Wavelet packet energy embedded autoencoder with dynamic weighting strategy for fault diagnosis under unknown working conditions

Recent advancements in domain generalization techniques can be employed to address the issue of inaccessible target domain data in fault diagnosis. However, most methods require multiple source domain datasets to learn domain-invariant features. Simultaneously, target domain data are often collected under unknown working conditions (stationary or nonstationary), significantly increasing the difficulty of fault diagnosis tasks. In this regard, this paper proposes a wavelet packet energy embedded autoencoder (WPEEAE) with a dynamic weighting strategy (DWS) to address these challenges. First, a wavelet packet energy encoder is designed, encoding wavelet packet energy into tensors with values of 0/1, enhancing feature sparsity. Second, an autoencoder is utilized to embed the wavelet packet energy corresponding tensors into a low-dimensional space consistent with the data sample features. A DWS is introduced to balance the importance between the general features corresponding to the data samples and the wavelet packet energy embedded features. Finally, a sparse regularized activation (SRA) function with feature constraint capability is developed and integrated into a classifier operating in low-dimensional space. Due to the incorporation of wavelet packet energy in the proposed model, fault diagnosis tasks can be accomplished in the testing phase with only the original data. The effectiveness of the proposed approach is validated in datasets containing constant and time-varying speed fault conditions. The code can be found at https://github.com/John-520/WPEEAE-DWS.