Embedded-Oriented Open-Switch Fault Diagnosis and Localization for Dual Y-Connected PMSM Based on Deep Learning

The open-switch fault diagnosis speed of multiphase machines is generally no less than 1/4 of an electrical cycle, severely affecting the timeliness of fault-tolerant control intervention. To address this issue, this paper first proposes a cascaded triggering algorithm based on logical judgment, integrating optimized open-phase characteristic values with harmonic subspace current trajectories to enhance the stability of the logic-based method. The proposed data-driven method treats open-phase characteristic values and harmonic subspace currents as input features to reduce the computational burden of extracting open-circuit features in deep learning networks. Additionally, to improve the computational speed of the neural network on embedded platforms, a multi-target classification network is designed based on a simplified gated unit, and a lightweight solving algorithm for the activation function is applied. The dataset used in the deep learning network is labeled with the assistance of a designed Autoencoder while preserving the transition data between pre- and post-fault states. The effectiveness, accuracy, and fault localization speed of the proposed algorithms are validated through the dataset and the experiment on the motor bench. The results demonstrate that the proposed data-driven method can accurately locate the open switch fault within 10 sampling periods