Physics-Guided Sparse Domain Adaptation for Robust Rotor Defect Detection: Integrating System Identification, Synthetic Data Generation, and Maximum Mean Discrepancy Transfer Learning

Abstract

Rotor machinery underpins critical industrial processes across power generation, petrochemical, and aerospace sectors, yet experimental investigation of fault conditions is frequently impractical due to operational safety constraints, prohibitive costs of controlled fault induction, and the rarity of specific fault types in operational data. This paper proposes a physics-guided sparse domain adaptation (PGDL-SDA) framework that resolves the fundamental data scarcity challenge in rotor defect detection by integrating three complementary components. First, a rotor dynamic model is developed through system identification that extracts modal parameters (eccentricity and bearing damping coefficients) from limited experimental observations, enabling physics-faithful synthetic signal generation across a broad range of defect severities and operating conditions. Second, a hybrid deep learning training strategy combines labeled synthetic data from physics-based dynamic simulations with unlabeled experimental data from a physical test rig, exploiting the large labeled synthetic dataset while adapting toward the real-world signal distribution. Third, sparse domain adaptation employs Maximum Mean Discrepancy (MMD) minimization to reduce the distributional shift between synthetic and experimental feature spaces, Domain-Adversarial Neural Networks (DANN) for domain-invariant feature extraction, and entropy regularization to enforce sparse, confident predictions on unlabeled target-domain samples. Experimental evaluation on a custom rotor test rig with four fault conditions (normal, mass unbalance, shaft misalignment, and bearing outer race fault) demonstrates classification accuracy of 94.8% at 1500 RPM, outperforming CNN (81.2%), LSTM (82.1%), CNN+MMD (84.7%), DANN (86.3%), and non-sparse PGDL (89.5%) baselines. Cross-speed generalization experiments confirm the framework maintains above 91% accuracy under speed variations of +/-25%, establishing PGDL-SDA as a robust and practical solution for industrial rotor condition monitoring without requirement for large experimental fault datasets.

Keywords: physics-guided deep learning; rotor defect detection; system identification; domain adaptation; maximum mean discrepancy; DANN; synthetic data generation; condition monitoring