Abstract: Welding is a coupling process influenced by many factors, it is not enough to detect the welding quality only by single type of welding data.Moreover, the current welding quality detection model based on deep learning is designed to meet the requirements of industrial real-time and embedded equipment, which does not give full play to the performance of deep learning.Aiming at the above problems, a welding quality detection method based on multimodal tensor fusion was proposed.The welding data set fine division scheme was proposed to solve the difference of welding status classification within class; MobileNetV2 bottleneck module was used as feature extraction unit to reduce the parameters of the model.The welding features of different modes were extracted from each welding modal subnetwork, and then the welding multimodal features were coupled by tensor fusion, and a tensor attention module was proposed, which can suppress a large amount of redundant information generated by tensor fusion.The model was tested on a welding multimodal data set.The results show that the proposed multimodal tensor fusion quality detection model has higher accuracy than the single-modal model, and the tensor fusion and the tensor attention module increase the accuracy of the model.