Machine-Learning-Based Joint Defect Prediction Using Temperature Distribution of High-Frequency Induction-Brazed Copper Joints

This study introduces an innovative application of Convolutional Neural Networks (CNN) to predict internal defects in brazing joints of copper pipes, utilizing infrared thermal imaging data. The research focused on high-frequency induction heating brazing of phosphorus deoxidized copper tubes (C1220), a material commonly employed in heat exchangers. By analyzing the temperature distribution during the brazing process, we developed a prediction model for identifying joint defects. A comprehensive dataset comprising 400 sets of 80x80 pixel thermal images was collected across a range of high-frequency currents (28-37A) and voltages (218-359 V) for copper tubes of varying outer diameters (6.35, 9.52 mm, 12.71, and 15.88 mm). The quality of the joints was evaluated based on the penetration depth of the brazing material, with labeling performed on each frame of the temperature data. The model was trained and validated using 39,767 instances of temperature distribution, which included 20,354 samples of normal joints and 19,413 of defective ones. The results demonstrated that the model could not only accurately determine the outer diameter of the copper tubes but also detect defective joints with remarkable precision. The model achieved an accuracy of 99.535%, a recall of 99.536%, and an F1_score of 0.9959, all within an average processing time of approximately 12.29 s. This level of performance signifies a substantial advancement in the field of welding and joining, suggesting that the proposed model can serve as a reliable tool for defect prediction in brazing joints. The implications of this research are profound, offering fresh perspectives in defect detection and quality control in manufacturing processes.