Wheel Defect Detection Using Attentive Feature Selection Sequential Network With Multidimensional Modeling of Acoustic Emission Signals

Wheel defects, such as flats, can have severe impacts on vehicles, seriously hindering railway running stability and safety. Thus, it is crucial to continuously monitor the condition of wheelsets using techniques such as acoustic emission (AE). In this article, we propose an attentive feature selection sequential (AFSS) module inspired by attention-based mechanisms, which can be integrated into deep sequential neural networks, such as bidirectional gated recurrent units (biGRU). The feature sets were extracted from the AE signals acquired with various wheel defects using Mel-frequency cepstral coefficients (MFCC) and Gammatone cepstral coefficients (GTCC). Next, irrelevant and redundant features were removed, and the optimal feature subset was obtained with the simultaneous training of the AFSS modules and trunk classifier network. The proposed method was compared with several state-of-the-art (SOTA) feature selection (FS) algorithms. The evaluation was conducted using experimental wheel defect datasets. Considering the imbalance in the datasets, the methods were further enhanced using an ROC curve-based threshold rescaling strategy. The experimental results demonstrate that AFSS outperforms other algorithms in terms of accuracy and generalization without an excessive increase in computational costs; moreover, AFSS exhibits superior stability and noise resistance, as confirmed by false-alarm rate (FAR) tests in practical defect-free validations. Overall, the proposed method combines the benefits of the learning capabilities of sequential neural networks and the embedding of FS in classifiers; thus, this study provides theoretical guidance for employing FS techniques in sequential network modeling and wheel defect detection tasks.