shaft bear-ing race fault, and a medium to severe output shaft race

Fig. 5. 1000 Hz Horizontal acceleration spectra (at outpu

defect. The bearing faults, gear looseness and misalignmentfaults were detected by both analysis modes of VES. Otherdetected bearing faults include a loose fit between the bear-ing, shaft and housings. Inspection of the gear teethrevealed that although the gears were pitted and showedscuffing marks, the teeth profile had not changedsignificantly.

The contamination laboratory test resulted in the gearsbecoming severely worn by a polishing action, causingexcessive looseness and a misalignment secondary fault.The looseness was found to be a combination of the worngear teeth as well as the worn bearings and shafts. The VESanalysis revealed that one or both input and output bear-ings developed ball faults, as well as gear misalignmentand backlash. The horizontal 4000 Hz spectra is shown inFig. 6. Evidence of loose and eccentric gears was alsodetected. The gear eccentricity may have come about dueto preferential wear of the laboratory gearbox, as the gearshave a greatest common divisor higher than 1.

The capability of the expert system to assess the condi-tion of a multistage-reduction gearbox operating in anindustrial environment was determined by analysing thedata obtained from a two-stage reduction spur gearboxoperating in the agricultural industry. The gearbox waspowered by a 0.55 kW four pole flange mount electricmotor, and used to operate a grain auger of 50 mm dia-meter. All four gears were of the spur gear design, withconsecutive reduction ratios of 3.33:1 and 1.44:1, givingan overall reduction of 4.8:1. The intermediate and outputshafts were supported in brass plain bearings, while theinput pinion was mounted directly to the motor shaft.

Data from two gearboxes was collected, in order toobtain amplitude levels of a gearbox in good condition

t gear) of worn spur gears with a bent output shaft.

Fig. 6. 4000 Hz Horizontal acceleration spectra of the contaminated laboratory test, taken at the output shaft.

S. Ebersbach, Z. Peng / Expert Systems with Applications 34 (2008) 291–299 297

and of a gearbox in the wearing out stage. The operatinghours of the two gearboxes were approximately 300 hand 3000 h, respectively. The faults detected by the expertsystem were

• Reduction 1: Loose output gear, misalignment, prefer-ential wear.

• Reduction 2: Possible eccentric pinion and output gears.• Possible lubrication problem of bearings, or other

machine resonance.

Dismantling and visual inspection of the gearbox com-ponents confirmed the loose output gear, misalignmentand preferential wear of reduction 1. The loosening ofthe press fit between the gear and shaft resulted in the gearlooseness and misalignment, as shown in Fig. 7. The loose-ness of the output gear is evident by the rubbing marks ofreduction 2 pinion on the side of the gear. Possible geareccentricity of the reduction 2 gears could not be confirmedby visual inspection, as the gears had not worn sufficientlyto show typical wear marks. The plain bearings were foundto have a worn and scratched surface, which resulted in apolished shaft, as shown in Fig. 8. The bearing wear mayhave been caused by a low oil level, as the gearbox had aleaking input oil seal.

6. Discussion

The developed expert system has achieved the designobjectives to allow high throughput vibration analysis con-dition monitoring to be performed by non-expert technicalstaff. This is a new concept in condition monitoring, tradi-tionally performed by experienced staff in a manual man-ner. Due to the increasing importance of efficient plant

operation from economic challenges, machinery includedin condition monitoring programs is also increasing. TheVES software will allow pre-processing of routine data,enabling experienced staff to focus on monitoring severefaults in critical machinery.

The development of this expert system has focused onallowing technicians to perform vibration data interpreta-tion of common machine components with high throughput,rather than on developing new interpretation techniques invibration analysis. While innovative techniques includingtime spectra (Andrade, Esat, & Badi, 2001), signature detec-tion using various transforms (Matthew, Ma, & Zhang,2005), and probabilistic techniques (Matthew et al., 2005;Chen, Du, & Qu, 1995; Wang, Ismail, & Golnaraghi, 2001;Andrade, Esat, & Badi, 2001) are commonly used, ampli-tude–frequency spectra analysis is still commonly usedby maintenance engineers for vibration data interpretation(http://www.rockwellsoftware.com/emonitorodyssey; http://www.commtest.com/index.cfm). The use of proven indus-try analysis techniques in the analysis algorithm thereforeallows the construction of an expert system with reliable datainterpretation. The use of proven analysis techniques alsoallows operators analysis transparency, while operators areunable to trace analysis conclusions using new techniques.

The VES presents a new development for the use of arti-ficial intelligent systems for vibration data analysis inmachine condition monitoring. Although an expert systemfor vibration analysis has recently been developed (Yanget al., 2005), VES is significantly different by allowing datainterpretation with very limited operator input, as well ashigh fault detection ability using multiple analysis tech-niques. VES has been designed to limit the amount ofinformation required from the operator per machine analy-sis, which has been achieved by using a text file containing

Fig. 7. Wear marks on the output gear of reduction 1. (a) Wear on side ofgear teeth, typical for misalignment, (b) wear extending across whole ofwidth of gear teeth, (c) wear marks of neighbouring pinion gear, indicatinglooseness.

Fig. 8. Scratched surface of brass bush supporting intermediate driveshaft.

298 S. Ebersbach, Z. Peng / Expert Systems with Applications 34 (2008) 291–299

all machine specific information, entered beforehand. Thisis also useful for routine analysis as the information willnot need to be re-entered for subsequent analyses, signifi-cantly increasing throughput of the condition monitoringlaboratory.

