An edge oriented architecture for state based remaining useful life prediction in hydraulic systems

Predictive maintenance in hydraulic systems is constrained in practice by the absence of run-to-failure data, limited component-level observability, and disconnected data pipelines between PLC-level operational events and machine-learning-based prognostics. This paper presents an edge-oriented prototype architecture that integrates these elements into a prototype-level state-based predictive maintenance workflow for hydraulic and fluid power systems. The central methodological contribution is a formal state-transition Remaining Useful Life (RUL) formulation, which defines RUL as the remaining time, or number of hydraulic cycles, before the system transitions from nominal to PLC-supported abnormal or critical operation. This definition provides an operational prognostic target without requiring complete run-to-failure trajectories. Supporting this formulation, an OPC UA-based acquisition and ETL framework aligns high-frequency hydraulic sensor data, multi-resolution storage, and PLC-based event logs into a structured data backbone for the investigated prototype. The framework is evaluated on a real PLC-controlled electro-hydraulic prototype operated across 15 repeated sessions at 100 ms resolution. Latent operating states are inferred using K-Means and Hidden Markov Models and evaluated against PLC-based operational reference logs. K-Means achieved an accuracy of 0.737 and a binary anomaly F1-score of 0.938, while the HMM achieved an accuracy of 0.749, a macro F1-score of 0.635, and a binary anomaly F1-score of 0.933. For RUL estimation, XGBoost achieved a test MAE of 3.916 min and a median prognostic horizon of 34.068 min for the first abnormal transition ( R²=0.911 ), while LSTM-Attention achieved the best critical RUL performance with a test MAE of 3.467 min, a median prognostic horizon of 39.662 min, and R²=0.964 . SHAP, PDP, and attention-based analyses indicated that the models relied mainly on hydraulically plausible variables, including cylinder velocity, pressure ratios, oil flow rate, and cycle-level descriptors, supporting physical consistency without claiming causal inference. The results provide prototype-level evidence that structured event logging, unsupervised state modeling, and state-transition RUL estimation can be integrated into a coherent hydraulic PdM architecture under practically relevant PLC-integrated constraints, including the absence of run-to-failure labels and restricted component-level observability.