In order to improve the generalization evaluation accuracy and the interpretability of existing partial discharge depth diagnosis models

the focus coefficient and its composite index are proposed. The class activation mapping (CAM) of the training sample is calculated by improved gradient-weighted class activation mapping (Grad-CAM++)

and the focus coefficient is obtained by dividing the weighting coefficient by the adjustment factor after the convolution of the CAM and phase resolved partial discharge (PRPD). On this basis

the composite index of evaluation accuracy and focus coefficient is constructed

and the comprehensive evaluation of the final model is carried out. Seven kinds of multi-source defect PD data are collected through the 110 kV true gas insulated switchgear (GIS) platform

and 12 in-depth diagnosis models are constructed. Six sets of test datasets are constructed from experimental data and field data to verify the effectiveness of the evaluation method proposed in this paper. The experimental results show that the focus coefficient can effectively quantify the visual analysis results of the CAM and improve the confidence of the diagnostic results. The generalization correlation of the composite index (γ=10%) built upon the comprehensive focus coefficient is 81.01%

which is 8.74% higher than that of the accuracy index. The accuracy of the diagnostic model selected by composite indicators is 97% under the field-collected unfamiliar dataset

which is 9.1% higher than that of the traditional optimization method.