International Journal of Machine Tools & Manufacture 44 (2004) 759766 www.elsevier.

com/locate/ijmactool

A real-time predictive maintenance system for machine systems

Dheeraj Bansal a,, David J. Evans b, Barrie Jones a
a

Mechanical Engineering Department, School of Engineering and Applied Science, Control and Dynamics Research Group, Aston University, Birmingham B4 7ET, UK b Neural Computing Research Group, Aston University, Birmingham B4 7ET, UK Received 27 October 2003; received in revised form 21 January 2004; accepted 2 February 2004

Abstract This paper describes a novel real-time predictive maintenance system for machine systems based upon a neural network approach. The ability of a neural network to learn non-linear mapping functions has been used for the prediction of machine system parameters using the motion current signature. This approach avoids the need for costly measurement of system parameters. Unlike many neural network based condition monitoring systems, this approach is validated in an o-line proof of concept procedure, using data from an experimental test rig providing conditions typical of those used on production machines. The experiment aims to classify ve distinct motor loads using the motion current signature, irrespective of changing tuning parameters. Comparison of the predicted and actual loads shows good agreement. Generation of data covering all anticipated machine states for neural network training, using a production machine, is impractical, and the use of simulated data, generated by an experimentally validated simulation model, is eective. This paper demonstrates the underlying structure of the developed simulation model. # 2004 Elsevier Ltd. All rights reserved.
Keywords: Neural network; Motion current signature; Machine parameter; Predictive maintenance

1. Introduction Much emphasis has been given to the research and development of high speed machine systems possessing high levels of programmability and recongurability. Motion cycle demands in such machines are high; motors and drives are required to perform under demanding conditions, hence the need for a preventive maintenance system. Modern machines typically use some form of direct current (DC) motor and the process described is based upon such motors. The DC motor is one of the rst machines devised to convert electrical power into mechanical power. Recent improvements in DC machines, and specically the emergence of the brushless motor have lead to its wide usage due to its high torque and small size when compared to induction motors.
Corresponding author. Tel.: +44-121-359-3611; fax: +44-121-3335809. E-mail address: bansald@aston.ac.uk (D. Bansal).

The need for new high-performance motors, with highly sophisticated capabilities, has produced an abundance of new types and sizes of DC motor. Nowadays, DC motors are widely used in many machine applications; with this there is a need for high reliability supported by an eective maintenance system. Recent studies have demonstrated that the technique of predictive maintenance approach can ensure high reliability and performance [1,57,10,16,20]. This paper introduces an eective, real-time, predictive maintenance system based on the motion current signature. The aim of the proposed system is to localize and detect abnormal electrical conditions in order to predict mechanical abnormalities that indicate, or may lead to the failure of the motor [12,13]. Conventional techniques require an accurate mathematical model in order to predict the dynamic behavior in response to a command input [24,8,9,11,15]. This requires that system parameters such as load, backlash and torque are accurately known. Tracking of the system parameter changes during the operation

0890-6955/$ - see front matter # 2004 Elsevier Ltd. All rights reserved. doi:10.1016/j.ijmachtools.2004.02.004

760

D. Bansal et al. / International Journal of Machine Tools & Manufacture 44 (2004) 759766

Nomenclature Kp Ki Kd Pcmd PA Perr Perrp Pint Pdi V Vgain Igain Kv Ka Vcmd VA Verr Verrp Verrint Verrdi Icmd Tcmd JL GT FT JM Kt FTm Ts position proportional gain position integral gain position dierential gain position command actual position position error (current sample) position error (previous sample) position error integral position error dierential shaft velocity velocity proportional gain velocity integral gain velocity feed-forward gain acceleration feed-forward gain velocity command actual velocity velocity error (current sample) velocity error (previous sample) velocity error integral velocity error dierential current command torque command load inertia external torque friction torque motor inertia motor torque constant motor friction torque time slice

