As a core of power electronic systems, three-phase voltage source inverters' power devices are prone to open-circuit faults, causing system waveform distortion and load abnormalities. Traditional diagnostic methods suffer from low accuracy and poor real-time performance. To address this, this paper proposes a real-time diagnosis and location method based on a wavelet-Transformer fusion network. It uses wavelet transform to extract multi-scale time-frequency features of fault signals and combines Transformer's self-attention mechanism to mine temporal correlation. Data preprocessing, feature selection, and comprehensive location strategies optimize model performance. Experimental results show that the proposed wavelet-Transformer fusion network achieves 95.6% overall diagnostic accuracy, 96.2% fault location accuracy, and an average end-to-end response latency of 8.4 ms under the main test setting. Compared with single-module base-lines, the proposed framework provides a better balance between diagnostic accuracy and online deployment efficiency, offering a practical solution for intelligent operation and maintenance of power electronic equipment.

wavelet transform, transformer model, inverter, power device, fault diagnosis