Tool Wear Prediction Based On Multi-Channel One-Dimensional Convolution And Attention Mechanism

With the development of the manufacturing industry,various machining technologies have become more mature and intelligent.In the milling process of workpieces,tools are indispensable components.Tool wear will directly affect the quality of workpieces or even lead to processing accidents.Therefore,this paper proposes a tool wear prediction method based on deep learning,which predicts the tool wear according to the sensory data collected during the milling process.The staff can choose whether to replace the tool according to the predicted wear value to avoid the accidents.The main research contents of the paper are as follows:To solve the problem that tool wear dataset may include abnormal data due to manual measurement errors.This paper proposes an abnormal data detection method based on LSTM auto-encoder.The LSTM auto-encoder is used to learn the information between normal data samples and reconstruct the entire dataset.The data are arranged in descending order of reconstruction error.The top k largest reconstruction error data do not conform to the information of normal data.So they are regarded as abnormal data.The experimental results show that the reconstruction error of the LSTM auto-encoder under the three tool wear datasets has reached 10^-3,and the three anomaly detection indicators（precision,recall,F1 value）are all over 80%,which proves the effectiveness of the method.Considering the multi-dimensional and the large-scale characteristics of the milling data,this paper proposes multi-channel one-dimensional convolution to extract features.The features of different dimensions are sent to the corresponding channels for one-dimensional convolution.Then,a channel attention mechanism is added to perform weighted multi-feature fusion.From the time dimension,the one-dimensional convolution performs feature extraction and data down sample.From the multi-feature dimension,the channel attention mechanism performs weighted multi-feature fusion by their influence to tool wear.After feature extraction,the LSTM module is used for time series prediction.Last,two dense layers are used for logistic regression to get the predict tool wear value.Compared with the traditional tool wear prediction method,the tool wear prediction method based on the deep learning model doesn’t require manually design features,doesn’t require too much expert knowledge.Instead,it used an end-to-end model to obtain the predicted wear value.Compared with the traditional deep learning model,the proposed model considers the time series characteristics and multi-dimensional characteristics of the sensory data,and can process multi-dimensional long-term series data as well.It is demonstrated in the experiment results that the loss values（MAE,RMSE）of the proposed model for all tools are lower than 2.52μm（MAE）and 4.90μm（RMSE）,prospectively.Such results outperform traditional CNN/LSTM model to prove the novelty of this model.