Deep Learning Based Tool Wear Prediction Method

In industrial production,the large amount of data obtained on the basis of monitoring and data acquisition systems contains a wealth of historical information about the industrial process,which reflects the specific operation of the industrial process and the problems that arise,as well as the trend in the dynamics of one or more indicators over time.Industrial equipment data is therefore a good indicator of key changes in the industrial process and can thus be used as a basis for fault prediction.In particular,accurate prediction of tool wear can optimise production decisions and avoid losses due to tool wear in high-speed milling processes,where wear affects workpiece quality,productivity and manufacturing costs.In this study,a Feature Extraction and Long Short-Term Memory（LSTM）based tool wear prediction model（FE-LSTM）was first constructed to predict the wear of tool.The FE-LSTM model first analyses and extracts time-domain features,frequencydomain features and time-frequency-domain features embedded in industrial data,and makes full use of the time-series characteristics of industrial data,so that it can learn complex implicit relationships between the extracted features adaptively,and has longdistance dependent learning capability,which can avoid the gradient explosion and gradient disappearance problems.To solve the shortcoming that one-way LSTM cannot take into account the future moment sequence information,a Bi LSTM-AM model based on Bidirectional LSTM and Attention Mechanism was finally built in this study.The bidirectional long short-term memory neural network structure allows the model to have both past and future information,and integrates attention mechanism that allows the model to focus more on the formation of key features during the learning process of long-range information,thus improving the relevance of parameter weights to the training target.In order to validate and evaluate the effectiveness of the proposed tool wear prediction model,three different comparative experiments were conducted in this study based on time series data generated during high-speed milling processes.Experimental results show that the Bi LSTM-AM model reduces mean square error MSE by 70.81%and 47.97% compared to the traditional machine learning models Support Vector Regression（SVR）and Back Propagation Neural Network（BPNN）respectively,and reduces MSE by 79.69% and mean absolute percentage error MAPE by 5.977%compared to the deep learning model Convolutional Neural Network（CNN）.The experimental results indicate that the Bi LSTM-AM model can achieve a more accurate and reliable prediction of tool wear.