Multi-scale Transformer-CNN domain adaptation network for complex processes fault diagnosis

Abstract

Despite the breakthroughs in deep neural network-based fault diagnosis, the model mismatch problem owing to the changes in data distribution remains challenging. To fuse deep features for cross-mode feature modeling, a Transformer-convolutional neural network (TrCNN) based multi-scale distribution alignment network is proposed. In the source domain stage, a concatenated structure of Transformer and convolutional neural network (CNN) extracts deep diagnostic information by combining global and local approaches. In the transfer stage, alignment is performed on the complex features extracted from different CNN substructures at multiple scales. Multi-scale feature alignment allows aligning information from various aspects while maintaining the discriminability of the data. The effectiveness and feasibility of the proposed method were demonstrated through experiments conducted on the Tennessee-Eastman (TE) process and industrial three-phase flow (TFF) equipment.