This paper proposes a convolutional neural network (CNN) -based insulation defect detection model for electrical equipment for high-voltage system safety monitoring. This method organizes image representation, multi-scale feature extraction, region localization, category discrimination and state mapping into a unified computing link for insulation defect recognition and risk expression for monitoring. Methods: A data set of 8640 images including insulators, casing, cable terminals, and lightning arresters was constructed. Six labels were set: crack, ablation, flashover attachment, electrodischarge burn, edge denuding, and normal surface, and three safety states: normal, concern, and alarm. The network combined residual convolution, feature pyramid fusion and two-branch prediction to complete regional positioning and category discrimination, and monitored the state through the output of the state mapping unit. Results: The model achieves 95.1% precision, 93.8% mAP, 92.6% recall, and 41 ms inference delay. Conclusion: The results show that the framework has a stable detection and monitoring ability under the conditions of complex background, multiple equipment objects and weak texture defects, and can provide a basis for inspection sequencing and alarm tips, which has good engineering adaptability.
