MTL-TBARNet: a multi-task learning method for transmission line fault diagnosis based on adaptive threshold residual denoising and residual attention mechanisms

Abstract

Abstract Rapid and accurate fault diagnosis is essential for ensuring the stable operation of transmission systems. However, the effectiveness of data-driven models in practical applications is greatly constrained by sample scarcity and noise interference. Furthermore, most current methods rely on separate models for fault classification and location, which fail to exploit the inherent correlation between these tasks. In view of this, we propose a multi-task fault diagnosis framework called MTL-TBARNet which integrates the multi-task learning integrated transformer, bidirectional gated recurrent unit (BIGRU), adaptive threshold residual denoising (ATRD), and residual attention mechanism (RAM) network. This framework integrates ATRD and the RAM, and utilizes the BIGRU and the Transformer encoder to perform classification tasks and location tasks respectively. The ATRD realizes threshold-based denoising by dynamically calculating thresholds and fuses the original input features through residual connections, which can retain the effective information of the original input while eliminating noise. The RAM can focus on the key features of small samples and alleviate the overfitting problem by using the Dropout layer. The experimental verification results show that in the face of different numbers of small samples and scenarios with high noise levels, the MTL-TBARNet exhibits better accuracy and stronger stability compared with the comparison algorithms. Meanwhile, the experimental results verify that the designed ATRD has excellent denoising performance and the designed RAM possesses outstanding performance in the context of small samples.