Research On Milling Tool Wear Condition Monitoring Method Based On Transformer

Based on the nature of manufacturing,vigorously promoting the combined development of manufacturing and intelligent technology is of importance to accelerate the upgrading and evolution of manufacturing industry.As an important end of machining,the identification of the wear state and the timely replacement of the milling tool have an important influence on the machining efficiency of the machine tool and machining accuracy of the workpiece.Compared with the traditional tool wear condition identification method which relies on manual experience,the research on tool wear condition monitoring method based on sensor signals collected in the process of machining is of great significance to fully tap the tool processing potential,effectively guide industrial production,and drive the green and intelligent transformation of manufacturing industry.Therefore,based on the multi-sensor data collected in the milling process,this paper carried out a research on the monitoring method of milling tool wear state.The main research content of this paper is as follows:The evolution of multi-sensor signals in single milling and tool life cycle were analyzed by visualization and piecewise mean splicing methods;Aiming at the problems of large amount of heterogeneous multi-source sensor signal data,redundancy and low information capacity of single data point,three-axis vibration sensor signal was selected as the subsequent data source,and its data validity was verified by converting vibration signal to audio signal;Based on the data preprocessing methods such as invalid signal elimination and data subinterval division,and combining with the short-time Fourier transform method,the extraction process of time-frequency domain characteristic data was established,which not only reduced the length of time series data,but also improved the information capacity of one single time step.The data set for three-fold cross validation was established,which laid a solid data foundation for model training and validation.In view of the insensitivity of convolutional neural network to time step sequence and the complex structure and low training efficiency of recurrent neural network,a tool wear state monitoring method based on Transformer was proposed combining the characteristics of three-axis time-frequency characteristic input data;the model effects under different hyperparameter settings were compared,and the optimal setting was determined;visual analysis was used to verify the effective encoding of the input data by the self-attention mechanism,and the effect of the position encoding module on the model was discussed;under the condition of three-fold cross-validation,the effectiveness of the proposed method in monitoring data of different wear stages was verified,and the advantage of the proposed method in training accuracy and training efficiency was proved by comparison with the recurrent neural network model.Aiming at the phenomenon that Transformer model had poor classification effect on some validation sets,a tool wear state monitoring method based on Transformer-CNN was further proposed on the basis of visual analysis of output data of model encoder.A decoder model combining the data normalization method and 2D convolutional neural network was established;the influence of different data normalization methods and encoder model was analyzed by feature extraction pretraining method.Finally,the training and verification of the Transformer-CNN model was carried out on the whole training set,the results showed that the model achieves the overall accuracy of 95.2%,especially for the data in the serious wear stage had good identification ability.