Unsupervised subdomain contrastive adaptation for elevator fault diagnosis based on time-frequency feature attention mechanism segmentation

Existing elevator fault diagnosis algorithms have limited engineering applicability due to variations in working conditions and differences in equipment structures. To address this limitation, this study proposes an unsupervised subdomain adaptation method based on a time-frequency feature attention mechanism, LMMD-based subdomain alignment, and contrastive local alignment. This enables the application of the diagnosis model across different working conditions and equipment types. First, a novel time-frequency feature attention mechanism assigns weights to vibration signals of varying dimensions. Second, the time series is transformed to obtain a three-channel time-frequency diagram. This diagram is input into the proposed dimension-segmentation cross-channel multihead self-attention framework to extract high-dimensional frequency-domain fault features. These features are concatenated with the time-domain features to obtain a global feature representation. Then, the extracted high-dimensional features are sent to the classification module to obtain the predicted labels for the source and target domains. Finally, after confidence filtering, the true labels from the source domain and the prediction labels from the target domain are fed into a dynamically weighted multilevel feature alignment module to promote proximity between similar fault features across domains while enhancing separation among different fault types. The validity and superiority of the proposed method were demonstrated through simulation experiments conducted on two types of manned escalator systems under multiple working conditions. For the most challenging transfer task, the proposed method achieved higher accuracy on the target domain test set than DANN, ADDA, C-CLCN, TFA-CCN, and TFA-LCN by 26.87%, 24.72%, 11.44%, 28.94%, and 16.85%, respectively.