Physics-Guided Deep Learning for Fault Diagnosis in Electro-Mechanical Systems

Abstract

Electro-mechanical systems are critical in industrial operations, where unexpected failures result in catastrophic economic losses. Early fault detection through Prognostics and Health Management (PHM) technologies is essential. However, developing effective diagnostic systems faces a fundamental dilemma. Model-based approaches employ idealized representations that fail to capture real-world complexity and lack adaptability to evolving fault conditions. Conversely, data-driven approaches suffer from limitations: fault data are exceptionally scarce, these models exhibit black-box characteristics that obscure physical interpretation, demonstrate poor generalization beyond training scenarios, and frequently produce results inconsistent with physical laws. This research addresses these challenges through a physics-guided artificial intelligence framework that systematically integrates physical principles with data-driven methodologies across three complementary research thrusts. First, a scale-free health indicator addressing the capacity-dependency problem in electrical fault diagnosis was developed. Conventional motor current signature analysis indicators exhibit inconsistent performance across different motor capacities because motor impedance scales nonlinearly with capacity. The proposed Max-Min Crest Factor (MMCF) exploits morphological characteristics of current waveforms through wavelet-based denoising and max-min normalization. Validation through simulation, experimental testbeds, and a large-capacity industrial motor (100 W to 50 kW) demonstrated that MMCF enables unified diagnostic thresholds without individual calibration. Second, an interpretable deep learning architecture for bearing spall size estimation was established. Expert knowledge methods require accurate assumptions about spall-excited events that are difficult to predict under field conditions, while pure data-driven models suffer from black-box opacity and physics inconsistency. The proposed Frequency-Enhanced Neural Network (FENN) integrates modified Fourier convolution layers that simultaneously capture local temporal patterns and global frequency characteristics. The Hybrid Spall Size Estimator (HSSE) combines physics-based formulas encoding expert knowledge with data-driven feature learning, enabling physically interpretable predictions from raw vibration signals. Domain generalization experiments demonstrated successful transfer from simulation to experimental testbed, achieving practical accuracy on limited training data. Third, a physics-informed neural network framework for vibration signal generation was developed to address computational inefficiency and data scarcity challenges of conventional simulation approaches. The Bearing Runge-Kutta Physics-Informed Neural Network (BR-KPINN) directly integrates bearing dynamics governing equations through a physics-informed derivatives process and Runge-Kutta temporal integration. The two-phase training strategy incorporates physics constraints through composite loss functions. BR-KPINN achieves high validation accuracy (cosine similarity of 0.95) while reducing data generation time from hours to 1.93 seconds. When integrated with the diagnostic model, BR-KPINN-generated data improved diagnostic performance by 21.2% on unseen testbed conditions. This research demonstrates that physics-guided artificial intelligence frameworks can address the fundamental challenges in model-based and data-driven diagnostic approaches. By embedding physical principles into health indicators, neural network architectures, and training processes, the proposed methodologies achieve scale-invariant detection, interpretable predictions, and physics-consistent generalization under data-scarce conditions. These contributions provide a technical foundation for developing adaptive diagnostic systems capable of continuous improvement through field data integration, with applications in real-time maintenance, digital twin-based prognostics, and autonomous health monitoring of electro-mechanical systems.