To address the issues of insufficient feature extraction ability and low efficiency of manual parameter optimization in traditional single deep learning models for processing data under variable operating conditions, this study proposes a physics-prior-based time-frequency dual-channel model, CNBiG-AM. The model is designed to handle the different impacts of variable conditions on the manifestation of time-frequency domain fault features by employing a dual-channel time-frequency approach for complementary feature learning. At the same time, a lightweight and efficient model and attention collaboration mechanism are established, utilizing decision-level deep feature fusion for diagnostic decisions. The Sparrow Search Algorithm (SSA) is used as an intelligent optimization strategy to ensure the optimal performance of the complex model under variable operating conditions. Ablation experiments using the CWRU and HNU datasets demonstrate that CNBiG-AM outperforms baseline models in all metrics, with accuracy improvements of 6.5% and 5.5%, respectively, proving the necessity of the BiGRU-attention collaboration mechanism. After intelligent optimization using SSA, the model performance is further enhanced, with accuracy improving by 2.5% on the CWRU dataset and 3.5% on the more generalized HNU dataset. The t-SNE visualization graph shows a significant improvement in the feature space aggregation effect. This study provides an efficient and robust ensemble model for bearing fault diagnosis under variable operating conditions, demonstrating the effectiveness of dual-frequency separation and intelligent optimization algorithms.
Bu et al. (Sun,) studied this question.