Research On Quality Prediction Of Laser Cladding Process Based On PSO-BP Network And Multi-modal Information Fusion

As one of the core technologies in additive manufacturing,laser cladding technology realize the rapid formation of complex metal components.It can satisfies the high-precision equipment manufacturing demands that are difficult to achieve under traditional manufacturing processes.Furthermore,laser cladding technology is also an indispensable part of green remanufacturing technology,which is of great development significance for the extension of the service life of parts and the improvement of sustainable development.Laser cladding process is accompanied by a large number of complex physicochemical changes,and the melt pool contains a lot of key characteristic information reflecting the forming quality.Therefore,how to monitor the melt pool and predict the forming quality through the characteristic information of the melt pool so as to achieve early warning has become a research hotspots in domestic and foreign fields.In this study,the melt pool of the laser cladding process was taken as the object,and an off-axis visual monitoring system for laser cladding was built to achieve the effective extraction of the region of interest and its features in real time under complex background;On this basis,the prediction study on the morphology quality and dilution rate quality of laser cladding forming was carried out,which can lay a foundation for guiding the process parameters to optimize the processing quality and improve the forming accuracy.The specific study contents areas follows:(1)A melt pool region of interest selection and its feature extraction method based on improved YOLOv3 network is proposed.According to the requirements of high-precision localization and fast detection of the melt pool ROI in the laser cladding process,we propose a method to extract the melt pool region of interest based on the improved YOLOv3 network,and lightweighting the backbone part of the YOLOv3network by using group convolution,thus reducing the amount of parameters of the YOLOv3 network by nearly 50%,in addition,the PANet structure is introduced in the Neck part of the network to improve the localization accuracy of the model,finally,based on the improved YOLOv3 target detection network performs well on the melt pool region of interest detection task,the m AP@0.75 is about 94.33%,with a detection time of 14.75ms.The extracted melt pool region of interest is used to obtain melt pool features by gamma transform,adaptive threshold segmentation,and Canny edge detection.According to the bounding box of the melt pool region of interest detected by the improved YOLOv3 algorithm,combined with the characteristics of the temperature distribution in the melt pool area,the redundant background temperature photographed by the thermal imager was eliminated,and the temperature field in the melt pool area of interest was obtained.(2)A time seires prediction model of melt pool feature for laser cladding process is established.According to the characteristics of melt pool feature timing signals with drastic fluctuations,the melt pool feature time series prediction model is established based on VMD-DAGRU,which improves the feature expression capability by decomposing the fluctuating and complex melt pool feature information through VMD,and improves the feature extraction capability of the model by DAGRU which introducing the dual attention mechanism and GRU unit,and performs well in the prediction task of melt pool features.Compared with LSTM,TCN,DARNN and other models,the VMD-DAGRU model performs the best in melt pool feature time series prediction.(3)A quality prediction model for laser cladding forming geometry based on PSO-BP neural network regression is proposed.The regression model between melt pool characteristics and forming geometry quality is established by BP neural network,and the model is optimized by PSO particle swarm algorithm to improve the accuracy of regression prediction.The prediction errors of width and height of the optimized BP neural network model after optimized by PSO algorithm are 0.0775 and 0.065,respectively,the MSE errors are 0.0103 and 0.0047 respectively,and the~2 are 0919and 0.8842 respectively.The results show that this PSO-BP neural network is effective in predicting the geometric quality of the cladding layer forming.(4)A quality prediction model for laser cladding forming dilution rate based on feature-level multimodal fusion is proposed.According to the problem that the single modal data of the melt pool cannot characterize the forming quality well and the prediction model has poor robustness and low accuracy,a convolutional neural network based on Inception structure is established as the feature extractor of the multimodal information of the melt pool,which can realize the effective mining of the detailed feature information and the global feature information inside the melt pool,which further realize the richer expression of the melt pool on the quality of the cladding layer;A feature fusion module based on MFB-FC is established to fuse the multimodal features extracted by Inception structural convolutional neural network with the physical features of the melt pool to enhance the correlation between different modal features,provide better information complement to the network,and further enhance the expression of the fused features on the forming quality of the cladding layer.After comparison,the proposed feature-level multimodal fusion model is found to have good performance in dilution rate quality prediction,with an accuracy of 93.33%on the dilution rate classification prediction task and an overall average accuracy of 93.79%.