Open-Circuit Fault Diagnosis for Charging Modules Based on Transfer Light Gradient Boosting Machine

Deep learning methods have been widely employed to diagnose faults in power converters. However, it is challenging to diagnose multiple faults in two-stage charging power modules. Besides, diagnostic models often excel only under the exact operating conditions on which they were trained. To enable rapid and accurate open-circuit fault (OCF) diagnosis for charging pile power converters operating under varying conditions, this paper proposes an improved Light Gradient Boosting Machine (LightGBM) framework based on transfer learning. Measured waveforms are first segmented via a sliding window, from which eleven concise time-domain features are extracted and fed to the computationally efficient LightGBM for fault classification. To address the prevalent class- imbalance problem encountered in real-world fault data acquisition, this paper proposes a channel-attention-based Wasserstein generative adversarial network with a gradient penalty for data augmentation. Domain adaptation from one working condition with sufficient labelled data to other few-labelled conditions is realized through a novel dynamic re-weighting scheme from the perspectives of instance weights and feature mapping. Furthermore, a new loss function is established to integrate Maximum Mean Discrepancy for aligning the feature spaces of source and target domains, with cross-entropy for reducing the source-domain classification error. Experiments on a fast-charging power module demonstrate that the proposed lightweight method achieves an average diagnosis accuracy of 99.16% for both single- and multi-switch OCFs, and a diagnosis speed of about 13 ms across diverse load and grid conditions. It also achieves an accuracy of over 98.68% in the target condition with merely ten labeled samples, outperforming state-of-the-art alternatives. Moreover, the proposed algorithm maintains robustness under abrupt load transients and severe external noises. Compared to existing deep learning methods and state-of-the-art transfer networks, the proposed method cuts training time by one order of magnitude while maintaining the highest accuracy.