Another unique feature of VES is the use of triaxial fre-quency spectra, demodulated spectra and time domain anal-ysis techniques in the developed knowledge base, forimproved fault detection and major fault identification when

numerous faults are found. Major fault identification is alsoaided by the use of a unique confidence factor based onamplitude and detection frequency, and allows the operatorto assess the likelihood that the detected fault is actuallylinked to the frequency signature identified in the vibrationdata. The ability of VES to diagnose faults of numerouscommon components, including bearings, spur gears, belts,and pumps, using triaxial and demodulated frequency andtime domain analysis techniques makes VES a new develop-ment in utilising expert systems for vibration analysis.

The integrated knowledge base and peak detection algo-rithm were tested using both laboratory and industry vibra-tion data, which demonstrated the high fault detectionefficiency and accuracy of the analysis expert system. Theability of assessing the condition other components includ-ing centrifugal pumps and power transmission belts allowsVES to be used for health monitoring of the diverse plantoften found in manufacturing and minerals processingoperations.

It is anticipated that the developed VES expert systemwill be combined with an expert system for oil analysis inorder to provide an integrated artificially intelligent systemfor comprehensive machine fault diagnosis. The develop-ment of VES was an integral part of this research goal.

7. Conclusion

The article outlines the development of an expert systemfor the condition monitoring of fixed plant, using provenindustry analysis methods traditionally performed in amanual manner. The developed vibration analysis expertsystem (VES) has been specifically designed to allow it to

S. Ebersbach, Z. Peng / Expert Systems with Applications 34 (2008) 291–299 299

be integrated into a planned comprehensive analysis expertsystem, which will be used to analyse vibration, oil andwear debris analysis data and provide a single correlatedcondition report. The interface and data handling opera-tions of the software therefore reflect this goal.

The development of the expert system has allowed theverification of vibration analysis techniques commonlyused in industry, including triaxial spectra, time domain,and demodulated spectra, for machine condition monitor-ing. The VES analysis algorithm has been tested using lab-oratory data collected from a single reduction spurgearbox, and a two-stage reduction spur gearbox operatingin the agricultural industry. The robust fault detectionalgorithm of VES successfully identified the gear faults thatoccurred, which included bent output shaft, continuousoverload, contamination, as well as loose and misalignedgears. The successful completion facilitates the design ofa comprehensive machine condition monitoring expert sys-tem utilising oil, wear debris and vibration analysis tech-niques, which is currently in development.

Acknowledgements

Acknowledgment is made to the Australian ResearchCouncil for funding the ARC linkage grant (LP0348873).Acknowledgements also go to Geoff Savage from QNI/BHP Billiton for his help with industry contacts and pro-viding industry data, as well as John Williams of Shell Di-rect, and Paul Schelten of PDS Condition MonitoringServices for their expertise in vibration analysis. The helpprovided by Industrial and Technical Services, in termsof both project funding and feedback used in the knowl-edge base design, was very appreciated. The authors alsowish to thank Nicole Kessissoglou for providing valuedcomments during the editing phase of this paper.

References

Andrade, F. A., Esat, I., & Badi, M. N. M. (2001). Gear conditionmonitoring by a new application of the Kolmogorov-Smirnov test.Proceedings of the Institution of Mechanical Engineers, 215(6), 653–661.

Andrade, F. A., Esat, I., & Badi, M. N. M. (2001). A new approach totime-domain vibration condition monitoring: Gear tooth fatigue crackdetection and identification by the kolmogorov-smirnov test. Journal

of Sound and Vibration, 240(5), 909–919.Chen, Y. D., Du, R., & Qu, L. S. (1995). Fault features of large rotating

machinery and diagnosis using sensor fusion. Journal of Sound and

Vibration, 188(2), 227–242.Islam, S. M., Wu, T., & Ledwich, G. (2000). A novel fuzzy logic approach

to transformer fault diagnosis. IEEE Transactions on Dielectrics and

Electrical Insulation, 7(2), 177–186.Li, B., Chow, M. Y., Tipsuwan, Y., & Hung, J. C. (2000). Neural network

based motor rolling bearing fault diagnosis. IEEE Transactions on

Industrial Electronics, 47(5), 1060–1069.Matthew, J., Ma, L., Zhang, Z. (2005). Some recent advances on condition

monitoring research. In: Proceedings of the twelfth internationalcongress on sound and vibration, Lisbon.

Neale, M. J. (1995). The tribology handbook. USA: ButterworthHeinemann.

Peng, Z. K., & Chu, F. L. (2004). Application of the wavelet transform inmachine condition monitoring and fault diagnostics: A review withbibliography. Mechanical Systems and Signal Processing, 18(2),199–221.

Rich, E., & Knight, K. (1991). Artificial intelligence (second ed.). USA:McGraw Hill.

Taylor, J. I. (2003). The vibration analysis handbook (second ed.). Florida:Vibration Consultants Inc.

Toms, L. A. (1998). Machinery oil analysis – Methods, automation and

benefits (second ed.). Virginia: Coastal Skills Training.Wang, W. Q., Ismail, F., & Golnaraghi, M. F. (2001). Assessment of gear

damage monitoring techniques using vibration measurements.Mechanical Systems and Signal Processing, 15(5), 905–922.

Wang, Z., Liu, Y., & Griffin, P. J. (1998). A combined ANN and expertsystem tool for transformer fault diagnosis. IEEE Transactions on

Power Delivery, 13(4), 1224–1229.Yang, B. S., Lib, B. S., & Chiow Tanc, A. C. (2005). VIBEX: An expert

system for vibration fault diagnosis of rotating machinery usingdecision tree and decision table. Expert Systems with Applications, 28,735–742.