Research On Tool Wear Monitoring And Remaining Useful Life Prediction Based On Deep Neural Network Fusion

As a key component of machine tool processing,tool wear directly affects machining accuracy and workpiece surface quality.Accurate tool wear monitoring and residual useful life(RUL)prediction can accurately feedback the machining state and make full use of tool useful life,so as to improve the machining quality and reduce the production cost.The actual machining conditions are complex,and direct monitoring of tool wear is facing great challenges.Establishing an accurate nonlinear mapping relationship between sensor signals and wear values and RUL through deep learning algorithms is an important path to achieve tool wear monitoring and RUL prediction.The amount of tool wear signal data is huge,and the internal features are hidden deeply.How to accurately extract the deep information contained in the data is the bottleneck problem to achieve accurate monitoring and prediction.This thesis takes the wear signal data of milling cutter and turning tool as the research object,improves the network structure through network fusion,enhances the feature extraction ability of the model,and realizes high-precision monitoring and prediction.The main research contents are as follows:(1)Data preprocessing.The signal characteristics and variation rules of the data set are analyzed by means of visualization.According to the characteristics,the samples are preprocessed by invalid data elimination,data downsampling,noise reduction processing,and working condition parameter integration,which lays a data foundation for tool wear monitoring and RUL prediction.(2)A deep network fusion model is constructed to achieve accurate monitoring of tool wear.Aiming at the problem of low monitoring accuracy caused by insufficient wear feature extraction and inaccurate data temporal relationship mining,a bidirectional Gated Recurrent Network(Bi GRU-MHSA)model integrated with Multi-Head Self-Attention mechanism is proposed.The Bi GRU is used to mine the correlation information between the tool wear signals.On this basis,the MHSA mechanism is integrated to fully extract the information details and hidden features,and the supervised learning is used to train the model.The validity and accuracy of the model are verified by comparative experiments.(3)A deep network fusion model and a transfer learning model are constructed to achieve accurate tool RUL prediction.Under the single working condition,aiming at the problem of low prediction accuracy and weak generalization of the model,the Temporal Convolutional Network(TCN-Bot Net)model integrating Bot Net mechanism is proposed.Under the premise of inhibiting gradient disappearance and network degradation,the key feature extraction is realized by Bot Net mechanism.The TCN is used to construct the longterm dependence of time series information,and the mapping relationship between signals and RUL is established.Under the multi working conditions,a transfer learning model based on fine-tuning mechanism is proposed to solve the problem of small sample data scarcity.The common features are transferred from the source domain to the target domain and the parameters are fine-tuned to further improve the generalization ability of the model.A comparative experiment was designed to verify the performance and accuracy of the TCNBot Net model.The NASA data set is used to verify the applicability and advantages of the transfer learning method in tool RUL prediction.(4)The tool wear monitoring experimental system is established.Based on the Python language,the experimental system is built using the Py Charm compiler.The system can monitor the wear value according to the input data and predict the RUL.