An online parameter identification scheme for flow quality monitoring in solenoid valves

Abstract

Condition monitoring of solenoid valves is essential for ensuring system reliability across a wide range of industrial applications. Traditional strategies rely either on physics-based models, which often depend on uncertain parameters and fail to capture all failure modes, or on data-driven methods, which suffer from limited interpretability and robustness under varying conditions. This paper proposes a sequential framework that combines model-based parameter identification with machine-learning classification. Key parameters, i.e. coil resistance and piecewise-linear inductance, are extracted from an equivalent RL-circuit using an online Total Least Squares identification scheme. These parameter features serve as inputs to a Feedforward Neural Network classifier, which distinguishes between three flow quality states: Good, Leaking, and Stuck. Its classification performance is validated against a Support Vector Machine with a Gaussian kernel. This sequential approach improves interpretability by grounding the analysis in physical parameter features, and efficiency by reducing the size of training data. Experimental validation is performed on 48 solenoid valves subjected to accelerated lifetime testing. The proposed method achieves high classification accuracy, reaching 91.9% on a balanced test set and 99.1% on an unseen valve. The method shows reduced predictive capabilities in the transition regime between Good and Faulty states, where gradual degradation occurs. An ablation study shows the necessity of inductive features for reliable multi-class fault detection. Only using a resistance-based feature leads to binary fault classification.