Modern Temporal Convolutional Networks for Intelligent Fault Diagnosis of Machines Under Multiple Working Conditions With Noisy Labels

Machinery fault diagnosis plays an important role in enhancing equipment reliability and ensuring production safety. Varying working conditions and the occurrence of label noise make it difficult to learn patterns from complex data distribution. In this work, loss function named sigmoid symmetric cross entropy in modern temporal convolutional network (SGSCE-MTCN) is proposed to solve the rotating machinery fault diagnosis with noisy labels under multiple working conditions. Vibration data is first preprocessed to time-frequency sequence. Then, MTCN leverages depthwise separable convolution to extract cross-variable and cross-feature information after capturing long-term time series features. An adaptive loss is designed to improve model’s robustness to label noise, the probability of the predicted output processed by sigmoid function is taken as the dynamic parameter for adjusting the loss weight. Comprehensive experiments are conducted on two datasets with multiple working conditions, the proposed MTCN model outperforms other methods and reaches an accuracy over 95%, experiments with label noise indicate that SGSCE enables the model to perform fault classification under varying levels of noise.