Tool wear prediction in milling CFRP with different fiber orientations based on multi-channel 1DCNN-LSTM

In the machining of carbon fiber reinforced polymer (CFRP) components, tool wear grows rapidly due to the highly abrasive property of carbon fibers, resulting in unfavorable part quality such as delamination and fiber pullout. The tool wear progression, cutting forces, and their quantitative relationship are highly dependent on the fiber orientation. This paper predicts the tool wear progression based on cutting force signals in milling unidirectional (UD) CFRP with the fiber orientation effect. A deep learning model based on multi-channel 1D convolutional neural network (CNN) and long short-term memory (LSTM) is used, with the benefit of considering the unique force variation features due to the CFRP anisotropy. Milling experiments with the fiber orientations of 0 & DEG;, 45 & DEG;, 90 & DEG;, and 135 & DEG; under different feed rates and cutting speeds were performed. The multi-channel 1D CNN obtains the force signal features from individual univariate time series, and combines the time-series information from all channels as the final feature representation in the final layer of the network. The LSTM layer extracts the temporal and spatial characteristics of the dynamic force signals. Moving window and distribution analysis techniques are applied to the predicted tool wear results to reduce the effect of force signal disturbance caused by CFRP inhomogeneity. The model achieves the tool wear prediction with R-2 of 95.04% and MAE of 2.94 & mu;m. Moreover, benefiting from its high feature representation capacity, the proposed method shows higher prediction accuracy for different cutting conditions with more than 25% improvement compared to other commonly used data-driven methods, including 1DCNN, 2DCNN, LSTM, BPNN, and SVR.

Automated summary

This study predicts the tool wear progression in milling unidirectional (UD) CFRP components by using a deep learning model based on a multi-channel 1D convolutional neural network (CNN) and long short-term memory (LSTM). The model takes into account the fiber orientation effect and achieves high prediction accuracy with R-2 of 95.04% and MAE of 2.94μm. It outperforms other commonly used data-driven methods, including 1DCNN, 2DCNN, LSTM, BPNN, and SVR, with an improvement of over 25%.