Automatic Fault Diagnosis of Internal Combustion Engine Based on Spectrogram and Artificial Neural Network Sandeep Kumar Yadav Indian Institute of Technology Kanpur Department of Electrical Engineering Kanpur, 208016 India sandeepy@iitk.ac.in Prem Kumar Kalra Indian Institute of Technology Kanpur Department of Electrical Engineering Kanpur, 208016 India kalra@iitk.ac.in Abstract: This paper presents a signal analysis technique for internal combustion (IC) engine fault diagnosis based on the spectrogram and artificial neural network (ANN). Condition monitoring and fault diagnosis of IC engine through acoustic signal analysis is an established technique for detecting early stages of component degradation. The location dependent characteristic fault frequencies make it possible to detect the presence of a fault and to diagnose on what part of the engine the fault is. The difficulty of localized fault detection lies in the fact that the energy of the signature of a faulty engine is spread across a wide frequency band and hence can be easily buried by noise. To solve this problem, the spectrogram for an integrated time frequency pattern extraction of the engine vibration is proposed. The method offers the advantage of good localization of the acoustic signal energy in the time frequency domain. Statistical parameters like, kurtosis, shape factor, crest factor, mean, median, variance etc. are used for feature extraction in time-frequency domain, and artificial neural network (ANN) was employed to identify the faults in IC engine. Experimental results show that the proposed method is very effective. Key Words: Fault Diagnosis, Acoustic Analysis, Internal Combustion Engine, 1 Introduction Several techniques have been reported in the literature for engine fault detection. Based on the type of signals, they can be classified into acoustic signal analysis, temperature measurement, cylinder pressure analysis, electrical current analysis, and vibration measurement [1]. The vibration signatures of normally aspirated engine contain valuable information on the health of combustion chamber components. It cold be used to detect the incipient faults in the engine [2]. Vibration and acoustic emission signals are often used for fault signal diagnosis in mechanical systems, since they carry dynamic information from the mechanical elements [3-7]. These signals normally consists of a combination of the fundamental frequency with a narrow band frequency and the harmonics [8]. Most of these are related to the revolutions of the rotating system since the energy of acoustic and vibration signal is increased when mechanical element is damaged or worn out. An acoustic and vibration based signal analysis techniques used for fault signal analysis include power spectra in time domain or frequency domain, and they can provide an effective technique for machinery diagnosis provided that is the assumption that the signal is stationary [9-13]. The stationary assumption that is wide spread in all time series analysis methods may not hold at the within rotation scale due to variable loading conditions or faults of the signal that is measured during one shaft rotation may be severely nonstationary. From the time frequency characteristics of such a signal, valuable information about machine or component condition can be extracted. There have been several claims that incipient and developing faults in machine components can better be characterized in the time frequency domain [14-15]. Short duration transients effect due to intermittent or transient vibration give rise to sudden and brief changes in signal amplitude or phase; this will hardly be visible in the spectrum, because energy is dispersed in the temporal averaging process. In this study, the spectrogram is used for an integrated time frequency signature extraction for the analysis of different faults in the internal combustion engine. Various statistical features like Kurtosis, Variance, Skewness, Shape factor, Root mean square value, Absolute mean, Zero crossing rate, Maximum Peak, are computed from the information obtained from the spectrogram representation of different faults. These statistical features were used for fault classification using neural network. Back-propagation algorithm was used with feed-forward neural network ISSN: 1790-5117 101 ISBN: 978-960-474-175-5
