Research on mechanism-based modeling and migration learning for cable local discharge diagnosis under small sample conditions

In actual environments, it is challenging to obtain multiple types of local discharge signals, and diagnostic models built on such small data sets suffer from severe overfitting issues. To address the issue of small samples, a new partial discharge pattern recognition method based on the local discharge mechanism model and migration learning is proposed. First, a mechanistic model of cable local discharge is developed and diverse local discharge samples are generated. Second, the one-dimensional local discharge signal is converted into a two-dimensional colour image using the Grammer angle field transformation method to fully exploit the temporal characteristics of the local discharge signal. Third, the diagnostic knowledge acquired from the simulated data is then applied to real-world scenarios using parameter transfer strategies and convolutional neural networks (CNN). Finally, a real-world case of local discharge diagnosis is then utilized to validate and discuss the efficacy of the proposed method. On the basis of simulation data and the CNN parameter transfer strategy, the results demonstrate that the method can learn from more generalizable aspects, resulting in a significant advance in the performance of local discharge diagnosis.