Research On Thermal Error Modeling Of Heavy Duty Machine Tools With Imbalanced Monitoring Data

Heavy-duty machine tools play an important role in machining automation.In recent years,with the advancement of production technology and the increase of product complexity,the machining accuracy requirements have become higher and higher.Research shows that the most critical error source of heavy-duty machine tools is thermal error,so it is necessary to take effective measures to reduce thermal error.Data-driven thermal error modeling methods are widely used in compensation systems to improve the machining accuracy of machine tools.However,in practical applications,the data collected by the machine tools are subject to experimental conditions,resulting in uneven data sets available under different conditions.For the working conditions with less data,if traditional thermal error modeling is performed directly,there will be problems such as low model prediction accuracy and overfitting.Therefore,it is of great significance to study to improve the prediction accuracy of the thermal error model in an uneven data environment.This article takes the heavy-duty gantry drilling machine ZK5540 A as the research object,and proposes a thermal error modeling method based on feature migration for data imbalance under multiple operating conditions,and proposes a thermal error model optimization method based on data enhancement and multi-source migration.The main research work is as follows:(1)Research on the collection,preprocessing and imbalance of monitoring data for heavy-duty machine tools.Firstly,analyze the structural characteristics and heat source distribution of the machine tool,determine the measurement point arrangement of the temperature field,and design an online monitoring data acquisition system.Then for the quality problems in the monitoring data,consider the time and space characteristics of the data comprehensively,use the autoregressive and k-nearest neighbor algorithm to process the outliers,and introduce the comprehensive evaluation index into the locally weighted scatterplot smoothing algorithm to achieve adaptive denoising.Finally,the data imbalance under multiple operating conditions and its impact on thermal error modeling are analyzed to lay the foundation for subsequent research on thermal error modeling in an imbalanced data environment.(2)Research on thermal error modeling of convolutional neural networks based on feature transfer.Aiming at the problem that the thermal error model prediction accuracy is reduced and there is overfitting due to data imbalance under multiple working conditions,the similarity measurement method is first used to analyze the data similarity of heavy machine tools under static,idling and processing conditions to prove that the data under different working conditions is transferable.Then,in order to solve the problem of data imbalance under multiple working conditions,a source domain with sufficient data is used to pre-train the CNN,and the MK-MMD and CORAL domain adaptation methods are used to construct a mapping matrix between the source domain and the target domain to make the source and target domains have the greatest correlation,extract more effective features,and fine-tune the model under the target domain to improve the prediction accuracy of the model.Finally,the validity of the thermal error model of transfer learning is verified through comparative experiments.(3)Research on thermal error model optimization method based on data augmentation and multi-source migration.Due to the lack of processing data of heavy machine tools in practical applications,and the static and idling data sets are similar to the distribution of processing data sets,using only single source migration not only results in waste of resources,but also is not conducive to more comprehensive extraction of feature information and limits the improvement of model prediction accuracy.Aiming at this problem,it is proposed to use the static and idling data domains as the source domains to assist in processing the data domains.During model training,Wasserstein distance is used to update the weights of the source domains to ensure the forward transfer of data knowledge in the two source domains.And during the model training process,the Mixup method is used to enhance the idling and processing data with a small amount of data to further improve the prediction accuracy of the thermal error model.Finally,the validity of the model is verified by comparison experiments.