Fault classification of photovoltaic module infrared images based on transfer learning and interpretable convolutional neural network

In the operation of photovoltaic (PV) power plants, infrared cameras are commonly utilized for monitoring the operational status of PV modules. This study focuses on the performance improvement and complexity reduction of convolutional neural network (CNN) when used for fault classification based on infrared images of PV module. By implementing the transfer learning strategy on some famous CNN models, it is observed that the number of convolutional layers has weak impact on the classification results. Therefore, a transfer-learning-based depth reduction approach for CNN models (TLDR-CNN approach) is proposed, and the VGG16 model is employed for verification. Then, a multi-scale feature extraction module (MSFE module) is developed for efficiently replacing the convolutional layers to reduce model complexity and improve classification performance, and several representative model configurations are employed for convolutional layer replacement. Experimental results demonstrate that the application of the developed MSFE module significantly outperforms the baseline model on both classification performance and model complexity. Specifically, the modified model with a reduction of 5 convolutional layers exhibits notable improvements over the training results, with an accuracy increase of 0.90%, precision increase of 0.98%, F1 score increase of 6.89%, and a Matthews correlation coefficient increase of 1.01%. Finally, the interpretability of the above outperformance is also provided by using the Grad-CAM method. The generated CAM images show that the modified model concentrates its weights more on the regions crucial for the model to learn, so the features can be extracted more efficiently.