Turning Chatter Detection Using a Multi-Input Convolutional Neural Network via Image and Sound Signal

Abstract

In mechanical cutting and machining, self-excited vibration known as “Chatter” often occurs, adversely affecting a product’s quality and tool life. This article proposes a method to identify chatter by applying a machine learning model to classify data, determining whether the machining process is stable or vibrational. Previously, research studies have used detailed surface image data and sound generated during the machining process. To increase the specificity of the research data, we constructed a two-input model that enables the inclusion of both acoustic and visual data into the model. Data for training, testing, and calibration were collected from machining flanges SS400 in the form of thin steel sheets, using electron microscopes for imaging and microphones for sound recording. The study also compares the accuracy of the two-input model with popular models such as a visual geometry group network (VGG16), residual network (Restnet50), dense convolutional network (DenseNet), and Inception network (InceptionNet). The results show that the DenseNet model has the highest accuracy of 98.8%, while the two-input model has a 98% higher accuracy than other models; however, the two-input model is more appreciated due to the generality of the input data of the model. Experimental results show that the recommended model has good results in this work.