A novel few-shot deep regression domain adaptation method with wavelet scattering attention mechanism for fault prognostics

Deep transfer learning (DTL) has shown substantial potential in fault prognostics. However, most current DTL methods still suffer from insufficient run-to-failure training data in source domain. Moreover, model interpretability is crucial for trustworthy predictions in few-shot scenarios. To address these problems, this paper proposes a novel few-shot deep regression domain adaptation method with wavelet scattering attention (WSA) mechanism. This method adopts domain-adversarial neural network as the baseline architecture. In the feature extractor, the WSA mechanism is designed by introducing wavelet scattering transform (WST) into a deep convolutional neural network. Under different decomposition scales and rotational orientations of WST, WSA obtains a set of deep time-frequency features with frequency-level meanings, each frequency channel of which is dynamically weighted with different significances. Through domain-adversarial training with an LSTM regression prediction and a temporal tendency regulariser, the temporal features with various frequency significances can be extracted to raise a fine-grained representation of degradation information. Experiments on the MNIST handwritten character dataset first visually verify the effectiveness of WSA, while a set of prognostic experiments on two rolling bearing whole-life datasets validate the superiority of the proposed method in prediction accuracy and the frequency-aware interpretation under few-shot setting.