necessitates costly instrumentation, which is dicult to justify on a production machine system. The approach presented here allows deviations from the normal dynamic behavior to be predicted accurately based upon the interpretation of the motor current characteristic. The system uses a neural network which learns the non-linear mapping function between system parameters and motion current signature and avoids the need for costly measurement of system parameters. The neural network has been shown to be an eective technique for performing motor fault detection [14,1719]. Successful implementation of neural networks for motor fault detection and diagnosis, up to an accuracy of more than 90%, has been achieved [18]. A fundamental requirement for the successful implementation of a neural network is the availability of relevant, information-rich training data. This application requires machine parameters to be varied to cover all anticipated machine conditions; which ensures that any parameter variation can be interpolated [2123]. While an ideal solution would be to utilize training data

from a real production system, this is impractical for a number of reasons: 1. a large number of sensors would be required to collect data relevant to all machine parameters; 2. machine faults are rare and unlikely to occur; 3. it is impractical to scan the entire range of machine operations. This motivates an alternative approach for developing a system which generates the training data using simulation models. Simulation modelling is a valuable alternative provided that the model is fully validated against a real production machine [13]. The model can be used to generate the training data covering all anticipated machine conditions. The simulation model developed for this system is the result of a joint University and Rockwell Automation research programme. Initially, the primary objective was to develop a tool that would be used to optimize the size and selection conguration for motors, drives and systems [2433]. Enhancements were added to provide a

D. Bansal et al. / International Journal of Machine Tools & Manufacture 44 (2004) 759766

761

capability for simulation and advanced analysis, incorporating models for mechanical aspects such as resonance and backlash. The authors anticipate that the neural network trained using the training data generated from a validated simulation model will have the same level of capability as a neural network trained using the training data from a real machine. Proof of concept is therefore carried out with a simplied experimental setup, which is representative of a production machine axis.

2. Proof of concept The experiment aimed to classify ve distinct motor loads given the motion current signature. The p-gain was varied from 0.1 to 0.3, in steps of 0.01, and v-gain was varied from 5.0 to 15.0, in increments of 1.00. Five readings of motion signature were taken per conguration per p-gain per v-gain. In total, 5000 signatures were collected with a sample size of 900. The data were divided into three groups: training data (1/2 patterns), validation data (1/6 patterns) and testing data (1/3 patterns). All the groups were normalized to zero mean and unit variance. Principal component analysis (PCA) [34] was used for feature extraction. The eigen-values were arranged in descending order, and the plot of the eigen-value against its rank was generated (Fig. 1). This plot can primarily be used to locate the number of principal components where large eigen-values cease and small eigen-values begin [36]. The aim was to identify k principal components whose inclusion returns sucient information. Fig. 1 highlights three such points (4, 10 and 14) where there was a noticeable step in the value of eigen-value. The cumulative contribution of information by the rst

4 PCs to the data was 60% and hence it was not considered against 10 with a contribution of 85% and 14 with a contribution of 96%. The analysis with 14 principal components was adopted. This reduced the dimensionality of data from 900 to 14. A multi-layer perceptron (MLP) with a soft-max activation function, was trained using scaled conjugate gradient (SCG) optimization [34,35]. Early stopping regularization was used to avoid over-tting and number of hidden units was varied from 5 to 50 in steps of 1 in order to access model order. The minimum validation error (obtained by using validation data set on trained neural network) was plotted against the number of hidden units in the MLP (Fig. 2). The plot indicated that error was a minimum for 15 hidden units. Hence, a neural network of 14 inputs, 15 hidden units and ve outputs (1-of-N coding) was used. Each output of the neural network corresponded to a class which represented a load (Table 1). A confusion matrix was used to evaluate the performance of the classier [34,35]. A matrix showing the relation between true and predicted classes was generated. Diagonal entries of the matrix indicate the correct classications and the o diagonal entries show misclassications. The rows of the matrix represent the true classes and the columns represent the predicted classes. The response of the neural network to the test data was obtained and the confusion matrix was plotted (Table 2). The confusion matrix showed 97.59% correct classication for the test data. The o diagonal entries showed that 40 out of 1662 samples were misclassied. A graph of correct classication percentage for a range of loads is shown, Fig. 3. Fig. 3 shows that all the signatures of loads 2 and 4 were correctly classied. This is supported by all zero o diagonal entries in rows 2 and 4 of the confusion matrix.

Fig. 1.

Plot of eigen-values vs. number of PCs.

