Health state assessment of complex electromechanical systems based on AINN and frequency domain causal network

The reliability of complex electromechanical systems is crucial for industrial and national operations. However, accurately assessing the health state of these systems presents significant challenges due to high-dimensional data, nonstationary conditions, and limited labeled data. To address these challenges, we proposed an unsupervised health state assessment method that integrated an Autoregressive Informed Neural Network (AINN) with Generalized Partial Directed Coherence (GPDC). Our approach, AINN-GPDC, introduced an Iterative Variable Selection (IVS) strategy to establish correlation models among high-dimensional parameters and leveraged GPDC to construct frequency-domain causal networks, effectively handling nonstationary and nonlinear data. The method provided a comprehensive evaluation of system health by constructing health indicators based on graph density and graph entropy, enabling precise detection of system state changes. The proposed framework was validated on the N-CMAPSS and real-world satellite datasets, demonstrating superior performance in early anomaly detection and failure prediction. Our findings highlight the potential of AINN-GPDC as a robust tool for health monitoring, particularly in complex, unlabeled, and dynamic environments.