Abstract:

The problem of estimating eccentricity severity in induction machines under different and unseen conditions is addressed in this paper. It is challenging to achieve accurate and reliable estimations for fault severity. Unimodal approaches often fail to capture complementary information across heterogeneous sensors to achieve accurate results, while deep-learning models frequently lack interpretability and tend to overfit when the dataset is small, resulting in unstable results, especially when the motor is operating at varying conditions. To overcome these issues, we propose a physics-informed multimodal deep learning framework that integrates physics-informed and statistical features extracted from current, vibration, and air-gap signals. We first design a unimodal regression backbone with learnable scaling and decorrelation penalties to extract compact latent representations. These representations are then fused using a cross-modal attention mechanism, enabling adaptive information sharing across modalities. Model interpretability is enhanced using SHapley Additive exPlanations (SHAP) to identify physically meaningful features. Experimental results demonstrate that our framework achieves the best or second-best performance across different test settings, with explanatory power improvements of 4.4%–10% over unimodal baselines and greater stability compared to conventional fusion strategies.