A novel physical constraint-guided quadratic neural networks for interpretable bearing fault diagnosis under zero-fault sample
You Keshun, Yingkui Gu, Lin Yanghui, Wang Yajun
Abstract
To increase confidence in intelligent damage assessment models, research often emphasizes interpretability but neglects essential physical fault mechanisms. Addressing this gap, we propose a novel hybrid learning framework combining a dynamic physical model with a data-driven Quadratic Neural Network and Bidirectional LSTM (QNN-Bi-LSTM). Acceleration and velocity equations of the rolling element, derived physically, are embedded as constraints into the loss function, guiding learning. This Physical Constraint-Guided (PCG) learning aligns physical laws with features from real-world vibration data. To enhance generalization under zero fault samples, a genetic algorithm adaptively balances physical and cross-entropy losses. Experiments show the model achieves near 100% accuracy no-load and 95% heavy-load with zero fault samples, exhibiting strong robustness. Compared to conventional models, it significantly enhances interpretability, credibility, and reliability. This highlights a new direction for NDE and condition monitoring, especially for insufficient fault scenarios. Integrating physical principles and nonlinear feature learning offers a promising, generalizable solution for intelligent damage assessment in critical systems like rolling bearings.