Detection of rail squats from axle box acceleration: Optimization of a machine learning algorithm

Abstract

The detection of rail surface defects, such as squats, is vital to maintaining the structural health of railway tracks. Squats are a type of rail surface defect which results in large contact forces between the wheel and rail. Squats are self-sustaining; the increase in impact forces results in further rail deterioration. Squats are currently detected through various methods such as ultrasonic measurements, eddy currents, and human inspection. A promising alternative method of squat detection is the use of axle box acceleration (ABA) measurements. Even light squats are visible in the acceleration signal. The application of ABA data to automatically detect squats is an area of current research. Using data and knowledge from a previous study of acoustic squat detection on the German railway, the thesis aimed to optimize a squat detection algorithm based on machine learning. Measured and simulated axle-box acceleration data were supplied from the previous research. The thesis investigated different methods of preprocessing acceleration data to improve the machine learning algorithm. Two algorithms were used- logistic regression and neural networks. The different methods of preprocessing data were spectrogram images, scalogram images, time-averaged wavelet power, and scale-averaged wavelet power. To test the results, the data was divided into a training and a testing set. Furthermore, leave-one-out validation was conducted for the measured squats. Finally, the trained algorithm was tested on two 250 m test sequences of railway track. Issues were found with distinguishing insulated rail joints from squats. Furthermore, a higher success rate often led to a higher rate of false alarms. In these cases, the algorithm failed to generalize to new data. The final algorithm and method of preprocessing with scalogram images found 100% classification of medium to large squats and 87% classification of small squats. The algorithm found a total of 4 false alarms on the two test sequences, one of which was an insulated rail joint. Although the final optimization did not find increased success in identifying small squats in comparison to the previous study, the use of ABA to identify squats was consolidated. Areas of further research are training the algorithm on tracks with varying track dynamics as well as testing other algorithms such as convolutional neural networks.