Research On Injection Molding Process Optimization Of Thin-walled Parts Based On Transfer Learning

Injection molding is a complex thermal-rheological molding process,many factors can affect the final injection molding quality,especially for thin-walled parts,which are more prone to many quality defects such as warpage and deformation.Integrating finite element(CAE)software,DOE experimental methods,proxy regression models,and optimization algorithms play a vital role in avoiding a lot of disorderly exploration,shortening cycle time,reducing costs,effectively improving the quality of plastic parts,and minimizing defects in plastic parts.However,this method will fail when new data is predicted based on the original data surrogate model.Due to the differences in structure,materials,etc.between the new product and the original product,the process data of the two are different,so the original model cannot be applied to the process optimization of the new product.For complex thin-walled injection molded parts,using numerical simulation alone cannot effectively reduce production workload and improve quality.However,there are deviations between the physical environment settings such as device structure,stroke,time,and speed response,and the ideal simulated environment settings,resulting in a huge "gap" between actual and simulated data.Therefore,the surrogate model based on the numerical simulation data cannot accurately predict the actual production data.In addition,the existing optimization algorithms have certain defects.For example,the search efficiency is low,and it is easily "premature",and converges to local extremes.For these reasons,the following research is conducted in this paper to address the data differences and the shortcomings of the optimization algorithm.(1)This paper presents an in-depth study of the complex molding mechanism of thinwalled injection molded parts.Taking a thin-walled interior part of a car as the object,the process parameters of holding pressure and temperature and the mechanism of warpage of the injection molded part under their interaction are studied separately.The range of major factors and process parameters is determined by simulation pre-analysis.The extreme difference analysis and analysis of variance were performed to determine the influence of each factor on the molding quality and to obtain a preliminary process solution.The correctness of the mechanism analysis is verified,and a theoretical basis is provided for the next data collection and process optimization.(2)To address the problem of catastrophic forgetting that cannot be accurately predicted when the fitted source model is applied to new distributed data,model-based(Regular Transfer NN,Finetune)and instance-based(Tr Ada Boost R2)migration learning methods are used in this study,respectively,and the feasibility of these methods is verified by the process dataset of block injection molded parts.The experimental results show that Finetune migration is the most effective among the three methods and also has a better stability.Finally,a detailed study of Finetune method is conducted to compare the migration effect of retaining different pre-training information,and the results show that the amount of retained pre-training information should be chosen according to the size of the difference between the source and target domains.(3)For the problem that the optimization-seeking algorithm is easy to fall into local extrema,this paper proposes the multi-domain orthogonal space co-evolution(MDOSE)algorithm based on the ideas of orthogonal design and particle swarm search algorithm.The optimal-seeking ability of the algorithm is verified by calculating the global optimal extrema of the standard function CEC2017.The results show that MDOSE has a better and faster finding ability for single-peak,multi-peak,and mixed functions than several other traditional optimization-seeking algorithms.Finally,MDOSE is used to optimize the process of a thinwalled interior part of a car.The results show that the Z-directional warpage deformation of the part is reduced from 1.022 mm to 0.0857 mm,and the optimization effect reaches 91.6%,which has a very good process optimization effect.(4)Integrating these methods mentioned above to build a set of process recommendation systems conforming to manufacturing makes it easy for users to manage process cases,input source domain data,a small amount of target data to fit the target model,and provide a process solution conforming to the requirements for users to use for reference,saving costs such as mold trials in production.Finally,the practicality of the system is verified with the simulation data of U-shaped thin-walled injection molded parts deformation and the actual data.