APPLYING DEEP LEARNING TO DETECT THERMAL DEFECTS IN BUILDING ENVELOPES

Currently, China has many energy-intensive buildings, where thermal defects in the envelope structure are responsible for a considerable amount of energy wastage. However, traditional methods for detecting these defects have limitations when it comes to large-scale and precise detections. Therefore, we propose a thermal defect detection system for building envelopes based on deep learning, which leverages the exceptional image analysis capabilities of deep learning to provide efficient and accurate services for thermal defect detection. Through case examination, our system achieves a Pixel Accuracy (PA) of 93.311%. The paper also analyzes the impact of different types of datasets on detection efficacy and concludes that a multimodal dataset combining infrared images and visible maps yields superior recognition outcomes. Furthermore, verification confirms that integrating infrared images into the multimodal dataset significantly enhances the recognition rate of thermal defects compared to other categories. The findings from this study provide valuable insights into advancing and applying thermal defect detection technology in building envelopes, offering a novel approach and methodology for energy-saving retrofits and detecting thermal defects in existing buildings.