Prediction of bidirectional milling forces based on spindle current signals by using deep learning algorithms

Abstract

Milling force, which can determine factors such as tool wear and chatter, is the most crucial and accurate variable for monitoring the milling process. However, its application remains limited due to the utilization of expensive and invasive measurement sensors. Predicting the milling force based on low-cost and highly accessible spindle current signals is therefore an effective solution. However, the existing method is generally inaccurate and cannot predict the milling forces in two main directions, namely along the tool feed direction and perpendicular to the tool feed direction. Thus, a data-driven bidirectional milling-force prediction method based on the screened spindle current features is proposed. First, 40 milling process samples including force and current signals are experimentally established. Accordingly, the signals’ time domain, frequency domain, and time–frequency domain features are calculated, and current features with high stability and high correlation with force signals are derived. Finally, a deep learning algorithm Convolutional Network (CNN) combined with Residual Network (ResNet) is established to model the milling force based on the current. In predicting the average milling force using this deep learning algorithm, the percentage error is less than 4% along the feed direction, and less than 10% perpendicular to the feed direction.