Fig. 2. Plot of validation error vs. number of hidden units.

762

D. Bansal et al. / International Journal of Machine Tools & Manufacture 44 (2004) 759766

Table 1 The relationship between shaft masses and classes Description No load (shaft only) Small shaft Smooth shaft Big mass shaft Cog shaft Class 1 2 3 4 5 1-of-N coding 10000 01000 00100 00010 00001

Table 2 Confusion matrix using test data Predicted classes 1 Class 1 Class 2 Class 3 Class 4 Class 5 295 0 0 0 0 2 13 346 1 0 0 3 0 0 314 0 0 4 0 0 19 322 7 5 0 0 0 0 345

indicate a borderline case between class 2, 3 and 4. Similarly, the probabilities for all the misclassications were analyzed and were found to be the borderline cases. This justies further work to improve the process. The high value (97.59%) of correct classication indicates that the neural network based approach can use the motion current signature for predicting the machine parameters. However, the concept has been proved by classifying only inertia as a system parameter. Future work lies in identifying other parameters to be classied using the motion current signature. Additionally, the training data requirement increases exponentially as the number of system parameters to be classied increases. This multi-fold training data requirement can be fullled with the use of a simulation model. Thus, the next step was to design the simulation model which can map the system parameters to the motion current signature.

The probabilities of the misclassications of loads 1, 3 and 5 were analyzed and were found to be the borderline cases between the predicted and the true loads. For example, the classier assigns a probability to the sample for each of the ve classes. The sum of all the probabilities is maintained to be 1, due to the use of the soft-max activation function in the MLP. The sample is then classied as a sample of class x, if its probability for class x is maximum. A signicant dierence between the highest and the second-highest value of the probability indicates a satisfactory classication [35]. For instance, one of the samples of class 4 was found to have a probability of 0.07 and 0.93 for class 3 and 4, respectively. This indicates a satisfactory classication of this sample to class 4 due to a large dierence between the probabilities of class 3 (0.07) and class 4 (0.93). However, the sample of class 3, which was misclassied as class 2 (row 3 and column 2 of Table 2), had the probability of 0.38, 0.35 and 0.27 for class 2, 3 and 4, respectively. These values of probability clearly

3. The simulation model The simulation model generates the motor current and velocity characteristics on the basis of system parameters, and is capable of modelling a fault condition which cannot be generated in an on-line environment. As such, the model will be used to provide the mapping of system parameters and the current-time characteristic for all anticipated motor conditions. The mapping will then be used to generate training data for the neural network. 3.1. Block diagram of simulation model The simulation model is of a closed loop form as shown in Fig. 4. It contains three connected PID loops in sequence: a position loop, a velocity loop and a current loop. The input and output of the model are the position command and the motor position, respectively. The position loop calculates position error by subtracting motor position (obtained as feedback) from the position command. The output of the position loop is the velocity command, which is calculated using the position loop gains and the position error. The velocity command is carried forward as an input to the velocity loop. The velocity loop calculates velocity error by subtracting the velocity feedback (obtained by dierentiating the motor position) from the velocity command. The output of the velocity loop is the current command, which is calculated using the velocity loop gains and the velocity error. The current command is carried forward as input to the current loop. The current command is ltered using a notch and a low pass lter within the current loop. The ltered cur-

Fig. 3. Percentage of correct load type classication.

D. Bansal et al. / International Journal of Machine Tools & Manufacture 44 (2004) 759766

763

Fig. 4. Block diagram of the simulation model.

rent command is then used by the amplier to drive the motor. The following section explains the algorithm used in the simulation model. 3.2. Simulation model algorithm Initially, the motion prole is broken into smaller motion proles called the sub-prole, of a time slice duration. The time slice is a small time interval normally equal to the encoder count. Each sub-prole is then fed into the simulation model recursively; the algorithm of the simulation model is based upon the time slice approach. Following; we describe the steps of the simulation model algorithm. 3.2.1. Initial calculations The following set of calculation is required before the position command of the sub-prole can be used in the PID loops. The total inertia (J) of the sub-prole is given by J JM JL 1

