You are on page 1of 9
vs Industrial fault diagnosis: pneumatic train door case study N Lehrasab, H P B Dassanayake, C Roberts, S Fararooy and € J Goodman’ Department of Electronic, Electrical and Computer En: Abstra University of Birmingham, UK A practical, robust method of fault detection and diagnosis of a class of pneumatic train door commonly found in rapid transit systems is presented. The methodology followed is intended to be applied within a practical system where computation is distributed across a local data network for economic reasons. The health of the system is ascertained by extracting features from the trajectory profiles of the train door. This is incorporated into a low-level fault detection scheme, which relies upon using simple parity equations. Detailed diagnostics are carried out once a fault has been det for this purpose neural network models are utilized, This method of detection and diagnosis is implemented in a distributed architecture resulting in a practical, low-cost industrial solution, It is feasible to integrate the results of the diagnosis proce directly into an operator’s maintenance information system (MIS), thus producing a proactive maintenance regime, Keywortk: fault diagnosis, condition monitoring, train doors NOTATION c polynomial coefficient 9 RBF centres D damping e extension J cost function K spring constant m first-order model coefficient (gradient) or the number of inputs to the RBF M mass n number of data points my number of nodes in the hidden layer oO primary residuals Po initial pressure pt) dynamic pressure Ap average drop in pressure Q total airflow volume r secondary residuals ' temporal throw-time or activation delay 6 stopping time to starting time u force or applied pressure v region velocity x set of design parameters y displacement The MS was received on 12 September 2001 and was accepted after ‘eision for publication on 27 March 2002 “Corresponding author: Deparment of Electronic, Electrical and Computer Engineering, University of Birmingham, Edgbaston. Birming- hham B15 207, UK Ms weights of a linear combiner by positive scalar widths ° exponential function Abbreviations EP embedded processor FDI fault detection and isolation FMEA failure mode effects analysis MIS maintenance information system RBF radial basis funetion SOFM self-organizing feature map STME single-throw mechanical equipment TMS train management system 1 INTRODUCTION Many researchers have successfully applied fault diag. nosis methods in large industries such as power generation, aerospace and chemical processing [1-3] Due to both the safety and financial implications of failure in these industries, it is necessary that fault diagnosis be carried out on-line using expensive dedicated monitoring systems. However, little attention has been paid to more prosaic assets such as simple mechanical systems; examples in the rail industry include automatic doors, point machines and level crossings, This paper presents a practical, complete 176 N LEHRASAB, H P B DASSANAYAKE, C ROBERTS, S FARAROOY AND C J GOODMAN fault detection and isolation (FDI) system for a pneumatic train door mechanism. The method followed is equally applicable to a host of other classes of single- throw mechanical equipment (STME) [4] Even in prosaic assets, if incipient faults can be detected with sufficient notice, significant benefits ean be gained. Early detection and diagnosis of faults prior to failure potentially eliminates or reduces downtime, thus increasing reliability, passenger satisfaction and economy. In addition, the detection of faults leading to wrong-side failures provides an increased level of safety protection. In other industries, the integration of an FDI approach using stand-alone systems as part of the general maintenance management strategy has led to ‘ore cost effective and responsive maintenance [8]. In case studies, such as the train door, it is not economically viable to implement stand-alone diagno- sis systems. Therefore, a distributed architecture is adopted to enable reliable, robust FDI methods to be implemented at a fraction of the cost of multiple stand- alone units, 2 PNEUMATIC TRAIN DOOR CASE STUDY, Train doors are a significant cause of delay to metro and, mainline rail services. The failure of one single door panel on a train at a station often causes delay 10 the failed train as well as any following scheduled trains. It follows that incipient fault detection