Voltage Injection-Based Fault Diagnosis for Power Devices with Multi-Module Neural Networks

This research proposes a machine learning-based method for diagnosing single and compound open-circuit faults in inverter power devices. The objective is to diagnose faults before system operation by injecting voltage space vectors and analyzing the resulting current. A voltage injection pattern is proposed for effective diagnosis, and a multi-module neural network structure based on the injection pattern is employed. The proposed model achieved 100% accuracy on the test set, thus diagnosing switch faults accurately. The model is also implemented in the STM32G431RBT6 digital signal processor, controlling an inverter consisting of a 5A, 600V rated IGBT power module. In operation, the model requires a computation time of 168.84 µs and a model size of 4.698 kB. The results demonstrate superior diagnostic accuracy, memory usage, and computation time compared to existing neural network-based techniques.