Multiple characteristic signals are generated when partial discharge is caused by internal insulation defects in gas insulated switchgear(GIS). However

field conditions are often complex

with significant noise interference. In such challenging scenarios

collected signals are prone to noise interference

increasing the difficulty of signal processing and fault diagnosis. To address these challenges

a GIS insulation fault diagnosis method was proposed combining long short term memory(LSTM)with convolutional neural network(CNN)and Bayesian optimization for support vector machines(BO-SVM). The original ultrasonic and ultra-high-frequency signals generated by partial discharge were simultaneously fed into both the LSTM and CNN channels. Feature fingerprints extraction was performed from one-dimensional temporal signals using LSTM and image signals converted from one-dimensional temporal signals to two-dimensional signals using CNN

and the two types of feature values were fused in the fusion layer of CNN. Finally

the executed results were input into the BO-SVM classification layer

and the kernel function γ and penalty parameter C of support vector machines(SVM)were iterated using BO to obtain the optimal solution. Compared to traditional methods that extract features using a single channel

the proposed approach demonstrates significant advancements in terms of accuracy and robustness. The results show that the LSTM-CNN feature fusion and BO-SVM model exhibit notable advantages in insulation recognition performance

achieving an accuracy of 90.34%. This method provides a reliable solution for GIS insulation fault identification based on ultrasonic and ultra-high-frequency signals in complex environments.