Cable defect diagnosis based on multi-scale composite-dynamic supervised contrast network

Deep learning has achieved remarkable success in cable defect fault diagnosis. However, in actual engineering, the count of intact cable instances is invariably significantly greater than the flawed sample set, which substantially restricts the evaluation effectiveness of these techniques. Therefore, the document proposes a dynamic normalized supervised contrast network (MCAM-DNSCN) with a multi-scale composite attention mechanism to identify cable defect faults. First, a precision current sensor (TMR) based on tunnel magnetoresistance effect technology is used to capture the weak pulse current signal generated by partial discharge of high-voltage cables. It has sensitivity and low energy consumption characteristics, and shows significant advantages in weak magnetic field detection. Then a multi-scale adaptive feature extractor (MAFE) with bough significance modulation faculty is designed to obtain the phase-resolved partial discharge map (PRPD) and serve as comparative learning to effectively mine the characteristics of TMR partial discharge pulse current signals. Secondly, a multi-scale composite attention mechanism is crafted to adjust the feature weights output by MAFE, thereby improving the precision and confidence of defect identification. Then, an adaptive standardized guided comparison penalty metric for unbalanced conditions is presented, which equilibriums the contribution of each category of samples in the loss calculation through category normalization, and dynamically adjusts the weight according to the accuracy during the training process. Optimize the learning focus of the model and enhance the fault identification ability. Ultimately, the posited MCAM-DNSCN was employed in the diagnosis of cable defects and experimentally compared with other neural network models. The empirical outcomes demonstrated the superior performance of this method.