Monitoring of railway tracks maintenance needs using dynamic responses collected by an in-service train

Abstract Railway systems are critical components of transportation networks requiring consistent maintenance. This paper proposes a novel data-driven approach to detect various maintenance needs of railway track systems using acceleration data obtained from a passenger train in operation. The framework contains four modules. Firstly, data pre-processing and cleansing are performed to extract useful data from the whole dataset. Then, condition-sensitive features are extracted from the raw data in three different domains of time, frequency, and time–frequency. In the third module, the best subset of measurement features that characterize the state of the tracks are selected using the analysis of variance (ANOVA) algorithm which eliminates irrelevant characteristics from the feature set of responses. Finally, a multilabel classification algorithm based on the cascade feed-forward neural network (CFNN) is used to classify the type of maintenance needs of the track. An open-access dataset from a field study in Pennsylvania, USA, is used in this study for validation of the proposed method. The results indicate that employing a CFNN can achieve 95% accuracy in identifying two maintenance activities, tamping and surfacing, using time-domain features. Moreover, an extensive analysis has been conducted to evaluate the influence of various feature extraction and selection methods, diverse classification algorithms, and different types of accelerometers (uni-axial and tri-axial) on the accuracy of the proposed method.