as a classifier [16, 17]. The live acoustic signals were acquired under seven conditions which includes six fault states and the normal operating state. For the present study, four microphones were placed at different locations of the IC engine. Data for each microphone output was analyzed using proposed Spectrogram-ANN method and finally adopted the majority voting scheme for decision making. Rest of the paper organized as follows. In section 2, the proposed technique for feature extraction and classification has been discussed. Section 3 gives the overview of types of faults and experimental setup. In section 4, performance of the proposed has been tested on real life data sets. Conclusion are given in section 5. 2 Proposed Method The proposed method for automatic fault diagnosis used in this study is shown in Figure 1, where frequency features are extracted using spectrogram. The use of spectrogram for display and visual analysis is preferred due to its positivity and its energy representation characteristics [18]. To get the insight of the effects of time domain signals for different faults are transformed into the time frequency domain via STFT (Spectrogram) [19]. Spectrogram representation for different faults is shown in Figures 2 to 7. Figure 8 shows the frequency band of activity for the studied faults. 2.1 Feature extraction The basic purpose of time-frequency analysis is to define a function which will describe the energy density of a signal simultaneously in time and frequency and is commonly used in applications like speech, sonar, acoustic and vibration signals. STFT is very popular in describing the signal in time-frequency domain and is widely used in non-stationary signal analysis. The concept of STFT is to cut the signal into a suitable overlapping time segments (using windowing method) and do the Fourier analysis of each time segment to find the frequency content in it. The aggregation of such spectra represents how the spectrum is changing in time and is called the spectrogram [20, 21]. STFT of a signal x(t) is written as: ST F T x (τ, ω) = x(t)w (t τ)e jωt dt (1) Where ω denotes the frequency in a window w(t) around t = τ. The spectrogram is the squared magni- Figure 1: Automatic fault diagnosis procedure using spectrogram with neural network tude of the STFT of the signal x(t). P spectrogram,x (τ, ω) = ST F T x (τ, ω) 2. (2) The window w(t) allows the localization of the spectrum in time, but also spreads in frequency in accordance with the uncertainty relationship [22], leading to trade off between time resolution and frequency resolution. The major problem with the application of spectrogram is its non-uniqueness, means any window may be applied to form a valid spectrogram [18]. By seeing the spectrogram representation of different fault signals, we were able to make following inferences: Cam chain noise pattern were seen to appear in the band of 0.7 khz to 2 khz in the form of horizontal line across the spectrogram. Another additional observation made was that as the engine speed was increased, the cam chain signature shifted upwards in the spectrogram. Pattern of primary gear problem appears in the band of 2 khz to 3 khz. This manifests in the form of a continuous horizontal line at around 2.5 khz, which when observed in isolation, produces a whining kind of noise (primary gear ISSN: 1790-5117 102 ISBN: 978-960-474-175-5
Feature number Time-domain features 1. Absolute mean ē = 1 N N i=0 e i 2. Maximum peak value e p = max(e) 3. Root mean square e rms = 1 N N i=1 e2 i 4. Square root value e r = ( 1 N N i=1 ei ) 2 5. Variance var = 1 N N 1 i=1 (e i ē) 2 6. Kurtosis β = 1 N N i=1 e i 4 7. Crest factor C = e peak e rms 8. Shape factor S f = erms ē 9. Standard deviation e std = 1 N Table 1: Statistical features N i=1 (e i ē) 2 Figure 2: Spectrogram Representation of CCN Fault whine), and when small vertical lines emanating from this root are observed, it indicates the presence of gear damage problem. Pattern of cylinder head noise was found to be co-located with that of tappet. In case of severe / high degree manifestation of the problem, their signatures were observed to bear resemblance to that of tappet noise but dispersed /smeared slightly in time axis. In most cases of cylinder head noise, it was seen that there was a distinct activity in the band of 3.5 khz to approximately 4.5 khz, disjoint from the tappet signatures. Figure 3: Spectrogram Representation of CHN Fault Pattern of tappet noise appears in the band of approximately 4 khz to 12 khz and it manifests in two simultaneously occurring patterns. A base at around 4 khz on the spectrogram with equally spaced vertical lines are emanating from this root. The height and thickness of these vertical lines were found to be proportional to the severity of the problem. Pattern of MRN fault is overlapping with in the band of CCN fault, and pattern of PGW fault is overlapping with that of the PGD fault. Frequency spectrum section between 0.2 khz to 6 khz is chosen for investigating the fault patterns, because patterns of the studied faults are very intense in this region. The spectrum or time series signal of each frequency band between 0.2 khz to 6 khz after spectrogram was processed to compute the features like Shape factor, Crest factor, Kurtosis, Variance, Root mean square, Absolute mean, Square root value, Standard deviation, and Maximum peak value. Mathematical formulation of statistical features are given in Table I. Figure 4: Spectrogram Representation of PGD Fault 2.2 Classification A feed-forward neural network (FFNN) was designed and trained on the above mentioned features to classify IC engine condition. 