A physics prior spectral embedding network expansion method for domain-incremental fault diagnosis under variable working conditions

Key rotating components such as bearings and gears operate under complex and variable working conditions, which significantly increases their risk of failure and directly compromises the reliability of the mechanical system. Existing methods for solving fault diagnosis under variable conditions, such as transfer learning, struggle to achieve satisfactory performance across both previously learned and newly encountered domains. As a result, models tend to suffer from severe catastrophic forgetting of previously acquired knowledge when adapting to new domains. To address this issue, this paper proposes a domain-incremental fault diagnosis method based on a Physical Prior Spectral Embedding Network Expansion framework (PPSENE). First, a Physical Prior Spectral Embedding Network (PPSEN) is developed. Specifically, global instance compression is employed to normalize time-domain signals, thereby reducing distribution discrepancies across different domains. Then, a real-valued Fast Fourier Transform is applied to transform signals from the time domain to the frequency domain, facilitating physically interpretable spectrum input. A logarithmic magnitude compression mechanism is introduced to mitigate the adverse effects of extreme amplitudes on feature learning. Meanwhile, frequency domain instance normalization is utilized to alleviate the scale drift of spectral energy across domains, enabling effective physical prior modeling of vibration signals. Finally, a multi-layer perceptron is adopted for deep feature extraction, yielding discriminative fault representations. Based on this, a Physical Prior Spectral Embedding Network Expansion framework is constructed for domain-incremental fault diagnosis. By expanding the spectral network, parameter isolation across different domains is achieved, which enhances the stability of the model during incremental learning. In addition, a knowledge distillation loss is introduced to align the output distributions between the new model and the old model. The proposed method is validated on a comprehensive gear transmission test rig dataset and a high-speed railway wheelset bearing dataset. Experimental results demonstrate that the proposed approach effectively alleviates catastrophic forgetting in domain-incremental fault diagnosis scenarios while maintaining superior diagnostic performance under varying working conditions, the average accuracy exceeds 98 % on both datasets.