Online chatter detection using self-evolving automated machine learning fusion

Abstract

Computer Numerical Controlled (CNC) Milling is used to remove excess metal from a blank to produce the final shape of the workpiece. One limitation of high material removal rates with reduced cost in machining operations is self-excited vibrations called chatter. Chatter results in poor surface finish, damage to the workpiece and machine tool’s spindle, and causes accelerated tool wear. Chatter detection and prevention have been one of the major fields of study in manufacturing to improve the quality and productivity of machining operations. This research proposes a self-evolving, online method to detect and avoid chatter in milling operations by fusing deep learning with the knowledge of chatter theory. The process is monitored by collecting vibration data during machining. Windows of data are converted into Short-time Fourier Transforms and processed through a Convolution Neural Network to identify five machining states: Air cutting, entrance to and exit from the cut, stable cut, and unstable cut with chatter. A base state detection model is built by modifying and training AlexNet architecture using experimental data with known states. A specialized Deep Learning architecture is designed based on the base model, using an Automated machine learning process, with high state detection accuracy and low complexity in mind. In parallel to the machine learning model, chatter detection with a physics-based model is executed to increase the robustness and accuracy of chatter detection. The forced vibrations which always occur in milling are removed by Kalman filter, and the occurrence of chatter is detected using an energy-based method. The hybrid system detects the chatter with a 98.8% success rate. The system has a built-in online self-improving capability. The system stops the machining process and commands a new spindle speed to force the machine to operate in chatter-free zone.