2.2.1 Classification using ANN In this study, the feed-forward neural network, which has got the back propagation algorithm, was used for classification of features for different fault classes. In- ISSN: 1790-5117 103 ISBN: 978-960-474-175-5
Figure 5: Spectrogram Representation of PGW Fault Figure 8: Activity band for different faults 3 Experimental Setup Figure 6: Spectrogram Representation of TAPPET Fault Figure 7: Spectrogram Representation of MRN Fault puts propagate from input layer to output layer via one hidden layer. By trial-and-error method, the number of nodes in the hidden layer is chosen as 12. The output nodes are decided by the number of fault patterns available for the analysis as: Healthy: [1 0 0 0 0 0 0] T ; CCN: [0 1 0 0 0 0 0] T ; CHN: [0 0 1 0 0 0 0] T ; PGD: [0 0 0 1 0 0 0] T ; PGW: [0 0 0 0 1 0 0] T ; Tappet: [0 0 0 0 0 1 0] T ; MRN: [0 0 0 0 0 0 1] T. After the ANN is successfully trained; it would be ready to test the samples to identify their fault classes. Multiple sensory mechanism is designed for data collection process. Different faults occur in different parts of the engine and hence multiple sensory systems are used to capture fault with higher accuracy. Numerically, it was observed that significant improvement can be achieved when using multiple sensor data [23]. This validated our understanding that a faulted system is localized and can be detected by the neighboring sensor. Table 2.4 depicts origin of fault with respect to these four positions. Engine speed was kept to 3000 RPM (2750-3250 RPM). Figure 9 shows complete experimental setup, for condition monitoring implemented for real time testing of engines. In order to verify the effectiveness of the proposed technique, six different faults were seeded in the engine. In view of recording the data for these seeded faults four microphones were placed at four different locations of the engine. Faults and nearest position of accelerometers are given in Table II. LabVIEW 8.0 was used as a platform to communicate with the cdaq-9172 hardware. Using VI s in LabVIEW, an easy to use menu driven user interface was used to record the engine acoustic samples. The sampling rate of the data acquisition system was 50 khz. The seeded faults are described below: 3.1 Tappet noise Tappet is only that part of a rocker arm which makes contact with an intake or exhaust valve stem above the cylinder head of an IC engine [24]. As the cam rotates it creates both a sideways and a downward force on the tappet. Without a tappet (and the cam acting ISSN: 1790-5117 104 ISBN: 978-960-474-175-5
cam chain is under tension, it produces the cam chain noise. All Cam Chain noise faults were seeded by varying the tension on the Cam Chain using the tension adjuster. 3.3 Primary gear damage (PGD) The gear assembly is located within the crank case. It comprises a set of drive gear and driven gears assembly. Drive gear is also called primary gear. Any abnormality in these gears in the form of tooth damage, tooth profile error, eccentric/inclined bore, results in typical impact kind of noise. The PGD faults were seeded individually introducing defects in the drive gear and driven gears. Figure 9: Experimental setup S. No. Fault Nearest Position of Accelerometer 1 CCN Position 1 (P1), P2 and/or P4 2 CHN Position 4 (P4) 3 MRN Position 3 (P3) 4 PGD Position 3 (P3) 5 PGW Position 3 (P3) 6 TAPPET Position 1 and/or 2 Table 2: Fault and nearest position of accelerometers directly on the valve) the sideways force would cause the valve stem to bend. With a tappet the sideways force is transferred to the cylinder head so only the downward force acts on the valve stem. Technically, the nominal distance (clearance) between the tappet surface and the valve s contact surface was maintained by means of an adjustment screw on the rocker arm. Tappet noise appears whenever the tappet clearance is too high. Under ideal setting the inlet and outlet tappet clearance are kept close to 0.07/0.08 mm. Any deviation from these set values results in generation of tappet noise. To seed tappet faults the tappet clearances of both inlet and outlet port were deviated from their ideal settings. The tappet clearances of both inlet and outlet port were set to 0.08, 0.11, 0.13, 0.14, 0.15 and 0.20 (all in mm). 