Research On Continuous Optimization Method And Strategy Of Cutting Database And Model Based On Transfer Learning

With the rapid development of cutting technology and the wide popularization of computer technology,cutting database is becoming more and more important in today’s manufacturing industry.A good cutting database system can store a large number of relevant cutting data,and on this basis,it can realize digital management and make full use of manufacturing resources.In this paper,based on the product characteristics and industrial demand of modern manufacturing industry,a cutting database system for this industry is established.The research is mainly carried out from three aspects:the establishment of database system,the establishment method of neural network cutting force prediction model and the optimization method and strategy of prediction model.The establishment of the cutting database system includes the design of the database and the development of the application program.Based on the functional requirements and data requirements of the database,the conceptual structure and logical structure of the database are designed.After collecting the relevant data of cutting materials and tools,the database is established.This project uses C++and Python programming language to complete the development of the application program.The interactive functions between the application program and the database system are realized by C++language,and the prediction function of cutting force and other important process parameters is completed by Python language.The whole database system is developed by both of them.In order to reduce the data demand of neural network prediction model,this paper studies the training method of neural network prediction model.Combined with the relevant concepts and methods of transfer learning,the knowledge in simulation data is proposed to help improve the prediction performance of neural network prediction model in real cutting data,so that when the same error rate is achieved,less data can be used.In this paper,we combine the Fine-tuning method and multi-core MMD distance strategy to establish a neural network training method based on simulation data and real cutting data,and test the performance of the method through experiments.The experimental results show that,in a suitable range,in the same number of training samples,compared with the traditional BP neural network,this method can effectively reduce the prediction error rate of the network.By combining transfer learning method with cutting experiment design,the optimization strategy of prediction model is studied.In order to study the impact of different transfer methods on network performance,based on the two methods of Fine-tuning and multi-core MMD distance,this paper introduces a strategy mechanism,that is,in the network training process,according to the different input samples,the number of Fine-tuning layers of the network is dynamically determined,and Gumbel softmax sampling method is used to solve the derivation problem of discrete variables.According to the principle that the difference of machining conditions increases gradually,the cutting experiments are divided into three control groups.Each control group includes two groups of cutting experiments.The workpiece materials include aluminum alloy,aluminum lithium alloy and titanium alloy,The cutting tools include 3-edge cemented carbide cutting tools and 4-edge cemented carbide coated cutting tools.The control group was divided into source domain and target domain to ensure that the source domain of the three control groups were the same experimental group.The MMD distance is calculated for the data sets of the source domain and target domain of each control group,and the difference between the data sets of the control group is verified.Using the three sets of experimental data sets,from less to more set the sample number of training set,FNN,Fine-tuning and Policy Fine-tuning neural network are trained respectively.FNN is the traditional BP neural network,and the latter two are two different transfer network methods.By comparing the prediction error rates of the two kinds of transfer networks and FNN networks,the influence of transfer learning method on the prediction performance of neural network is summarized in the case of different processing conditions.According to the experimental results,the optimization method and strategy of neural network prediction model are summarized based on the classification principle of processing conditions difference.