LAUSR

Abstract Cutting tool is one of the major components to largely influence the overall machining process efficiency. There has been a significant amount of research conducted in developing tool condition monitoring systems with the application of various sensory systems and artificial intelligence techniques for processing and predicting tool wear. Although artificial intelligent techniques have shown significant results in tool condition prediction, the fundamental methodology adopted by these techniques for predicting tool condition still needs to be investigated. This study aims in applying and investigating deep learning techniques for tool condition monitoring in end milling process based on spectrogram features of audible sound collected during the machining process and employing a deep visualization technique to gain the knowledge on inner workings of the deep learning models in the tool wear prediction. Specifically, this study applies Convolutional Neural Networks (CNN) to develop a tool condition monitoring model and performs hyper-parameter tuning to improve prediction accuracy. In addition, the class saliency extraction technique has been used by this study to generate the saliency maps to visualize the specific features utilized by the CNN model for a given spectrogram of audio signal. Furthermore, the study examines the saliency maps of different tool condition levels to understand the intuition of proposed CNN model in selecting the specific localized features of spectrogram for accurate prediction.