Noise introduced during current signal acquisition severely degrades detection accuracy. Under such conditions, the salient features indicative of faults become profoundly obscured, making them difficult to isolate and identify reliably. To overcome this limitation, this study presents a Multi-Head Wide Kernel Deep Convolutional Siamese Network that merges multi-head attention with wide-kernel deep convolutional architecture. The model replaces the initial wide-kernel convolution with a multi-scale design using three parallel convolutional layers of varying kernel sizes, improving multi-frequency feature capture. A Dual Global Pooling Block is added to fuse both global and local temporal information, yielding richer sequence representations. The Singular Value Decomposition Multi-Head Attention module applies low-rank approximation to the attention weight matrix via singular value decomposition, emphasizing relevant features under noise while lowering computational cost. Finally, a Siamese Network projects the extracted features into a high-dimensional space, where classification is refined by measuring similarity between samples. Under composite noise with a signal-to-noise ratio of 10 dB, the model achieves 97.42% accuracy in residual current fault diagnosis, surpassing existing approaches. Additionally, the response time remains below 30 ms, meeting the requirements for real-time fault diagnosis and validating its practical application potential.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号