A new embedded vision system for monitoring tool conditions in production lines using a combination of direct and indirect methods

Abstract

Despite the high development of tool condition monitoring systems (TCMS), their assessment in harsher conditions, such as those found in production lines, remains an open question. To face this challenge, a methodology consisting of two independent hardware devices was proposed. The first consisted of sensors for indirect monitoring (electric current and vibration). The second was a new vision system, which was embedded in the automatic tool changer of the machine tool. It was demonstrated that the embedded vision systems provided clear images of step drill which, when used to train a convolution neural network (U-Net), allows the systems to achieve a 91 % accuracy. In addition, the use of feature vectors extracted from the electric current and vibration signals for training a traditional machine learning algorithm (random forest) achieved a 95 % accuracy. When the two approaches were fused, the accuracy increased to 96 %.