Project: Enhanced Rail Contribution by Improved Reliability (ERCIR)

Reference: STP 14/6/19

Last update: 15/01/2008 14:41:27

Objectives

To enhance the ability of rail services to run safely to time, without unplanned disruptions an with consistent ride quality so that they can realise their full potential as part of an integrated national transport system by reliably meeting connections. The enabling technical proposition is an advanced form of in-service condition monitoring of track and vehicles. If incipient faults, or ride quality deterioration, can be predicted before they result in disruption, or even danger, the role played by rail can be much improved. This is likely to be particularly important where inter-running of widely differing vehicles is planned.

To improve the performance of rail-related fault prediction and diagnosis systems. This requires a greater theoretical understanding of track/vehicle wear in order to enhance the interpretation of data collected from transducers. Furthermore, it is intended to assess the feasibility of whether modem data processing and communication methods are able to implement mathematical algorithms in such a way as to opertae on in-service vehicles. This will enable continuous monitoring to take place on a wide range of vehicles wherever they are operating.

Description

To establish the feasibility of an on-board, in-service monitoring and fault prediction system focussed on the vehicle-track interaction. A system architect using distributed processing an sensors will be developed to act as a generic monitoring and fault diagnosis system with the ability to be fitted to all types of vehicle.

Contractor(s)

University of Birmingham
Electronic, Electrical and Computer Enginerring, Edgbaston, Birmingham, B15 2TT
0121 414 4287

Contract details

Cost to the Department: £375,336.00

Actual start date: 06 December 2001

Actual completion date: 30 September 2004

Publication(s)

Estimation of parameters in a linear state space model using a Rao-Blackwellised particle filter.
Author: P Li, R M Goodall and V Kadirkamanathan
Publication date: 06/11/2004
Source: IEE Proc. Control Theory Appl., pp 727-738 Vol 151.

