Dynamic-Constrained PINN for Complex Machinery System Digital Twin Modeling and Fault Diagnosis

Intelligent applications of artificial intelligence techniques have become essential in industrial maintenance, yet challenges persist in obtaining labeled fault data under time-varying operational conditions and harsh environments. To address mechanical system modeling with undetermined states and fault diagnosis with missing fault data, this paper proposes a dynamic-constrained Physics-Informed Neural Network (DcPINN) framework for complex machinery systems digital twin modeling and fault diagnosis. A dynamic-constrained PINN architecture that integrates physical laws with neural networks for digital twin modeling based on limited measured data is constructed. Moreover, a novel parametric fault impact embedding mechanism is designed to generate synthetic fault signals under dynamic constraints for diagnosis model training. Experimental validation demonstrates the framework's dual capabilities: The dynamic-constrained PINN achieves accurate system modeling highly parameter estimation accuracy under undetermined mechanical states, and enables effective fault diagnosis with 89.3% average recognition rate without historical fault samples. Comparative studies show the proposed method outperforms conventional unsupervised diagnostic approaches, and this work establishes a new paradigm for PINN assisted predictive maintenance in complex machinery systems based on artificial technology.