Research On Intelligent Recognition Method Of Tool Wear Based On Big Data

As a direct implementer of the metal cutting process,the tool inevitably has wear problems during the machining process.The change of the state of the tool directly leads to a decrease in product quality and an increase in production cost.Researchers and processing companies have gradually established their own tool monitoring systems for this problem.In the face of the massive data that the system has monitored for a long time,it is necessary to analyze and mine it in an effective way to realize the value of the data itself.Intelligent identification of tool wear status based on monitoring data,on the one hand,can grasp the health status of the tool,guide production and reduce accidents;on the other hand,it can provide theoretical and practical basis for subsequent research.In recent years,scholars have studied the intelligent recognition of tool wear status based on big data analysis methods,and these methods have the following shortcomings in the application of tool wear status identification.:(1)Lack of outliers and noise processing on the massive data monitored.(2)Lack of a "violent" filtering process for massive data,without transforming big data into valuable small data.(3)In the characteristic engineering,there is no analysis of the main and secondary influencing factors of the collected signals under different working conditions,and the characteristic attribute does not contain the working condition data attribute.(4)Single or integrated machine learning models for intelligent recognition are mostly based on strong learner models,lacking differential complementarity for different data attributes.In view of the existing research deficiencies,this paper has carried out the following research work:(1)Build a stable tool monitoring platform and reasonable design experiments,which lays a solid foundation for feature extraction and model integration.(2)The improved quartile distance analysis method(IQR algorithm)and the unbiased risk estimation threshold wavelet method are used to complete the de-existing value and denoising of the monitoring data,and achieved good results.(3)Based on the idea of block algorithm,a data deduplication method based on decision coefficient and relative coefficient is proposed.This method implements a “violent” filtering process and selects the most valuable and representative small sample from the massive data set.(4)In the original data analysis and feature extraction work,the range analysis method is used to study the working conditions of the main and secondary effects of the sensor signal.After analyzing,the spindle speed condition is included in the characteristic attribute.(5)The feature dataset is characterized by principal component analysis(PCA)and recursive feature removal(RFE).The test results are consistent with the test results of all feature datasets.(6)The selected fusion model is fused by the learning fusion strategy.After experimental comparison,the recognition effect of the model fusion is better than the base model and other strong learner models,indicating that the model fusion is effective.