and diagnosis can be applied to the system and trains can be brought out of service and maintained prior to failure, thus reducing disruption to passenger and fines levied on train operating companies. Each train car has four or more automatic, sliding train doors. Usually these are either pneumatically or electrically driven using the train management system (TMS). The doors are centrally enabled by the driver; a passenger-operated open and close control is also provided at each individual door. Figure 1 shows the pneumatic train door actuation system. For each door panel pair a ‘master’ and ‘slave actuator configuration is used. An electrically activated solenoid valve controls the opening and closing of the door. When requested to open, air from the train car's central compressor flows into the ‘open’ cylinder, while air is exhausted from the ‘close’ cylinder. Each linear cylinder contains a piston head that is attached to the rotary door arm through a spring/damper arrangement. The hydraulic damper provides damping to the door movement, while the spring allows a pushback facility if obstructed. The reverse process is used to close the door: however, the trajectory profiles for opening and closing the doors are not identical, ‘A laboratory train door test rig was used for the purposes of collecting the data used for this paper. The Proc Inst Mech Engrs Vol 216 Part F: J Rail and Rapid Transit ‘Table 1 Pneumatic train door oper ating characteristics Operating characteriie Value Opening time () Closing time (5) Door velocity (is) Stall velocity (mis) Stalling force (N) Pushback (N) rig consisted of an actual train door mechanism, of the pneumatic type, operating a single door panel in a specially built housing. The typical operation character istics of the test rig are shown in Table 1. Currently, scheduled maintenance tasks are carried out on the train door either fortnightly or annually. Table 2 shows an abstract of a sample train door maintenance schedule. If an FDI system capable of determining the presence of an incipient fault is utilized, the frequency of certain tasks can be reduced or eliminated, as previously described by Lehrasab er al (61, In addition to the maintenance checks, primitive throw-time check facilities are currently incorporated into some door mechanisms. These systems utilize the TMS for acquiring and evaluating micro-switeh timings, which give an indication of the door throw-times. However, this does not take into account the current operating conditions of the door (e.g. air pressure) and provides no diagnostic information whatsoever. The current systems are also prone to failure as no redundancy is provided in the sensor scheme. An appropriate methodology for incorporating FDI within the pneumatic train door was based on failure mode and effect analysis (FMEA) worksheets gener~ ated during a reliability-centred maintenance study [7,8]. The outcome of the study was that it provided information regarding the failure modes that may be detected through measuring certain characteristics of the door operation. Angular displacement, airflow and air pressure sensors were selected together with methods for measurement of the door’s operating timings. Table 2 Pneumatic train door maintenance schedule Prequency Maintenance task Fortnightly Annually Passenger pushbuttons check X Door operation times check x Door walling check x Door pushack check x Magnetie doar locks check x Rough operation chesk x Door cock check x ODIO «© IMechE 2002 INDUSTRIAL FAULT DIAGNOSIS: PNEUMATIC TRAIN DOOR CASE STUDY i Bivion SHAT ARM PusHBack air ‘CONNECTION ‘SPRINGS HyoRAULIC DAMPER. a Fig. 1 Poeumatic train door actuation system 24. Characteristic features of the train door In order to detect faults in the system it is necessary to extract features that describe the system behaviour. The features are typically temporal, spectral or abstract Once the features are extracted they are modelled with respect to varying operating conditions. Deviations from the model predictions indicate a drift in the system dynamics, and potentially a fault 2.1.1 Temporal features ‘The temporal throw dynamics (ie. the dynamics of an open or close operation) are directly related to the air pressure supplied to the cylinders. It follows that the air pressure applied is an independent variable while airflow, angular displacement and throw timings are considered to be dependent upon the applied pressure. ‘The