协同分布对齐神经网络用于高性能变流器故障定位

Deep learning-based data-driven methods have demonstrated remarkable performance in power converter fault diagnosis. However, the prevalent issue of failure sample dependence, coupled with limited accuracy and robustness, restricts the applicability of these methods in real-world industrial systems. This research underscores a small-sample learning theory that enforcing strict cross-domain feature distribution alignment with shared feature extractors facilitates the extraction of both domain-invariant and fault-discrimination features. These features are conducive to improving fault diagnosis performance. Based on this theory, an asymptotic feature distribution align neural network with the embedded structure and parameter separation training is proposed for high-performance power converters fault location. The embedded structure implements asymptotic multifeature layer constraints to achieve strict cross-domain feature distribution alignment and designs parameter separation training with an asymptotic loss function to improve the training stability. Experimental results on multiple power converters demonstrate the effectiveness of the proposed method. Even under the challenging condition of zero samples, the method achieves excellent fault diagnosis accuracy for multiple open-circuit fault locations. A substantial comparison to state-of-the-art techniques shows the proposed scheme’s superiority in failure sample size, accuracy, and robustness.