Research on Real-Time Weld Spot Quality Detection Based on a Dual-Fusion Backbone Model

Abstract Resistive spot welding is one of the commonly used welding methods in the automotive manufacturing industry. The quality of weld spots affects the reliability and performance of automotive structures. In actual welding processes, weld spot defects may occur due to various influencing factors, posing potential threats to the overall reliability of the vehicle. To enable real-time detection of the weld spot quality in automotive assembled parts, this paper presents a model for real-time weld spot quality detection on automotive weldments, based on the RT-DETR framework, referred to as DCA-DETR. DCA-DETR employs a dual feature extraction fusion backbone design, which effectively reduces computational consumption on shallow feature maps and dynamically adjusts learning weights according to the input features, thereby enhancing the model's adaptability to different features. In addition, a Dynamic Adaptive Convolution (DAConv) module is designed to further improve the model's feature extraction capability. Finally, a Position-Adaptive Attention (PAA) mechanism is proposed to enhance the model's ability to capture local detail features such as weld spot defects. On the self-constructed weld spot dataset, DCA-DETR achieves a 94.8% [email protected], which is 1.8% higher than the baseline model. In addition, the number of parameter count and FLOPs are reduced by 33.1% and 41.7%, respectively, compared to the baseline model. The model's detection speed reaches 91 FPS, which can meet the real-time weld spot quality detection needs of automotive production lines.