throw trajectory of the door is adequately repeatable throughout the input pressure range to allow the derivation of a general model. The throw time is, inversely proportional to the input pressure applied, as can be seen in Fig. 2a ODIO! © EMeshE 2002 sclose: VGNGER ‘coor tever ‘SOLENOD VALVE ASSEMELY SOLENCIO CONNECTOR ELECTRICAL CONNECTOR ‘The temporal activation delay, ie. the delay between the solenoid activation and the physical movement of the door, gives a measure of the start-up torque required. Through analysis of throw times from experi- mental data it was found that the activation delay also has an inversely proportional relationship with respect to the input pressure, as can be seen in Fig. 2b. Both the temporal throw-time and activation delay characteristics are such that they can be modelled adequately by the general equation Hu) =e bx w where 1 is either the temporal throw-time or activation delay, w is the input pressure and x = {x1,x2, x5} is a set of design parameters chosen to minimize a cost function: J ¥ [ive som 45) = (ue) with the number of data points. In equation (1), x5 relates to the minimum throw-time of the door at high pressure (limited by the damper). The minimization ean Proc Insn Mesh Brats Vol 216.Part FJ Rail and Rept Transit v8 N LEHRASAB, H P B DASSANAYAKE, C ROBERTS, S FARAROOY AND C J GOODMAN 36 15 34 Throw-time Me 214 ‘Aetivation Del 32 fis : i i 12 bas pe Cpa d26 4 Ba fe a be worm | f08 = : fos u Door Closing + 7 merce 1 6 30 35 40 45 50 55 60 65 70 75 80 Pressure (pi) (a) Fig. 2 Table 3 Temporal models Type Ste Mode! Activation delay Openin Closing Throw-time Opening 5.150007 5 1.83 Closing 00 LTH be carried out using constrained parametric-optimiza- tion techniques (9} Table 3 shows the optimized temporal model param- eters for a nominal train door operation. It should be noted that the opening and closing parameters are different; this is due to the configuration of the actuator. Through experimental analysis it was found that the models provide a high degree of accuracy (+ 5 per cent over the entire operating range, thus justifying the choice of model equation) and can predict robustly the temporal features for a given operating point. 2.1.2 Spectral features As the door displacement profile, in general, is a smooth transition, the spectral features, derived from a discrete Fourier transformation, are primarily low-frequency to dc. components. However, higher spectral components provide an indication of the signal quality, ie, the noise level in the signal measurements, thus providing method for identifying sensor faults in the system ‘The spectral properties of the angular displacement profile can be used to ascertain sensor faults in the FDI system. This provides diagnostic capabilities for the monitoring system itself while reducing the false alarm rate. Figure 3 shows the failure mode of an intermittent high-resistance connection in the angular displacement Proc Inst Mech Engrs Vol 216 Part F: J Rail and Rapid Transit 06, 30 35 40 45 50 55 60 65 70 75 80 Pressure pe (b) Throw-time versus pressure sensor together with the corresponding power spectral densities acquired at 100Hz. It can be seen that as the severity of the fault increases, the magnitude of the spectral components around the Nyquist frequency (S0Hz) also increases. This method of spectral compo- nent analysis can therefore give an indication of the integrity of the acquired data Furthermore, when a passenger obstructs the door, the lower frequency spectral components are observed to increase, as shown in Fig. 4 2.1.3 Abstract features Various abstract methods can be used to generate additional features that indicate the state of health of the system. Four spatial regions with constant acceleration were found within the open and close velocity profiles. The region boundaries were placed at 0.02, 0.12, 0.5, 1.0 and 1, 18m along the trajectory of the door. The average Velocity within each region was calculated, The first region velocity is referred (0 as the start velocity, while the final region equates to the stall velocity (stopping velocity), Start and stall velocities are critical in ensuring passenger safety. A first-order static polynomial was used to model the region velocity over a varying pressure range. A consistent reduction in pressure during the throw of the door indicates air leaks within the system. The average drop in pressure, Ap, can be modelled using J perar—po ) where fo is the starting time, 1, is the stopping time, p(?) is the dynamic pressure acquired during the throw and Po is the initial pressure; Ap can also be calculated indirectly using the airflow measurements [10] ODIO «© IMechE 2002 INDUSTRIAL FAULT DIAGNOSIS: PNEUMATIC TRAIN DOOR CASE STUDY os 08 0.12 2 Initial Failure east 2 0. oe a 06 0.08| 03 3 04 0.06) a [_Healthy Mode 0.04! — 2 go 02] 2 o za 2 4 6 O08 2 4,6 Time Time Time on 012, 0.02 Nors i > = 2 o4 % 00 Initial Stage non a eae 08} y 006 g 0.06. 