3.2 Cam chain noise (CCN) The cam chain is the element within the engine which transfers the drive from the crank shaft to the cam shaft. The cam chain rides along two riders. The tension on the cam chain can be adjusted by pushing the slider inwards or outwards. This can be done using the tension adjuster, which is accessible. Whenever 3.4 Cylinder head noise (CHN) In an internal combustion engine, the cylinder head sits above the cylinders and consists of a platform containing part of the combustion chamber and the location of the valves and spark plugs. Any noise emanating from the cylinder head which is not produced by the tappet clearance is termed as the cylinder head noise. The component other than the valve and rocker arm assembly within the cylinder head is the cam shaft. Any defect on the cam shaft known as cam-lobe NC tends to produce the cylinder head noise. 3.5 Primary gear whining (PGW) The origin of gear whining is gear mesh (misalignment), which may be design, manufacturing assembly and operation- related cause. The level of perceived noise depends on the dynamic properties of all gear bore components and the interfaces between them. Gear whining signal had the character of a sinusoidal signal and its frequency is velocity dependant, the noise level decreases as the speed decreases. This noise can be captured more efficiently when the engine runs on the gears. 3.6 Magneto rotor The stator, in conjunction with the rotor (its appearance is like brass made hollow cylindrical object, Located under the left engine cover) that covers it, is responsible for the spark, lighting, and charging system. Under the rotor, there are 3 coils; the source, the pulsar, and the lighting/charge coils. The source coil allows a spark to be generated, the pulsar tells our engine when to make a spark at the spark plug, and the lighting generates the required voltage for headlight ISSN: 1790-5117 105 ISBN: 978-960-474-175-5
Category P1 P2 P3 P4 Maj. Vot. Healthy 97.67 93.02 90.69 90.69 93.02 CCN 83.33 76.67 80 83.33 83.33 CHN 86.66 83.33 80 90 86.67 PGD 87.80 82.92 92.68 95.12 87.80 Tappet 90 90 85 92.5 90 PGW 86.04 79.06 83.72 88.37 83.72 MRN 90.47 73.80 80.95 83.33 83.33 Table 3: Classification performance of Spectrogram- ANN approach on vibration data and the charge coil charges the battery. The actual image of the stator part on which rotor rotates is shown in Figure 10. The gaps around the pulsar coil are fixed. a performance of more than 73% for all the positions and various faults. For overall classification a majority voting (M.V) scheme (choosing at least 3 out of 4 outputs), and for this scheme accuracy is lying between 83% to 93%. The experimental results show that the Spectrogram-ANN technique can be used effectively in IC engine diagnosis for various faults through measurement of engine acoustic signals. 5 Conclusion In the present study, a prototype of an expert system for fault diagnosis of single cylinder four stroke IC engine using spectrogram for visualization of the fault patterns and feature extraction of vibration signals for different faults, features classification and fault diagnosis using artificial neural network is presented. The approach improves the conventional method which is performed according to the experience of technician. The conventional method is time consuming and personal wasting approach. The propose approach is automatic and online fault diagnosis system that that can improve the cost of fault diagnosis system and reduce human error. Acknowledgment Figure 10: Stator In the faulty condition this gap decreases and the coil starts rubbing and hence it produces a noise which is known as magneto rotor noise. This noise can also be generated if the pulsar coil s screws are loose, but this is a very rare condition. 4 Results This section presents the classification results for faults of IC engine using Spectrogram-ANN technique. For each of the six faults and normal categories features (as discussed in section 2) were extracted. In all, 600 acoustic signals were collected for different faults out of which 400 acoustic signals were used for training the neural network and features of remaining 200 acoustic signals were used for testing. Classification accuracy corresponding to all four sensors placed at four different positions are shown in Table 3. As summarized in Table 3, all the testing results have The research presented in this paper was supported by Technology Information, Forecasting and Assessment Council (TIFAC), Department of Science and Technology (DST), Government