Dynamically Mapped Attention Network for Similar Fault Diagnosis in Cascaded H‐Bridge Multilevel Inverters

ABSTRACT

Due to the topological symmetry of cascaded H‐bridge multilevel inverters, open‐circuit fault features in power switches are highly similar and susceptible to electromagnetic switching noise, significantly increasing the difficulty of fault diagnosis. This paper proposes a fault diagnosis method based on a dynamic Gramian Angular Summation Field (DGASF) and a global joint correlation attention (JCA) mechanism. A parameter‐learnable adaptive DGASF mapping model is developed, in which the mapping scale and phase factors are dynamically learned via a multilayer perceptron to project one‐dimensional time‐domain voltage signals into a high‐dimensional latent feature space. By doing so, the geometric separation between highly similar fault patterns is effectively enlarged in the feature domain. Building upon the mapped representations, a global JCA module is further introduced to exploit the physical coherence of signals through cross‐channel second‐order covariance statistics. In parallel, an asymmetric residual denoising path is incorporated to actively suppress endogenous switching noise that is strongly coupled with fault‐related features, thereby enhancing diagnostic robustness. Experimental results show that, under extreme conditions of 40 samples per class and a signal‐to‐noise ratio (SNR) of 2 dB, this method can achieve a diagnostic accuracy of 79.17%, which is 9.73% higher than existing methods.