duration equal to the time slice and dx/dt is the angular acceleration of the sub-prole. The friction torque, F, of the system can be found by F FT FTm S where S is a constant, and is: 1. equal to 1, when the motion is in the clockwise direction; 2. equal to 1, when the motion is in the anticlockwise direction; 3. equal to 0, when the motor shaft is stationary. FT represents friction torque of the sub-prole and FTm is motor friction torque. Adding acceleration torque, AT, gravitational torque, GT, and friction torque, F, we obtain T GT AT F where T is the total torque of the sub-prole. 3.3. Position loop Please note that all the calculations hereafter are in the time domain and a superscript of t is used to denote it. The position command of the sub-prole is written in the form
t 1 t Pt cmd Pcmd Vff Ts

where JM and JL denote the motor inertia and load inertia, respectively. The total inertia is then used to calculate the acceleration torque (AT) of the sub-prole using AT J dx dt 2

where x is the angular velocity, dt is the sub-prole

764

D. Bansal et al. / International Journal of Machine Tools & Manufacture 44 (2004) 759766

1 in which the parameter Pt cmd is the position command t is the shaft velocity of of the previous sub-prole, Vff the sub-prole and Ts is the time slice. The actual position, Pt A , is then subtracted from the position comt mand, Pt cmd , to calculate the position error, Perr , at time t, so that t t Pt err Pcmd PA

The current command from (13) is passed as an input to the current loop. 3.3.2. Current loop t , is scaled and ltered The current command, Icmd using a notch and a low pass lter. The ltered value of current command is used by the amplier to drive the motor. Finally, the motor position acts as feedback for the analysis of the next sub-prole, thereby, completing the closed loop control system. 3.4. Sample simulation to demonstrate the eectiveness of the model The simulation model is designed to generate the torque command and simulated torque on the same plot. This facilitates visual comparison and error measurement of actual and simulated torque feedback. The model also generates the velocity feedback which is mapped against the motion prole. A sample simulation generated using the model is shown in Fig. 5. The simulation is generated using the motion prole shown in Fig. 6 and machine parameters shown in Table 3. The motion prole (Fig. 6) and the machine parameters (Table 3) can be used to replicate the simulation process by an alternative model for verication purpose.

6 Pt diff ,

The dierential position error, shows the rate of change of position error. Subtracting the position 1 error of the previous sub-prole, Pt err , from the post ition error, Perr , gives
t t 1 Pt diff Perr Perr

7 Pt int , is given by 8

Similarly, the integral position error,

t 1 t Pt int Pint Perr

The position loop is based upon a PID controller with velocity feed-forward gain Kv. The velocity feedforward gain, Kv, generates a velocity command signal proportional to the derivative of the position command, Pt cmd . Using (6)(8) to form a PID equation, the t , takes the form velocity command, Vcmd
t t t Pt Vcmd err Kp Pint Ki Pdiff Kd t Kffv Vff

The velocity command from (9) is carried forward as an input to the velocity loop. 3.3.1. Velocity loop t Within the velocity loop, the actual velocity, VA , is t subtracted from the velocity command, Vcmd , so that
t t t Vcmd VA Verr

10

t where Verr is the velocity error of the sub-prole. Subtracting the velocity error of the previous sub-prole t 1 t (Verr ) from the velocity error (Verr ) gives the dierential velocity error, t t t 1 Verrdiff Verr Verr

11
t Verrint ,

Similarly, the integral velocity error, lated using

t t 1 t Verrint Verr Verrint

is calcu12

Fig. 5.

Motion prole used for sample simulation.

The acceleration feed-forward gain, Ka, generates a t , proportional to the derivative current command, Icmd of the velocity command. Taking the values of the t , the dierential velocity error, velocity error, Verr t t Verrdiff , and the integral velocity error, Verrint , from (10)(12) to form a PID equation, leads to  t t t Verr Vgain Verrint Igain Icmd  t  Vff t Dgain Kffa Verrdiff 13 Ts

Fig. 6. Sample simulation generated using the simulation model.

