Tool Wear Prediction Method Based on Attention Mechanism

The wear state of a tool affects the surface quality and processing stability of the workpiece；hence，accurate monitoring of its wear amount has a positive role in ensuring processing reliability and maintaining the continuity of production and processing. To further improve the generalization performance and accuracy of the tool wear prediction model，a tool wear prediction method based on a multiscale convolutional bidirectional gated recurrent unit-attention(MSCBGRU-A) neural network is proposed here，which integrates the attention mechanism and is composed of feature expansion，multiscale convolution，bidirectional gated recurrent unit(GRU)，attention，and regression modules. First，cutting force，acoustic emission，and vibration signals are taken as input signals. Next，these signals generate the tool wear output characteristic maps of multiple scales through the multiscale convolution module. The characteristic maps output by multiple convolution channels are input to the connection layer to connect in two ways，head-to-tail and stacking，to obtain two kinds of output data. Then，the two kinds of output data are input to the bidirectional GRU and attention modules，respectively. In addition，the bidirectional GRU module learns the dynamic changes of the output characteristic graph to obtain the time series characteristics. Next，the attention module assigns weights to the output of the multiscale convolutional neural network to strengthen the characteristics that contribute more to the tool wear prediction results. Finally，the regression module predicts the wear value of the tool. Furthermore，the mixed domain attention mechanism，that is，the convolutional block attention mechanism (CBAM)，is introduced through comparative experiments to obtain the MSCBGRU-CBAM model. It is proved that the attention mechanism can adaptively focus on features more relevant to tool wear by drawing an attention weight map of CBAM. Compared with other deep learning models，the MSCBGRU-CBAM model has the highest prediction accuracy. Compared with MSCBGRU models without the attention mechanism，the root mean square error decreases by 19.3%，MAE decreases by 17.7%，and R2 increases by 2.7%. © 2024 Beijing Institute of Technology. All rights reserved.