Search

In engineering practice, deep learning is commonly used for small-sample fault diagnosis of mechanical equipment under heavy noise interference. In recent years, many researchers have used attentional neural operators based on CNNs to extract feature representations from complex data. However, the current attentional neural operator-based fault diagnosis method still has the following limitations: (1) CNN performs poorly in global feature extraction and is unable to capture the global dependencies of long sequential signals. (2) Neural operators have a serious information loss problem in multi-level feature computation, which is unfavorable for processing sensor signals that are long and contain complex noise. To address these problems, we innovatively design an efficient attentional neural operator, and also propose a new lightweight model called CSDANet. Firstly, the spatial interaction between local and global in vibration signals is facilitated by introducing a multi-level convolution and layer scaling strategies; secondly, We designed a feature channel with long-term memory capacity to cope with the adverse effects of information loss on the model; Finally, a vibration signal feature extractor is proposed to obtain richer feature mapping from the original one-dimensional sensor signals, which significantly improves the ability of the attention module to perceive multi-frequency domain noise features while consuming less computational resources. The results of two experiments demonstrate that the proposed approach accommodates advantages of lightweight and robustness in small-sample fault diagnosis tasks, compared to the existing mainstream Transformer and CNN fault diagnosis frameworks. [ABSTRACT FROM AUTHOR]