Diagnosis of brake friction faults in high-speed trains based on 1DCNN and GraphSAGE under data imbalance. (15th February 2023)

• Record Type: Journal Article
• Authors: Zhang, Min; Li, Xianjun; Xiang, Zaiyu; Mo, Jiliang; Xu, Shihao
• Is Part Of: Measurement. Volume 207(2023)
• Publication Date: 2023-02-15
• Subjects: Fault diagnosis -- Brake friction faults -- One-dimensional convolutional neural network -- GraphSAGE; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.112378
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: The issues of brake friction fault showing data imbalance are solved. An intelligent diagnosis algorithm based on 1DCNN and GraphSAGE network for brake friction fault diagnosis under data imbalance is proposed. The proposed network can effectively identify brake friction faults without extending fault data and adding cost sensitivity to the network. The superiority of the proposed method is verified. Abstract: A braking friction fault diagnosis method based on one-dimensional convolutional neural network (1DCNN) and GraphSAGE network is proposed to solve the problem of fault imbalance samples in actual high-speed train braking friction operation, taking into account the correlation between different fault features. To begin, the original sample is created using the friction interface state characterisation parameters such as vibration noise, vibration acceleration and friction coefficient. Second, the graph is built using the sample's characteristics as well as the Jensen-Shannon divergence between each sample. The 1DCNN is then used to extract and compress the graph node features; Next, the GraphSAGE is used to aggregate the information of each node in the graph, compensating for the neural network's inability to learn the features of small samples and ensuring that all kinds of fault information are fully extracted. Finally, GraphSAGE outputs the braking friction fault state category to realise braking friction fault diagnosis with imbalanced data. The proposed network was tested using various imbalanced data sets and it was discovered that even with fewer fault samples and more normal samples, the network can still achieve at least 93.83% effective diagnostic accuracy. The effectiveness of the proposed network for each braking fault identification is further verified using precision, recall, F1 score and t -distribution stochastic neighbour embedding ( t -SNE) visualisation. The superiority of the proposed network is validated when compared to the imbalanced data processing method and other state-of-the-art networks, indicating that the proposed network can achieve more effective fault diagnosis under imbalanced data without data expansion and large changes to the network, providing a new feasible method for research in this direction.