Milling Tool Wear Status Identification And Prediction Based On Deep Learning

With the development of information technology,the degree of automation and process centralization in machinery and equipment manufacturing is increasing,and the products are changing from "manufacturing" to "intelligent manufacturing".To meet the high efficiency and quality of the products produced,the manufacturing system needs to be smart enough to monitor the manufacturing process.As the most sensitive and vulnerable part of the manufacturing process,tool wear has a direct impact on the quality and accuracy of the workpiece,and can cause huge economic losses and unexpected safety problems if not replaced in time for scrap.Therefore,early detection and monitoring of the wear process is necessary for cost optimization and maintenance safety of the machining process.In this paper,we propose a tool wear status identification and prediction model based on deep learning method with milling cutters as the research object,and verify the superiority of the proposed method through experiments.The main research contents are as follows.(1)The current status of domestic and international research on tool wear status identification and tool wear status monitoring and prediction is introduced,the shortcomings of the current model based on deep learning methods are analyzed,and the PHM2010 high-speed milling tool data set is introduced;the tool wear mechanism is analyzed,and the force signal,vibration signal and acoustic emission signal are determined as the monitoring signals of the tool wear monitoring and prediction model,and the cutting force signal is used as the monitoring signal of the tool wear The cutting force signal is used as the monitoring signal for tool wear identification.(2)The above monitoring signals are extracted in the time domain,frequency domain and time-frequency domain,and the strong correlation feature signals with tool wear are screened,the screened features are normalized,and the screened features are dimensionalized by the principal component analysis(PCA)method to obtain ten feature spaces that are highly sensitive to tool wear.The above features are the input for the machine learning comparison model in Chapter 4 and Chapter 5;the continuous wavelet packet transformation is introduced to make the theoretical basis for the conversion of 1D signal to 3D time-frequency image in Chapter 5.(3)In order to monitor the tool wear during cutting process and predict the future wear state in real time,a tool wear monitoring and prediction method based on a custom Dense Net and GRU integrated model with multi-sensor feature fusion is proposed;considering the influence of multi-sensors on tool wear,a heterogeneous asymmetric convolution kernel is added to the model to explore the strong correlation features between signal and tool wear;to explore the influence of historical information on tool In order to explore the influence of historical information on tool wear,an "expansion" scheme is used in the one-dimensional convolutional kernel;the current popular machine learning methods are compared with various deep learning methods for monitoring and predicting tool wear values,and the high accuracy of the proposed method for tool wear monitoring and prediction is demonstrated.(4)A data augmentation-based tool wear state recognition framework is proposed,which incorporates deep learning methods such as generative adversarial networks,maximum mean difference,continuous wavelet transform,and convolutional neural networks;conditional Wasserstein generative adversarial network with gradient penalty is proposed to generate artificial samples;a generative data evaluation metric is proposed,and the generated high-quality data is The original data samples are expanded to obtain the enhanced data;the enhanced data set is transformed into timefrequency images using continuous wavelet packet transform;a convolutional neural network based on the Inception-Res Net framework is proposed for tool wear status identification.It is demonstrated experimentally that the method has a higher accuracy compared with the method of unbalanced data.