D. Bansal et al. / International Journal of Machine Tools & Manufacture 44 (2004) 759766 Table 3 Machine parameters for sample test rig simulation Parameter name Motor parameters Motor type Motor inertia Motor torque constant Motor peak torque Motor poles Motor rms torque Motor damping factor Motor friction torque Motor maximum speed Load parameters Load inertia External torque Friction torque Tuning parameters Position proportional gain Position integral gain Velocity proportional gain Velocity integral gain Value H-4075-R 0.00068 0.74 30.0 6 10 0.068 0.14 3000 Units kg m2 Nm A Nm Nm Nm/krpm Nm rpm

765

Ongoing work is now concerned with implementation of the method on a real world production machine.

Acknowledgements The authors would like to thank Dr. Gurdial Singh (Managing Director, Compro Technologies, New Delhi, India), Graham Elvis (Commercial Engineer, Rockwell Automation, Crewe, UK), John Durrant (Former manager, Rockwell Automation, Crewe, UK), Roger Brookes (Former manager, Rockwell Automation, Crewe, UK), and Prof. Seamus D. Garvey (The University of Nottingham, UK) for their time and suggestions. Special thanks to the Motion Projects Team, Compro Technologies, New Delhi, India and Dr. Patrick Littlehales for their invaluable support and inputs. The authors would also like to thank Compro Technologies for their nancial support to the project.

0.0119 0.052 0.321

kg m2 Nm Nm

1000 0 2000 0

1/s 1/ms s 1/s 1/ms s

Fig. 5 suggests that the torque feed-back obtained from the model is very similar to the torque command with an RMS error of 6%. This reinforces the argument that the simulation model is capable of mapping the motion current signature to the system parameters. However, further experiments to validate the simulation model against a real production machine are planned for future work. The validated simulation model can then be used for generating reliable training data for neural network training, validation and testing.

References
[1] F.M. Discenzo, et al., Motor diagnostics: technological drivers leading to 21st century predictive diagnostics, in: Proceedings of the 1997 of International conference of Maintenance and Reliability, vol. 1, Knoxville, TN, pp. 30.0130.12. [2] R.C. Kryter, et al., Condition monitoring of machinery using motor current signature analysis, Sound and Vibration September (1989) 1421. [3] T.W.S. Chow, Condition monitoring of electrical machines using third order spectrum analysis, Conference Record 1996 IEEEIAS Annual Meeting, Lake Buena Vista, FL, vol. 1, 1996, pp. 679686. [4] D.R. Boothman, E.C. Elgar, R.H. Rehder, R.J. Woodall, Thermal trackinga rational approach to motor protection, IEEE Transaction Power Apparatus and System September/October (1974) 13351344. [5] S. Cambrias, S.A. Rittenhouse, Genetic guidelines for life extension of plant electrical equipment, Electric Power Research Institute Report EL-5885 July (1988). [6] R.W. Smeaton, Motor Application and Maintenance Handbook, second ed., McGraw Hill, New York, 1987. [7] P.J. Tavner, J. Penman, Condition Monitoring and Electrical Machines, Research Studies Press Ltd. Wiley, New York, 1987. [8] A. Keyhani, S.M. Miri, Observers for tracking of synchronous machine parameters and detection of incipient faults, IEEE Transaction Energy Conservation 1 (2) (1986), June. [9] A.K. Sood, A.A. Fahs, N.A. Henein, Engine fault analysis, part I: statistical methods, IEEE Transactions Industrial Electronics 32 (1985) 294300, November. [10] J. Penmann, et al., Condition monitoring of electrical drives, Proceedings of Institution Electrical Engineering Part B 133 (3) (1986) 142148, May. [11] P.J. Tavner, et al., An electrical technique for monitoring motor cages, Proceedings of the 1987 IEE International Conference of Electrical Drives, London, UK, 1987, pp. 4346. [12] D. Leith, et al., Condition monitoring of electrical machines by real-time expert system, Proceedings of the 1988 International Conference Electrical, Pisa, Italy, vol. 3, 1988, pp. 297302.

4. Conclusion The use of neural network approach for the condition monitoring system has been validated in an oline proof of concept procedure. It has been shown that the classication of the machine system parameters, on the basis of motion current signature, using a neural network approach is possible. Load inertia on a motor shaft has been classied into dierent categories on the basis of motion current signature. Overall correct classication percentage is found to be 97.59%. Furthermore, it is observed that the classication is not aected by changing tuning parameters and suggests that motion current signature is independent of tuning parameters. The independence of the classication process from the tuning parameters is a useful result since real-time changes in machine parameters are not coupled to the tuning parameter changes.

