Rail Vehicle Diagnostics

Overview

Research in the area of rail vehicle diagnostics has gained increased attention over the past several decades, resulting in improvements in both reliability and safety. In the freight rail industry, the primary focus has been on the development and implementation of ‘wayside’ equipment for detecting defective rail vehicle components. In North America, there are over 6,000 wayside detectors dedicated to identifying defective rail vehicle components, including wheels, bearings, and truck systems.

The majority of wayside monitoring systems installed in North America are hot-bearing detectors. These detectors utilize infrared non-contact temperature sensors for measurement of the bearing housing temperature as a rail vehicle passes over the detector. Alarms are issued when the bearing temperature exceeds a threshold (typically 105 ºC above ambient). This can be problematic, as the bearing temperature tends to rise only under severe faulty conditions.

A much smaller wayside network of acoustic bearing detectors has been deployed in North America. Acoustic bearing detectors use either acoustic emission (AE) or acoustic signal analysis for fault detection. Vibration transmission through the bearing structure and housing make acoustic techniques difficult for detection of inner raceway defects. Doppler shift and heavy noise have also been identified as key challenges in the wayside acoustic diagnosis of rail vehicle bearings.

Wayside systems have also been developed and implemented for detecting and monitoring truck performance and wheel condition. These systems operate by measuring the forces exerted onto instrumented sections of track. Wheel impact load detectors (WILD) use transient force measurements from passing wheelsets to determine wheel condition. Defective wheels with flat spots, shelling, or spalls tend to create impulsive forces that can be measured from the track response. Truck performance detectors (TPD) use longer sections of instrumented rail to monitor the lateral and vertical forces exerted onto the rail. These measurements can be used to determine geometric issues with the truck, or wheel wear that results in steady-state ‘hunting’ oscillations.

Onboard Rail Vehicle Monitoring

A much smaller body of work has been dedicated to ‘onboard’ rail vehicle diagnostics. This has primarily been due to the considerable investment in wayside infrastructure, but also due to the implementation cost of monitoring equipment. However, as the performance continues to increase and the cost continues to fall for sensing equipment, it is becoming much more practical. There is therefore a need to research and study the response of rail vehicle equipment under ‘normal’ and ‘defective’ conditions. The rail vehicle system exhibits a wide spectrum of nonlinear phenomenon resulting from component clearance, nonlinear contact forces, friction, and defects. Our aim is develop diagnostic techniques that consider this nonlinear behavior and utilize this information for more accurate and efficient detection. We are focused on developing techniques that ‘generalize’ the system, and can be applied to different applications.

Vibration-Based Roller Bearing Diagnostics

A considerable amount of literature is available in the general area of roller bearing diagnostics using vibration. We have implemented many of these techniques, and have extended them with application in freight rail vehicle bearing diagnostics. Figure – 3 shows vibration data collected from normal and defective bearings on a bearing test rig. The bearing defects are shown in Figure – 4. Figure – 5 shows statistical features and autoregressive model parameters extracted from vibration data from normal and defective bearings. No single feature provides the separation necessary to classify the system. As a result, a classifier is used to increase classification accuracy and address the complexities of the system.

A support vector machine (SVM) classifier was chosen for the present work, primarily for its generalization performance compared to other classifiers like an Artificial Neural Network (ANN). Techniques were also used to rank the extracted features based on mutual information between the given feature and target class. Classification accuracy as high as 100% has been demonstrated diagnosing bearings with different size and types of defects.

Model-based Bearing Diagnostics

The first step in model-based diagnostics is the development of an accurate model that sufficiently models the dynamics required for classification. This can be achieved by creating a model for a ‘normal’ system, and comparing physical systems to the predicted model behavior. Large deviations away from the normal system (model) are therefore indicative of a fault. Alternatively, models can be generated to predict the system response with a defect present. This often results in a complex, highly nonlinear model that can only be solved using numerical techniques.

A parametric study of the model response is required to truly capture all facets of the system behavior. This is very challenging for a nonlinear system that can only be solved using numerical techniques. We are seeking to develop quasi-linear models that will allow us to use linear techniques such as the frequency response to study the nonlinear behavior. In rotating systems, roller bearings introduce nonlinearities as a result of the discrete number of rolling elements (varying compliance vibrations), clearance, nonlinear force-deformation relationship, and defects. Using a quasi-linear technique, the nonlinearity is represented by a gain which is optimal for each magnitude of the input to the nonlinearity. The dependence of the output on the input magnitude is maintained, yielding a much accurate prediction of system behavior compared to linearized models.

We have applied this concept of quasi-linearization to a simplified roller bearing system with clearance. With external excitations from rotating unbalance, the harmonic input to the system is assumed to result in a harmonic output with a gain that varies nonlinearly with the input magnitude. The frequency response for the quasi-linear roller bearing system is shown in Figure – 6 for increasing values of the ratio of clearance to eccentricity. A phenomenon known as ‘jump resonance’ is shown to occur as the clearance to eccentricity ratio reaches some critical value. This nonlinear behavior results in two different values of the response for a single input frequency. We plan to extend this work to a system with defects, in order to parametrically study the response based on defect size and location.