Reliability diagnosis and fault prediction technique for three-phase inverters using artificial neural networks

Abstract

This study proposes an artificial neural network-based diagnostic technique for diagnosing the aging status of the Insulated Gate Bipolar Transistor (IGBT), a core component of a three-phase inverter system installed in a stratospheric solar drone, and predicting their remaining useful life (RUL). Due to environmental constraints that make it difficult to collect actual operating data, 15,625 training data sets were constructed through MATLAB/Simulink-based inverter system simulations that reflect various IGBT aging conditions. The learning data was processed to use the change rate of the inverter three-phase voltage and voltage components (positive sequence, negative sequence, zero sequence) according to the symmetric coordinate method, based on the change in the on-state resistance (Ron) value of the IGBT as input feature data. Based on this, we implemented and trained a classification model to predict the Risk Level and a regression model to predict the Combined Aging Index. The classification model achieved 94.49% overall accuracy and 97.61% accuracy for severe aging prediction, while the regression model attained a coefficient of determination (R2) of 0.8718, RMSE of 2.56, and MAE of 1.84, demonstrating strong predi0ctive performance. Feature contribution analysis further identified the reverse-phase voltage change rate as the most critical indicator, accounting for 34.2% of the predictive power in inverter aging diagnosis. This study verified the robustness of the model through simulations that reflected a noisy operating environment. Future validation using operational data and the inclusion of various failure modes can enhance the generalization performance of the predictive model. The proposed technique can be used as a diagnostic technology for predictive maintenance and assurance of reliability in aircraft inverter systems, which require high reliability.