Multi-Source Information Fusion Based Motor Fault Diagnosis via Deep Discriminative Residual Adaptation

Asynchronous motors play a key role in industrial equipment. To cope with the challenge of incomplete fault information due to a single signal, and the problem of missing target domain labels. this paper proposes a method that combines multi-source information fusion with an improved domain adaptation approach. First, an autoencoder is used to extract features from the collected vibration, current and temperature signals. The extracted features are concatenated and fused at the channel level. Then, a deep feature extraction network is designed by combining efficient channel attention and deep residual network, which can accurately and efficiently extract deep features from multi-source information. Finally, an improved domain adaptation mechanism is proposed, which significantly enhances the learning of discriminative features between two different domains. In addition, by incorporating the I -Softmax loss, the constructed Deep Discriminative Residual Adaptive Network exhibits stronger classification capabilities. The superiority and feasibility of the proposed method is verified by comparing several methods under the sample of asynchronous motor without labeling the target domain.