A contribution to online tool wear detection using deep learning methodology

Abstract

Condition monitoring is necessary in machining operation to increase productivity, improve quality and reduce downtime. Tool wear is one of the most common sources of machining problems which occurs due to high temperatures and forces of machining process. Therefore, industry demands reliable tool condition monitoring to address these requirements. This researc investigates a robust on-line tool wear monitoring system in milling operation. Force, vibration and current signals are used as fault indicators to develop a robust model. The main steps of designing an intelligent monitoring system are signal acquisition, signal processing, wear modeling and decision making which all are tackled in different steps of this research. A set of experiments are conducted with K2X10 Huron high speed CNC machine in LIPPS and Dynamo labs of ETS for validation of the research. Moreover, NASA-Ames tool wear benchmark data-set is used for further validation. In the signal processing step, time-frequency transformation is selected to reveal both time and frequency domain characteristics of the signals simultaneously. Wavelet packet transform as a well established algorithm is employed to transform the signals to time-frequency domain. Spectral subtraction method is leveraged on top of the wavelet transform for current signals to remove the steady state part of the signal and magnify fault signatures. Recent developments in machine learning algorithms especially deep learning methods result in significant improvement in automation of various tasks in different industries. Therefore, we employed convolutional neural network (CNN) as powerful deep learning algorithm for modeling the tool wear. Most of the common machine learning algorithms are implemented as well in a comparative approach and it is shown that CNN outperforms the baseline algorithms. Furthermore, this research focuses on scalability of the monitoring algorithms to make them more practical by introducing deep transfer learning in this application. Despite the advantages of machine learning algorithms, one of their main drawbacks is that they have large data requirements. For example, a monitoring system which is trained based on the data from an special machine and model is not reusable on another machine. Therefore, for each machine and task, considerable amount of training data is required. Transfer learning refers to reuse of a machine (deep) learning model which is trained on a problem for a related new problem. This method is successfully implemented and validated for tool wear monitoring and it is demonstrated that by leveraging the proposed framework of this research, robust monitoring algorithms can be achieved even with low amounts of data.

Date	25 May 2020
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Antoine Tahan (Supervisor) & Marc Thomas (Co-supervisor)