0.01 & aos a 0.04) } mal i 0.005 « 2 4060 2 4060 2 40 60 Frequency Frequency Frequency Fig. 3. Spectral properties used to detect sensor faults 0.16 05 \ 04 i Aormst Fos Lower Fraqueney Deviations i \ ‘cma show cbtcon i \\ m oz VV f \ oa cm} \\ prem ae — hi Signatre ‘iow Sigatre ° ° ° 1 2 34 o 1 2 3S 4 5 Frequency (He) Time(s) (a) (b) Fig. 4 Spectral effets of door obstruction 3. FAULT DIAGNOSIS PROCESS The application of FDI to low-cost, multiple assets such as the train door differs to the conventional FDI approach applied to large industrial applications in the following manner: In large industrial applications a stand-alone FDI system has an insignificant cost compared to the asset value, For low-cost assets a stand-alone system is often not economically viable, as the monitoring system cost is as much as the asset itself Most industrial applications o monitor continuous complex pi of FDI are used 10 rocesses. Many low- 180 N LEHRASAB, H P B DASSANAYAKE, C ROBERTS, S FARAROOY AND C J GOODMAN cost assets, such as those found in railway applica- tions, are operated infrequently over a short duration (eg. train door, point machines). Therefore it is not necessary to analyse the system continuously, thus reducing the overall computation overheads. 3. Industrial applications of FDI are usually applied 0 single instances of unique systems. Low-cost, multiple assets are usually distributed over a local geographi- cal area, typically within a 1km radius. In the train door example, itis feasible for up to 100 door panels (c.g. a metro train) to be distributed along a 400m long train (e.g. a passenger express train). Therefore, when implementing an FDI system for low- cost assets it is necessary to employ a distributed approach to acquisition and computation. A generic distributed fault diagnosis architecture for low-cost assets has previously been proposed by Dassanayake et al, [11]. This architecture divides the computational intelligence found in stand-alone systems across a three- level topology linked by data networks. The division of the intelligence results in the necessity for only one high- level computational and storage unit for a large number of distributed assets. This generic specification can be adapted for use on modern trains where train manage- ment systems (TMS) are currently used for control and low-level monitoring (Fig. 5). Each door controller incorporates a fieldbus commu- nication node. The door controller acquires data directly from the asset's transducers and generates Jow-level velocity-based control alarms for preliminary fault detection (Section 3.1), The control alarms together with their corresponding throw data are initially analysed to determine whether the detected fault is associated with the movement of the door (process fault) oF the actual monitoring system itself Gensor fault Upon detection of a process fault the acquired data and control alarm is communicated across the TMS to the local embedded processor (EP) to allow detailed diagnosis to be carried out. The EP also schedules sequential fault detection routed through the TMS to, probe for incipient faults; this process is interrupted if a control alarm is received. Typically. one EP may be used, to oversee the fault detection and diagnosis process of doors from one or two ears. A fault code correlating to a specific failure mode within the asset's FMEA is generated by the EP once a fault is fully diagnosed. The FMEA is developed from @ priori knowledge about the system failures. The FMEA details causes of failures and recommends preventative actions. The EP passes the fault code to the train’s central computer via the TMS. The central computer, in turn, communicates the fault code to the train operating company’s remote, office-based main- tenance information system (MIS). The MIS is able to schedule maintenance of the faulty door when the train is next removed from service or, if the fault is, considered severe, early withdrawal of the train can be planned. 