Lining surface defects adversely affect boiler operation and efficiency, and pose safety risks. The existing crack detection methods for refractory materials lack intelligence and quantitative precision. To address these limitations, well-designed crack detection methods are required for defect identification and maintenance. An explainable AI (XAI) model was developed to detect lining surface defects in a 700 MW circulating fluidized bed (CFB) boiler in Yunnan, China. The hybrid model combines deep learning (i.e., faster region-based convolutional neural network, FRCNN) and image processing to automatically identify and classify defects. To enhance explainability, FRCNN and Vision Transformer (ViT) are used to highlight the contribution of each feature to defect identification. The experimental results show that the hybrid model has high precision and recall rate in defect detection. A relative improvement of 28.01% in overall detection performance is achieved by the proposed ViT-FRCNN hybrid model, in comparison with the single detection method FRCNN. Meanwhile, an overall detection precision of over 95% is attained by the same hybrid model. This work supports CFB boiler operation and maintenance and offers innovative defect detection approaches for related fields. Future work will focus on lightweight Transformer adaptations to enhance computational efficiency for real-time industrial deployment. • Innovative XAI model for detecting surface defects in 700 MW CFB boilers. • Integration of FRCNN and ViT for enhanced defect identification. • Improved detection efficiency by 28.01% and accuracy over 95%. • Provides visual explanations for defect identification results. • Offers new ideas for defect detection in related industrial fields.
Liu et al. (Sun,) studied this question.