Thermal Spot Damage Detection and Classification Based on BS-DeepLabV3+

The accurate identification, quantitative measurement, and damage degree evaluation of thermal spot damage are key issues in photovoltaic module damage detection systems, which are of great significance for fault diagnosis and predictive maintenance of photovoltaic modules. In response to the problems of low accuracy and large model parameters in previous thermal spot damage detection methods, this study proposes an improved lightweight BS-DeepLabV3+ thermal spot detection model, and the model is based on the DeepLabV3+ semantic segmentation model, and its backbone network is improved to the optimized MobileNetV2 to reduce the number of model parameters. SENet(Squeeze and Excitation) attention mechanism is added after AtrousSpatial Pyramid Pooling(ASPP) enabling the model to adaptively select more important thermal spot feature scales and assign corresponding weights;Simultaneously incorporating BiFPN (Bi-directional Feature) into the decoder to retain edge information and suppress noise interference.The Pyramid Network module is designed to adapt to the characteristics of thermal spot damage. The experimental results show that the BS-DeepLabV3+ model proposed in this study can accurately identify thermal spot damage and detect the degree of damage. Compared with the original DeepLabV3+ model, the recognition accuracy of the BS-DeepLabV3+ model has been improved by 2.75%, and the MPA value has been improved by 4.97%. It has achieved significant improvements in key indicators such as accuracy and intersection to union ratio, while effectively reducing the number of parameters, fully verifying its superior performance and lightweight advantages in thermal spot detection tasks with different degrees of damage.