Research On Dynamic Stability Monitoring Method Of Molten Pool Based On Weak-supervised Instance Segmentation

With the wide application of laser deposition additive manufacturing technology in mining equipment,large ships,aerospace and defense industries,the monitoring of its manufacturing process to ensure the quality and performance of the formed parts and reduce the scrap rate has become one of the key concerns of the manufacturing industry.The melt pool is the result of the combined action of the laser beam,the powder and the substrate surface and embodies a plethora of valuable quality information.Therefore,this thesis focuses on the melt pool and carries out the research on the dynamic stability monitoring method of the melt pool based on weak-supervised instance segmentation,and constructs the dynamic stability monitoring model of the melt pool based on MAT-EWMA control chart to realize the early warning of forming defects.The specific research contents are as follows:(1)A weak-supervised learning-based pseudo-annotations generation method for melt pools is investigated.A feature pyramid network is added to the Polygon-RNN framework to fuse low-level details and high-level semantic information.And constraints are introduced according to the characteristics of the melt pool images to facilitate a better synergistic relationship between the network model and the actual melt pool situation,and to improve the accuracy of the method for generating melt pool pseudo-annotations in a targeted manner.The results demonstrate that the proposed method achieves a 73.6% Intersection over Union(IOU)score for generating pseudoannotations of the melt pool.Moreover,the annotation efficiency of the method improves by 76.8% compared to manual annotation.This approach is expected to enhance the efficiency and accuracy of melt pool image annotation,offering a viable solution to the issue of expensive manual annotation.(2)The study focuses on a melt pool feature extraction method based on the pseudo-annotations instance segmentation algorithm.The Mask R-CNN framework is enhanced by employing Mobilenet V2 as the convolutional backbone network and the Vi T model as the classification structure in order to strike a balance between the number of parameters,computation time,and segmentation accuracy.The experimental results demonstrate that the method enhances the test speed by 15.4% while slightly improving the segmentation accuracy.Then,this thesis extracts physical characteristics such as area,length,width,and rear drag angle of the melt pool in association with their physical significance.Moreover,two novel parameters,namely mass center difference and regularity,are proposed to quantify the heat accumulation effect and the degree of regularity of the melt pool shape,thereby providing a more comprehensive monitoring signal.(3)A dynamic stability monitoring model for the melt pool was developed.After performing feature engineering operations such as feature cleaning,processing,and selection,redundant information in the melt pool features is eliminated.Finally,representative features such as mass center difference,regularity,length,width,and rear drag angle are selected.In this thesis,a MAT-EWMA control chart is proposed for correcting the spatial distribution of feature data using a multivariate affine transformation and gradient descent algorithm.An improved Mann-Whitney rank sum test is utilized to compare controllable and runaway melt pool feature samples to address the issue of non-normal distribution,and the test sample from the previous moment is dynamically added to the reference sample to enhance the accuracy and efficiency of anomaly detection.The feasibility of the model is eventually verified through numerical simulation and laser melting tests.The results indicate that the model is extremely sensitive to anomalous data and is capable of detecting anomalies and signaling runaway at a lag of only 2 s,offering broad application prospects and important practical value.