Transferable Deep Learning for In-Situ Tool Wear Diagnosis

Abstract

Abstract Emerging deep learning (DL) techniques, which have demonstrated the superior capability to learn complex patterns and interrelations from multivariate data, provide promising solutions to characterize and model complex system that cannot be accurately described by conventional machine learning techniques. Hence, DL techniques have been extensively studied for condition monitoring, diagnosis, and remaining life prediction of manufacturing machine and components. One challenge associated with DL techniques is that the accuracy and reliability of DL models would vary significantly with the data amount, variety, and machine operating scenarios that are used to train the models. If the trained model is applied beyond the training scenarios, an internal covariate shift problem would occur and thereby damage the model reliability. To address this challenge, the DL models should not only extract hierarchical features from the input data, but also study the similarities and differences among data collected from different scenarios and include the discovered similarities in the feature extraction mechanism to generalize models to a broad application. This paper presents a trial to develop a transferable convolutional neural network (CNN) for in-situ diagnosis tool wear severity under different operating conditions.