3.1 Fault detection and preliminary classification ‘A simple parity equation based approach is used within the door controller to generate control alarms for fault detection and preliminary classification. The four region velocities [start velocity, intermediate velocities (x 2) and stall velocity] are modelled using first-order polynomials that are used to generate primary residuals: 0 = Haan ~ (mp + er) aus — (map +62) (mp +s) Veaatt — (map + 04) Yar a) where O are primary residuals, vis the region velocity, m and ¢ are the model coefficients and p is the initial pressure. The first-order models are adopted due to their MIS - Maintenance Information System TCC - Train Central Computer TMS - Train Management System EP. - Embedded Processor TDA- Train Door Actuator 1 (on T ae) [RR L oA T | ae ms me ms] rw ee | | eee) Train Divers Cabin Passenger Cariages Fig. § TMS use for distibuted fut diagnosis er ine ne soon eecigreee swears ween nore Fox. © DeAE 202 INDUSTRIAL FAULT DIAGNOSIS: PNEUMATIC TRAIN DOOR CASE STUDY 81 Table 4 Residual responses Presure Residuals sensor fault Ont 1 t o 1 8 ' a simplicity, as well as being easily adaptable for secondary residual generation. The primary residuals can be made insensitive to inaccurate measures of p by using the following transformation: () thus creating a set of secondary residuals, 1). Further- more, a third residual, rs, independent of the displac ent profile can be generated = (co +ep toy +ep') 6) where Q is the total airflow volume and cp... are third- order model coefficients. ‘At higher pressures, the uncertainties in the system, ie. friction, are insignificant; thus the residual responses, are of a lesser magnitude than at lower pressures. To accommodate the variability in the residual responses, with respect to the operating conditions, adaptive thresholds were generated. This. was achieved by implementing a simple fuzzy inference system with the door controller with heuristically set membership functions and rules; e.g. if pressure is high then the residual threshold is low. For an optimum, robust performance a fuzzy model can be extracted from historical data using clustering-based methods [12] OPEN gu 51 Fos Bos Box Boa 3s ‘ 2 48 50 55 Pressure (psi) ° 70-4 Velocity sensor fault Train door fault Airflow sensor fault Table 4 shows the residual responses to different types of fault, Each fault type can be clearly identified by unique binary responses, This fault detection and preliminary classification is thus passed to the EP. 3.2. Detailed diagnosis In order to further classify the detected fault the EP utilizes an accurate RBF neural network [13] to re~ model the train door displacement profiles, as shown in Fig, 6, Such a network can model any non-linear behaviour with a predefined accuracy, However, a trade- off exists between the complexity of the network and the required model precision. Although RBF networks calculate the optimum number of neurons for a given error goal and spread, the number of neurons assigned can be significantly reduced by pre-processing the data, eg. filtering and removing corrupt data, Similarly, an RBE network was used to model the airflow profiles. ‘The response of the RBF network can be expressed by Limywille ~ l-os where my are the weights of the linear combiner, || * || denotes the Euclidean distance, ¢; are the RBF centres, py are the positive scalar widths, m, is the number of nodes in the hidden layer, mis the number of inputs and, @ isan exponential function. By extracting the characteristic features from the actual measurements and the model outputs, symptoms, ln. Effectively, a p xq grid of neurons represents this two-dimensional plane. In this applica- tion, a 13-dimensional classification problem is reduced to & QWo-dimensional problem by selecting winning neurons in the neuron grid structure. By reducing the complexity of the input vector distribution better classification rates can be achieved: Therefore, a quantized residual scheme is used: Proc Inst Mech Engrs Vol 216 Part F: J Rail and Rapid Transit Quantized residuals ded od oo xX 1000 4 fl oo 0 0001 4 3 ro 10000 3 2 io 10000 7 4 io 1 0000 8 OO 0 Ot ot oo 1 0010 we 6 10690000 0 s gee pee LO GO Ce oo 00000 5 2 1, Symptoms are less than the thresholds; therefore quantize the residual value to 0. 