Abstract: Aimed at the problem of industrial process fault diagnosis with the mixed characteristics of multimode and multiscale, a fault diagnosis method based on multiscale temporal convolutional network is proposed. Considering the multimode distribution characteristics of process data, we used the local neighborhood standardization method based on cosine similarity to standardize the process data to eliminate the multimode characteristics. Aimed at the multiscale characteristics of the process data, the multiscale representation of the process data was obtained by variational mode decomposition, a temporal convolutional network model with attention mechanism was constructed for each component to extract features, and the multiscale features were fused to achieve multiscale feature extraction. On the basis of the feature extraction, the fault diagnosis was realized by a full connection layer. The effectiveness and feasibility of the proposed method are verified by Tennessee-Eastman (TE) process simulation experiments.