An Improved Physics-Guided Interpretable Inception Network and Its Application in Bearing Fault Diagnosis

In industrial applications, obtaining sufficient fault samples is challenging due to factors like sensor installation difficulties and limited fault occurrence. Consequently, small sample size becomes an inevitable obstacle for data-driven fault diagnosis. The Inception network demonstrates strong performance in bearing fault diagnosis, particularly under small sample conditions owing to its structural advantages. However, the commonly used Inception network suffers from poor interpretability and is prone to overfitting. To address these issues, this paper proposes a physics-guided Inception network (PG-Inception), leveraging the physical characteristics of bearing vibration signals to guide the design of its parameters and structure. First, the input length and size are designed based on rules derived from the fault characteristic frequency and shaft rotation frequency. Next, by analyzing the acceleration response within a single fault period and considering possible fault types, the convolution kernel size and the number of parallel branches in PG-Inception are determined. Finally, the effectiveness of PG-Inception is validated using the Case Western Reserve University (CWRU) and the Paderborn University (PU) bearing dataset. The experimental results indicate that under small sample conditions, PG-Inception achieves significantly higher accuracy, faster convergence, and stronger stability compared to the baseline Inception, validating the effectiveness and superiority of the proposed method.