Enhanced brazing defect detection in automotive air conditioning heat exchangers using a novel deep learning transformer model

Abstract Brazing, a critical procedure in heat exchanger manufacturing, is susceptible to defects resulting from inadequate process control. These defects have the potential to seriously impact the heat exchanger’s performance and longevity. Consequently, we propose a novel detection model for brazing defects in heat exchangers, the so-called heat exchanger detection transformer (HE-DETR), using real-time detection transformer (RT-DETR) as the reference model. First, we design a lightweight feature extraction module, namely Re-parameterization Convolution and GhostNet based Cross Stage Partial Network, which enhances the model’s deployability while preserving efficient feature representation. We then redesign the backbone network of RT-DETR based on this module for further enhancement. Second, we adopt the linear deformable convolution to optimise downsampling in the neck network, enabling the model to more flexibly adapt to variations in object scale while simultaneously decreasing computational demands. In addition, we present a new Inner-MPDIoU loss function that enhances the model’s detection performance and accuracy by integrating auxiliary bounding boxes with minimum point distance optimisation. Finally, we created a dataset of brazing defects in heat exchangers and conducted extensive experiments on it to confirm the effectiveness of the proposed method. The experimental results indicate that, in comparison to RT-DETR, HE-DETR decreases the parameter count and computational complexity by 32% and 23%, respectively, while attaining mAP50 of 83.2%. This overall performance surpasses the mainstream YOLO series of object detection models and demonstrates exceptional efficacy in detecting defects in heat exchanger brazing.