Attention-guided wavelet packet energy autoencoder with distribution consistency for single-source domain generalization fault diagnosis

Abstract Single-source domain generalization for fault diagnosis under unknown working conditions remains a challenging task due to the absence of target domain data during training and the domain shift caused by varying operating parameters. Existing methods often struggle to extract domain-invariant features from a single source domain, particularly when facing both stationary and unseen operating-condition shifts across different constant-speed domains. To address these limitations, this paper proposes an Attention-Guided Wavelet Packet Energy Autoencoder with Distribution Consistency (AGWPEA-DC) for robust fault diagnosis under unknown working conditions. The proposed framework comprises three key innovations. First, a dual-autoencoder architecture is designed to simultaneously learn general representations from raw vibration signals and physical features from wavelet packet energy, embedding time-frequency information into a low-dimensional latent space. Second, a wavelet packet energy attention (WPEA) mechanism is introduced to adaptively recalibrate the energy coefficients across different frequency subbands, emphasizing fault-relevant components while suppressing noise and irrelevant information. Third, a wavelet packet energy distribution consistency loss is developed to explicitly enforce intra-class compactness and inter-class separation in the energy feature space by maintaining dynamic class centers, thereby enhancing the discriminative power and generalization capability of the learned physical features. Extensive experiments are conducted on two bearing datasets comprising 12 cross-domain tasks under various unknown operating speeds. The results demonstrate that the proposed AGWPEA-DC consistently outperforms state-of-the-art methods, in terms of accuracy, F 1-score, and AUC.