766

D. Bansal et al. / International Journal of Machine Tools & Manufacture 44 (2004) 759766 [23] D.S. Kim, Y.S. Shin, D.K. Carlson, Carlson, machinery diagnostics for rotating machinery using back-propagation neural network, Proceeding of Third International Machinery Monitoring and Diagnostic Conference, Las Vegas, NV, 1991, pp. 309320. [24] G. Singh, H. Parikh, Stepper motor data management: a structured approach, Proceeding of IMCSD Symposium, 1993. [25] G. Singh, A.C. Leenhouts, SPASM: an integrated task oriented tool for designing step motor systems, in: Proceeding of IMCSD, 2001. [26] P.A. Littlehales, B. Jones, Software scheme for machine design, Proceeding of Drives and Controls Conference, Telford, 1997. [27] P.A. Littlehales, B. Jones, L.A. Smelov, N.J. Leighton, An interactive approach to the selection of servo drive components, Proceedings of Drives and Controls Conference, Telford, 1998. [28] J. Durrant, G. Singh, G. Elvis, J. Martin, A comprehensive sizing and selection program for motion control products, Proceeding of Drives and Controls Conference, 1998. [29] L. Antognini, M. El-Khoury, C. Dayer, L. Kreindler, et al., Design and implementation of high performance motion systems by new software tools, Proceeding of IMCSD Symposium, 1995. [30] A.C. Leenhouts, Prediction and measurement of pullout torque in step motor systems, Proceeding of the IMCSD Symposium, 1995. [31] D. Jones, Using the computers to select industrial brushless DC motors as a function of motor performance characteristics, Proceeding of IMCSD, 1991. [32] Standard Product Catalog, Electro-craft servo systems, 1997. [33] Motion control glossary, Rockwell Automation, June 2000. [34] C.M. Bishop, Neural Networks for Pattern Recognition, Oxford Press, 1995. [35] I.T. Nabney, NETLAB: Algorithm for Pattern Recognition, Springer, 2002. [36] K.V. Mardia, J.T. Kent, J.M. Bibby, Multivariate Analysis, Acamdemic Press, 1979.

[13] R. Isermann, Process fault detection based on modelling and estimation methodsa survey, Automatica 20 (1984) 387 404. [14] M.Y. Chow, R.N. Sharpe, J.C. Hung, On the application and design of articial neural networks for motor fault detection part I, IEEE Transactions Industrial Electronics 40 (1993) 181 188, April. [15] O. Moseler, R. Isermann, Application of model based fault detection to a brushless DC motor, IEEE Transactions Industrial Electronics 47 (2000) 10151020, October. [16] R.G. Herbert, Computer techniques applied to the routine analysis of rundown vibration data for condition monitoring of turbine-alternators, Proceedings of the International Conference on Condition Monitoring, Swansea, UK, 1984, pp. 229242, April. [17] G. Betta, C. Ligouri, A. Pietrosanto, An advanced neural-network based instrument fault detection and isolation scheme, IEEE Transaction on Instrumentation and Measurement 47 (2) (1998) 507512, April. [18] B. Lennox, G.A. Montague, A.M. Frith, C. Gent, V. Bevan, Industrial application of neural networkan investigation, Journal of Process Control 11 (5) (2001) 497507, October. [19] C.-C. Lin, H.-P. Wang, Performance analysis of rotating machinery using enhanced cerebellar model articulation controller (E-CMAC) neural networks, Computers and Industrial Engineering 30 (2) (1996) 227242, April. [20] S.D. Haddad, Condition monitoring and fault diagnosis for predictive maintenance and quality control, Proceedings of Third International Machinery Monitoring and Diagnostic Conference, Las Vegas, NV, 1991, pp. 18. [21] R.P. Lipmann, Pattern classication using neural networks, IEEE Communication Magazine 27 (11) (1989) 4764. [22] G.M. Knapp, H.P. Wang, Machine faults classicationa neural network approach, International Journal of Production Research 30 (1992) 811823.