Thermal Fault Diagnosis of Power Equipment through Improved Instance Segmentation Algorithm and Deep Neural Network Temperature Fitting

Electrical equipment failures caused by abnormal thermal conditions pose a severe threat to the safe and stable operation of power systems. This study proposes a thermal fault diagnosis method for power equipment based on instance segmentation and deep neural networks. Firstly, an attention mechanism module improves the instance segmentation model, increasing the average precision of masks and detection boxes by 2.5% and 3.6%, respectively, achieving highly accurate extraction of power equipment. The model's computational cost is only 12.1 GFLPOs. Secondly, a grayscale-temperature dataset was created, and a deep neural network was developed to accurately model the relationship between grayscale values and corresponding temperatures, surpassing conventional methods. The model achieved an impressive R-square value of 0.9443. Finally, thermal state information is extracted using kernel density estimation (KDE), and an automated thermal fault diagnosis algorithm is developed based on industry guidelines. The experimental results demonstrate that the proposed approach accurately identifies and provides early warning of the thermal state of electrical equipment, thereby enhancing the stability of power systems.