of India, under the project number TIFAC/EE/20070174. References: [1] A. J. C. Sharkey, G. O. Chandroth, Acoustic emission, cylinder pressure and vibration: A nultisensor approach to robust fault diagnosis, IEEE, International Joint Conference on Neural Networks, Como, Itali, vol. 6, pp. 223-228, 2000. [2] G. Chandroth, A.J.C. Sharkey and N.E. Sharkey, Vibration signatures, wavelets and principle component analysis in diesel engine diagnosis, in Proc. of Marine Technology ODRA, Poland, 1999. [3] G. Gelle, M. Colas, C. Serviere, Blind source separation: a tool for rotating machine monitoring by vibration analysis, Journal of Sound and Vibration, vol. 248, pp. 865-885, 2001. [4] K. Shibata, A. Takahashi, T. Shirai, Fault diagnosis of rotating machinery through visualization ISSN: 1790-5117 106 ISBN: 978-960-474-175-5
of sound signals, Mechanical Systems and Signal Processing, vol. 14, pp. 229-241, 2000. [5] Z. Geng, J. Chen, B. Hull, Analysis of engine vibration and design of an applicable diagnosis approach, International Journal of Mechanical Science, vol. 45, pp. 1391-1410, 2003. [6] Q. W. Wang, F. Ismail, F. Golnarghi Assessment of gear damage monitoring techniques using vibration measurements, Mechanical Systems and Signal Processing, vol. 5, pp. 905-922. 2001. [7] H. Zheng, Z. Li, X. Chen, Gear fault diagnosis based on continuous wavelet transform, Mechanical System and Signal Processing, vol. 16, pp. 447-457, 2002. [8] J. D. Wu, J. C. Chen, Continuous wavelet transform technique for fault signal diagnosis of internal combustion engines, Nondistructive Testing and Evaluation International, vol. 39, pp. 304-311, 2006. [9] P. M. Frank, Fault diagnosis in dynamic system using analytical and knowledge-based redundancy: a survey and some new results, IEEE Transaction Automation Control, vol. 16, pp. 459-171, 1990. [10] R. Isermann, Process fault diagnosis based on process model knowledge- Part 1. Principles for fault diagnosis with parameter estimation, ASME Journal of Dynamic Systems Measurement, and Control, vol. 113, pp. 620-626, 1991. [11] R. Isermann, Process fault diagnosis based on process model knowledge- Part 2. Principles for fault diagnosis with parameter estimation, ASME Journal of Dynamic Systems Measurement, and Control, vol. 113, pp. 627-633, 1991. [12] M. Biswas, A. K. Pandey, S. A. Bluni, M. M. Samman, Modified chain-code computer vision techniques for interrogation of vibration signatures for structural fault detection, Journal of Sound and Vibration, vol. 175, pp. 89-104, 1994. [13] K. Shibata, A. Takahashi, T. Shirai, Fault diagnosis of rotating machinery through visualization of sound signals, Mechanical System and Signal Processing, vol. 14, pp. 229-241, 2000. [15] W. J. Staszewski, K. Worden, and G. R. Tomlinson, Time-frequency analysis in gearbox fault detection using the wigner-ville distribution and pattern recognition, Mechanical systems and signal processing, vol. 11, pp. 673-692, 1997. [16] J. D. Wu, C. H. Liu, Investigation of engine fault diagnosis using discrete wavelet transform and neural network, Expert Systems with Applications, vol. 35, pp. 1200-1213, 2008. [17] A. Ngaopitakkul, C. Pothisarn, Discrete wavelet transform and back-propagation neural network algorithm for fault location on single-circuit transmission line, in Proceedings of the 2008 IEEE International Conference on Biomimetics, Bangok, Thailand, February 21-26, 2009. [18] G. Jones, B. Boshash, Window matching in the time-frequency plane and the adaptive spectrogram, in Proc. of IEEE-SP, pp. 87-90, Victoria, Canada, 1992 [19] H. Arabaci, O. Biligin, Neural network classification and diagnosis of broken rotor bar faults by means of short time Fourier transform, in the Proc. of IMECS, vol. I, Hong Kong, 2009. [20] A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete Time Signal Processing, 2nd edition, Prentice Hall, New Jersey, America, 1999. [21] L. Cohen, Time Frequency Analysis: Theory and Applications, Prentice Hall, New Jersey, America, 1994. [22] D. Gabor, Theory of Communication, Journal of IEE, vol.93, pp. 429-457, 1946. [23] G. Yen and K. Lin, Wavelet packet feature extraction for vibration monitoring, IEEE Transaction on Industrial Electronics, vol. 47, pp. 650-667, Jun. 2000. [24] Y. Li, P. W. Tse, Xin. Yang, J, Yang, EMDbased fault diagnosis for abnormal clearance between contacting components in a diesel engine, Mechanical Systems and Signal Processing, vol. 24, pp. 193-210, 2010. [14] P.J. Loughlin and G.D. Bernard., Cohen-Posch (positive) time-frequency distributions and their application to machine vibration analysis, Mechanical systems and signal processing, vol. 11, pp. 561-576, 1997. ISSN: 1790-5117 107 ISBN: 978-960-474-175-5