Discriminative Visual Feature Learning Of Strong Interference Fusion Welding Image

Fusion welding is an important metal material connection method,which is widely used in the processing of aerospace structural parts,automobile bodies,ship sections and other products.The fusion welding process is characterized by nonlinearity,strong coupling,and high dynamics.How to achieve closed-loop control of fusion welding through monitor the fusion welding process has always been an engineering problem.Computer vision based on deep learning has been regarded as an important means of welding status recognition and defect detection in recent years,and it is a hot spot of current academic research.However,practical engineering is faced with obstacles such as strong interference in fusion welding image,difficulty in learning discriminative visual features,and poor explainability of recognition models.In response to the above issues,the characteristics of fusion welding image are systematically studied.Frontier challenges are proposed from three aspects,including the enrichment of the sample space,the limitation of the hypothesis space,and the enhancement of the explainability.The status quo of visual feature learning of fusion welding image based on deep learning is analyzed and a review study is carried out.This paper conducts research from three levels including data,feature,and model.The main innovations are as follows:1.At the data level.Aiming at the problem of visual interference in welding image data,a method for learning visual features of welding image fusion with time-series data is proposed.An improved three-dimensional convolutional neural network(3DSMDA-Net) with a separable structure and a multi-dimensional attention mechanism is proposed.The network includes a deep separable and lightweight method for three-dimensional convolution kernels,and a multi-dimensional attention mechanism is used to compensate for the loss of accuracy caused by the separation operation.Verification results show that the recognition accuracy of the visual feature learning method used for fusion welding image with time-series data under visual interference is much higher than that of the non-temporal sequential model,the lightweight model volume is reduced to its 1/7,and the multi-dimensional attention mechanism balances the lightweight requirement and the accuracy of the model.This method incorporates the time series characteristics of the welding process,adds time-series information to support model decision-making for a single image,and enriches the sample space.2.At the feature level.Aiming at the problems of the lack of effective visual features of the welding status,the difficulty of extracting discriminative visual features,and the redundancy of visual feature representation,an integrated learning method oriented to diversified,discriminative,and redundant visual features is proposed.A coarse-grained regularization method(CGRCK)oriented to convolution kernels is proposed,which enlarges the differences of the visual features learned by convolution kernels,and prevents the model from overfitting.The label semantic attention mechanism(LSA) is designed,which uses the semantic discrimination between label texts to improve the learning ability of discriminative visual features.A feature fusion method based on long short-term memory(LSTM) neural network is proposed,which enhances useful features and forgets useless features.Verification results show that the integrated learning method for diversified,discriminative,and redundant visual features restricts the hypothesis space from the three independent perspectives of the width,depth,and output of the convolutional neural network,which are better at preventing the model from over-fitting,learning discriminative visual features,and is more adaptive to fuse the redundant visual features.3.At the model level.A class activation mapping explainable modeling method based on multi-scale fused features(CAM-MSFF) is proposed to address the need for transparency in vision-based fusion welding state recognition models.The CAM-MSFF model incorporates a multi-scale supervision mechanism to facilitate multi-scale feature learning and localization of fusion welding image.Three modules of feature compression,feature mapping and feature rescaling are designed to achieve multi-scale feature adaptive fusion.A class activation mapping method based on multi-scale fused features is proposed to obtain a class activation mapping map of the fusion welding recognition model,which explains the decision basis of the model from a visual perspective.Algorithm validation shows that CAM-MSFF has higher recognition accuracy and the decision basis of the model is easier to interpret.Finally,the paper uses four actual fusion welding image datasets(MIG-DT,CO₂-MPD,Laser-PS,Laser-DC) to verify the proposed method with engineering cases.The validation results show that the proposed methods have the advantages of high recognition accuracy,low computational complexity and lightweight model in engineering applications,which can meet the practical engineering tasks of fusion welding condition monitoring and recognition.