2. Symptoms are less than twice the thresholds; there- fore quantize the residual value to 0.25 (a soft 1), represented by X. 3. Symptoms are greater than twice the thresholds therefore quantize the residual value to I (a hard 1). Two different SOFMs with grids of 3x4 and 4x4 neurons were used to classify the quantized input symptom vectors. The results after 1000 epochs are shown in Table 6. In a test of 50 faults, 40 were classified accurately, as shown in Table 7. Exposure of the SOFM to a larger set of training data would improve the classification Table 7 Results of fault classification Desription Presented Wentifed Process faults 15, 0 66 Obstructions 10 8 80 Now-citical faults 10 9 90 Sensor faults 1s B x7 Total faults 50 40 80 ODIO «© IMechE 2002 INDUSTRIAL FAULT DIAGNOSIS: PNEUMATIC TRAIN DOOR CASE STUDY 83 performance. Furthermore, it is noted that a Neuro- Fuzzy system, which performs fault mapping as well as providing qualitative knowledge, may add strength to the application in future work [15] 4 CONCLUSION The paper demonstrates that by distributing the intelligence throughout the system low-cost assets can be economically monitored while the robustness and integrity of the fault detection and diagnosis is retained. Although the case study within the paper addresses the diagnosis of pneumatic train doors, the techniques can be extended to other subsystems in the train as well as track side assets. Low-level fault detection was carried out using abstract velocity models. Accurate modelling of the velocity and airflow profiles was achieved by the use of RBE neural networks. Temporal, spectral and abstract, features were extracted from the modelled profiles and compared with actual measurements to generate symp- toms. An SOFM was used to classify the symptoms. Preliminary results show that a fault diagnostic accuracy of 80 per cent was achievable. REFERENCES 1 Weerasinghe, M., Gomm, J. and Williams D, Neural networks for fault diagnosis of a nuclear fuel processi plant at different operating points. Control Engng Practice, 1998, 6(2). 281-289. 2 Patton, R.J. Fault detection and diagnosis in aerospace systems using analytical redundaney. [EE Computing and Control J.. 1991, 203), 127-136. 3 Bissessur, ¥., Martin, E:B. and Morris, A.J. Monitoring the performance of the paper making process. Control Engng Practice, 1999, 7(11), 1387-1368, ODIO! © EMeshE 2002 4 Lehrasab, N. and Fararvoy, S. Formal definition of single throw mechanical equipment for fault diagnosis. JEE Electronics Lett, 1998, 3423), 2231-2232, 5 Trave-Massuyes, L. and Milne, R. Gas-turbine condition ‘monitoring using qualitative model-based diagnosis. IEEE Expert Incelligent Systems and Applic., 1997, 12(3), 22-31 6 Lehrasab, N., Fararooy, S. and Allan, J. Intelligent carly failure warning sensor systems for train rotary door operators. In IEE Colloquium on Inrelligent Sensors Leicester, 19 September 1996, IEE Digest 196/261. paper 5 7 Moubray, J. Reliability-Centred Maintenance, 1997 (Butterworth-Heinemann, Oxford) 8 Roberts, C. and Fararooy, S. Remote condition monitoring Into the next millennium. In 6th International Conference ‘on Compurer Aided Design, Manufacture and Operation in the Railway and Other Advanced Mass Transit Systems, Lishon, Portugal, September 1998, pp. 715-722 (Wessex Institute, Southampton), 9 Borse, G. Namerical Methods with MATLAB, 1997 (PWS Publishing, Boston, Massachusetts), 10 Lehrasab, N, Generic fault detection and isolation for single-throw mechanical equipment. PAD thesis, University of Birmingham, 1998. 11 Dassanayake, H., Roberts, C. and Goodman, €. An architecture for system-wide fault detection and isolation, Proc. Insin Mech. Engrs, Part I, Journal of Systems and Control Engineering, 2001, 213111), 37-46. 12 Dassanayake, H. Fault diagnosis for new generation of {intelligent electric train door system. PhD thesis, University of Birmingham, 2002. 13 Moody, J. and Darken, C. Fast lear in networks of Jocally-tuned processing units. Neural Computation, 1989, 1, 281-294, 14 Khonen, T. The selforganising map. Proc. IEEE, 1990, 78(9), 1464-1480. 18 Roberts, C., Dassanayake, H. and Goodman, C. Diagnosis Of low-cost industrial assets using distributed models and data networks, In 10th International Conference on System-Modelting-Control 2001, Zakopane, Poland, 21-25 May 2001, Conference Proceedings Vol. 2, pp. 171-175 (Technical University Press, Eoid7, Poland). Proc Insn Mesh Brats Vol 216.Part FJ Rail and Rept Transit

You might also like