Detecting train driveshaft damages using accelerometer signals and Differential Convolutional Neural Networks

TL;DR

This work targets railway axle crack detection by converting vibration signals into time-frequency spectrograms and applying a differential 2D-CNN that uses a healthy-reference spectrogram to enhance generalization across wheelset assemblies. The architecture integrates a six-layer 2D-CNN with three inputs (signal spectrogram, reference spectrogram, and static features) in a shared-weight, dual-branch design, fused through an MLP classifier. Across WA1–WA3 datasets, the proposed differential CNN achieves the highest AUROC values (0.93, 0.86, 0.75) and outperforms LSTM, 1D-CNN baselines and domain-adaptation approaches, while removing the need for extra preprocessing steps like tsfresh. The approach demonstrates robust crack-detection performance and provides a practical, generalizable condition-monitoring solution for railway axles, with future work planned on attention-based interpretations to further elucidate decision regions.

Abstract

Railway axle maintenance is critical to avoid catastrophic failures. Nowadays, condition monitoring techniques are becoming more prominent in the industry to prevent enormous costs and damage to human lives. This paper proposes the development of a railway axle condition monitoring system based on advanced 2D-Convolutional Neural Network (CNN) architectures applied to time-frequency representations of vibration signals. For this purpose, several preprocessing steps and different types of Deep Learning (DL) and Machine Learning (ML) architectures are discussed to design an accurate classification system. The resultant system converts the railway axle vibration signals into time-frequency domain representations, i.e., spectrograms, and, thus, trains a two-dimensional CNN to classify them depending on their cracks. The results showed that the proposed approach outperforms several alternative methods tested. The CNN architecture has been tested in 3 different wheelset assemblies, achieving AUC scores of 0.93, 0.86, and 0.75 outperforming any other architecture and showing a high level of reliability when classifying 4 different levels of defects.

Figures (9)

• Figure 1: Bogie framework.
• Figure 2: Spectrogram representation of a vibration signal a) real part, b) imaginary part, c) magnitude.
• Figure 3: Model architecture. Each convolutional layer follows a $2\times2$ MaxPooling layer, hence, the image is divided by two at each layer. There are three inputs: the spectrograms, the reference spectrograms, and the static data. Thus, three embedding are calculated. Finally, they are all fused to compute the four possible defects: D0, D1, D2, and D3.
• Figure 4: 1D-CNN LSTM baseline architecture. There are three inputs: the signals, the reference signals, and the static data. Thus, three embeddings are calculated. Finally, they are all fused to compute the four possible defects: D0, D1, D2, and D3.
• Figure 5: Training ROC curve over WA1.