A Convolutional Neural Network Aided Physical Model Improvement for AC Solenoid Valves Diagnosis

This paper focuses on the development of a physics-based diagnostic tool for alternating current (AC) solenoid valves which are categorized as critical components of many machines used in the process industry. Signal processing and machine learning based approaches have been proposed in the literature to diagnose the health state of solenoid valves. However, the approaches do not give a physical explanation of the failure modes. In this work, being capable of diagnosing failure modes while using a physically interpretable model is proposed. Feature attribution methods are applied to CNN on a large data set of the current signals acquired from accelerated life tests of several AC solenoid valves. The results reveal important regions of interest on current signals that guide the modeling of the main missing component of an existing physical model. Two model parameters, which are the shading ring and kinetic coulomb forces, are then identified using current measurements along the lifetime of valves. Consistent trends are found for both parameters allowing to diagnose the failure modes of the solenoid valves. Future work will consist of not only diagnosing the failure modes, but also of predicting the remaining useful life.