Failure Prevention Based on Principal Component Analysis and Machine Learning for Wireless Power Transfer Systems

Wireless Power Transfer is growing in popularity as it represents an attractive charging method in many applications. As such, monitoring these devices to detect soft faults (which can possibly lead to destructive failures) becomes an important task to ensure continuous operation of these systems and to reduce unavailability periods. This paper proposes a study regarding soft fault diagnosis on an LCC-S compensated Inductive Wireless Power Transfer system. The method can be easily applied to any compensation topology. First, a significant dataset is generated in simulation, applying a Monte Carlo approach. The compensation components values are randomly varied inside the tolerance, to simulate a nominal operating condition, and outside the tolerance, to simulate the component aging. Principal Component Analysis is applied to an initial set of 27 features extracted from three voltages measured on the circuit. Different classification algorithms are then applied to the Principal Component Analysis outputs, to identify the system state and locate the possible faulty component. The results show that a low-complexity Feed-Forward Neural Network can achieve over 90% classification accuracy, exploiting only the first six Principal Components, corresponding to around 80% data compression with respect to the initial 27 features.