An Intelligent Diagnosis Method Using Multiscale Features for Compound Fault Diagnosis in Air Brake Systems

Neural network-based methods are effective for single fault diagnosis, but compound faults present challenges due to fault characteristics spanning multiple scales. Single-scale feature extraction leads to misdiagnosis from overlapping signals, and many methods treat compound faults as single faults, making accurate diagnosis difficult. This paper presents a multiscale convolutional neural network (MSCNN) for compound fault diagnosis. Raw signals are processed through the Res2Net block (RB) to reduce feature overlap. To minimize feature map redundancy, the attention block (AB) is used, featuring an one-dimensional simple and parameter-free attention module (1D-SimAM). Finally, a multilabel classifier categorizes the faults. Experimental validation on a high-speed train braking system simulation shows the proposed method outperforms four existing methods”,demonstrating superior performance.