Kernel Based Open CNN Algorithm for Known and Unknown Fault Diagnosis of Train Traction Motor

Time-frequency map analysis method based on short-time Fourier transform (STFT) and deep learning algorithm based on convolutional neural network (CNN) are the most popular techniques in the field of vibration signal-based fault diagnosis. However, the unknown faults that often exist in open real-world environment, which causes the fault diagnosis accuracy of CNN based on traditional closed-set assumption show disastrous degradation. To solve the above problem, a kernel-based open CNN (KO-CNN) algorithm is proposed in this paper for known and unknown fault diagnosis of train traction motor. Firstly, the STFT is used to transform high-frequency vibration signal into time-frequency map to analysis time- and frequency-domain information simultaneously; Then, the kernel auto-encoder (KAE) is used to improve the conventional CNN for constructing the KO-CNN algorithm. The KO-CNN uses the diagnosis label from the conventional CNN and the feature data from the last fully connected layer of CNN to build KAE model for all known categories to diagnose between known and unknown faults at the same time. Finally, the PHM-BEIJING2024 train traction motor dataset is used to conduct the experiment. The results demonstrate that the proposed KO-CNN has better diagnosis performance of normal state, known faults and unknown faults in open real-world environment compared to conventional CNN.