Summary of results

  1. The Enhanced Rail Contribution through Improved Reliability (ERCIR) project was aimed at investigating the feasibility of improving in-service condition monitoring systems for railway vehicles and track. The principal foci of the investigation were to be optimising the sensor set(s) and developing fault detection and identification processing algorithms that would enable developing faults to be identified before necessarily representing a threat. As a pre-cursor, data would be gathered on the prevalence of different faults and their susceptibility to detection by sensors suitable for fitting under normal trains.
    The project was initiated by collecting data on vehicle and track faults from the experiences of the project partners and by utilising their fault reporting databases. A Failure Modes, Effects and Criticality Analysis (FMECA) was undertaken to identify and prioritise the indentified faults. Based on the fault list generated by the FMECA, the development and implementation of the fault detection algorithms were targeted on the common faults (in terms of criticality and frequency of failures reported).
    The majority of vehicle faults reported on Alstom's FRACAS database are wheel profile and suspension related. Among these components, secondary lateral dampers and anti-yaw dampers were identified as most prone to failure. It is thus necessary to monitor the condition of dampers during normal service in an attempt to predict incipient failures. The damper failures identified in the FMECA primarily affect the lateral and yaw dynamics of the bogie and wheelsets.
    Measurements taken from the vehicle-mounted sensors are fed into a residual generator derived from a dynamic vehicle model. The residual signal is then used in a knowledge-based system to reveal fault information. Two different approaches to constructing the residual generator have been investigated, namely the state estimation and parameter estimation approaches. In the state estimation approach, the unknown and unmeasured states are estimated (using a Kalman filter) to predict the fault free system output that is compared to the observed output, whereas the parameter estimation-based approach combines theoretical modelling and parameter estimation to generate the residuals.
    For track faults, linear analysis shows that certain track geometry information should be observable from a small set of simple sensors. A bogie-mounted pitch-rate gyro allows reconstruction of mean vertical alignment for wavelengths above about 8 m, working down to lower speeds than a similarly priced vertically-sensing accelerometer. Deterioration in the vertical alignment over time can be monitored as well as certain faults such as voiding and cyclic top. Adding left and right axlebox-mounted, vertically-sensing accelerometers allows reconstruction of the left and right rail shorter wavelength vertical irregularity detail, as long as the vehicle travels sufficiently quickly, enabling detection of dipped joints and irregular crossings. Absolute track crosslevel is not observed very accurately from the bogie alone, but a useful estimate of crosslevel is possible using bogie-mounted roll- and yaw-rate gyros.
    A pitch rate gyro is then used to obtain the mean vertical alignment of the track at wavelengths longer than a few times the bogie wheelbase, with modest deterioration at wavelengths associated with the bogie pitch mode at high vehicle speeds. Axlebox accelerometers measure short wavelength vertical alignment up to wavelengths related to the bogie bounce, pitch and roll resonances. Similarly, the bogie roll rate gyro gives an approximation of the track crosslevel for longer wavelengths while the difference between the axlebox vertical motions fills in short wavelength detail. The absolute bogie roll can be computed using a combination of lateral-sensing accelerometer with yaw and roll rate gyros on the bogie. The vehicle speed is also required to perform the conversion between time and displacement along the track.
    Versines (also known as Mid-Chord Offsets) over various chord lengths have been used to isolate vertical alignment defects. Features such as voids can be obtained by looking at 8 or 16 m versines computed directly from observations from a bogie-mounted pitch rate gyro. Shorter wavelength features on the track, such as dipped joints and crossings, can be monitored using the 1 m versines computed from the accelerations from the axlebox.
    Five tests were carried out on T&W Metro, and one on a Coradia vehicle travelling form Chester to Llandudno and back. A second Coradia trial is still pending.
    The first trial was used for testing the equipment. During the second trial a vehicle with old dampers was used. On the third trial, data were collected on the same vehicle with a new set of dampers. Trial four used a vehicle with worn wheels and trial five was carried out after the vehicle underwent wheel turning. The comparison of data for the latter two trials allowed the investigation of the effects of wheel/rail conicity on the detection algorithms for damper failure detection.
    There has been continuous liaison with all the project partners. Alstom provided extensive access to databases, technical reports and vital support for mainline testing. This latter led to a technical paper to AEA form the UofB suggesting a possibly superior method of integration. Carillion provided access to a vast store of knowledge over track faults that was captured as part of the FMECA work. This, and further more recent liaison has highlighted the problems of S&C monitoring. This issue has so far proved very difficult to tackle with the sensor sets chosen for this project and is clearly an important area for further investigation. Nexus provided substantial practical guidance but, crucially, the opportunity for the first practical trials which undoubtedly focussed the project on realistic targets.
    This project has demonstrated the feasibility of monitoring key aspects of both track and vehicle condition from simple and robust sensor sets mounted on normal trains. Actual trials on two types of train have been undertaken and the data processed by a selection of algorithms that would form the core of the anticipated on-line, real-time fault detection system that is the ultimate aim of this work.
    A bogie-mounted pitch rate gyro can reconstruct mean vertical alignment at wavelengths above about 8 m. Adding axlebox accelerometers enables shorter wavelength reconstruction. Adding bogie yaw and roll gyros will give some estimate of crosslevel, but this particular parameter is better measured by adding displacement transducers.
    Track mean vertical irregularity, cross-level and twist can be estimated from a set of bogie- and axlebox-mounted sensors, without mechanical or optical displacement transducers. The use of the versine and signal processing algorithms developed in this research for the detection of voids, dips and wetspot were demonstrated. By comparing the same piece of track at different scales, the causes of poor track alignment can be distinguished. This method offers better understanding of track alignment than a conventional standard deviation method over a 1/8 mile checking for exceedences. In contrast to the model-based approach, fault detection and condition monitoring of track irregularities was carried out using a benchmarking approach. This approach, being data driven and independent of the model, allows 'quality' of the track to be evaluated.
    To achieve FDI of vehicle suspension faults (primarily damper deterioration), two approaches have been demonstrated, both by simulation and with real data from the trials. The two approaches are state estimation using a Kalman-filter based innovation approach and RBPF based parameter estimation. The Kalman-filter method is computationally efficient and has a rapid response to an abrupt fault and is thus suitable for using on-line. The RBPF based method is computationally intensive and can only be used where the detection time is of minor importance, hence is suitable monitoring damper drift,
    for example. A combination of both these methods is proposed for future development.