Infrared thermography serves as a vital tool for monitoring primary substation equipment, enabling real-time temperature distribution sensing and anomaly identification critical for early fault detection. However, in complex substation environments, traditional infrared detection methods exhibit limitations including insufficient localization accuracy, defect misclassification, and elevated false positive/negative rates, constraining their practical utility. To enhance detection accuracy and robustness, this paper proposes and validates an overheating fault diagnosis model for electrical equipment infrared images. The methodology first enriches data distribution through image enhancement and sample expansion to improve scene adaptability. It then introduces a joint preprocessing strategy combining Contrast Limited Adaptive Histogram Equalization and wavelet transform, effectively enhancing detail contrast while suppressing background noise. The network architecture incorporates a Global Attention Mechanism module to strengthen critical feature extraction, boosting performance in multi-class defect identification and localization. Validation on real-world substation infrared imagery demonstrates that the proposed model achieves a Precision of 96.7%, Recall of 94.6%, and mAP 50 of 98.6%. These outcomes represent respective performance enhancements of 5.6, 0.1, and 1.4 percentage points beyond the baseline YOLOv8 model. This work provides a valuable reference for intelligent infrared detection in complex industrial scenarios.