Abstract: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.

An Efficient Method for Wheel-Flattened Defects Detection Based on Acoustic Emission Technique

Online Wheel Flaw Detection System Based on Electromagnetic Acoustic Technique

A Novel Rail Defect Detection Method Based on Undecimated Lifting Wavelet Packet Transform and Shannon Entropy-Improved Adaptive Line Enhancer

Detection Method Based on Kalman Filter for High Speed Rail Defect AE Signal on Wheel-Rail Rolling Rig

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

Defect Prediction of Railway Wheel Flats based on Hilbert Transform and Wavelet Packet Decomposition

Wheel and axle defect detecting technique using phased array ultrasonic

An Improved AE Detection Method of Rail Defect Based on Multi-Level ANC with VSS-LMS

Hybrid Microphone Array Signal Processing Approach for Faulty Wheel Identification and Ground Impedance Estimation in Wheel/rail System

Detection of Wheel Rim by Immersion Scan of Phased Array Ultrasonic Flaw Testing

Acoustic emission detection of rail defect based on wavelet transform and Shannon entropy

Automatic railway wheelset inspection system by using ultrasonic technique

Research on Eddy Current Pulsed Thermography for Rolling Contact Fatigue Crack Detection and Quantification in Wheel Tread

Research on Automatic Defect Localization for Ultrasonic Normal Probe Detection on Railway Wheel

Railway Wheel Defect Dynamic Inspection System Based on Ultrasonic Technology

Identification of wheel surface cracks under high noise environment based on texture segmentation and combination descriptor

An Analysis of the Simulated Acoustic Emission Sources with Different Propagation Distances, Types and Depths for Rail Defect Detection

Railway Wheel Flat Detection Based on Improved Empirical Mode Decomposition

High-speed railway wheel polygon detection framework using improved frequency domain integration

The Early Stage Wheel Fatigue Crack Detection Using Eddy Current Pulsed Thermography

Bearing Fault Detection Based on Acoustic Emission Technique