Physics-informed deep hidden markov model and Wasserstein Generative Adversarial Networks for hydraulic system condition monitoring.

Hydraulic systems are vital in industrial settings, and reliable Condition Monitoring (CM) is crucial to preventing failures. This paper introduces a Physics-informed Deep Hidden Markov Model (PiDHMM) combined with a Wasserstein Generative Adversarial Network (WGAN) for enhanced fault detection. PiDHMM improves traditional Hidden Markov Models (HMMs) by embedding physical constraints into state transitions and leveraging a Convolutional Neural Network (CNN) to model emission probabilities and capture complex sensor behavior. To address data scarcity in rare failure modes, WGAN is employed to generate realistic synthetic sensor data. The proposed framework is validated on a multi-sensor hydraulic dataset with known failure events. Comparative results show that PiDHMM outperforms both standard HMMs and deep HMMs without physics constraints, achieving a significant increase in accuracy for fault classification. The inclusion of physics-informed transitions enhances temporal consistency and interpretability, while the WGAN-based augmentation addresses issues of data imbalance, further improving model performance. These results demonstrate that the PiDHMM-WGAN approach offers a more precise, interpretable, and robust solution for hydraulic system monitoring.