Deep Learning Driven Tool Wear Identification and Remaining Useful Life Prediction

Abstract

Tooling condition monitoring and prognosis has been researched and developed extensively considering its importance of supporting manufacturing systems. An effective system for tooling condition monitoring and prognosis can ensure good product quality, minimisation of tooling failure and optimisation of production cost. However, due to the complex mechanism in manufacturing processes, the tooling condition depends on different factors (i.e., machining parameters and tooling/product materials). Therefore, the conventional experiment-based methods are expensive, time-consuming, and less effective. In recent years, the industrial Internet of Things (IoT) related technologies and the state-the-art artificial intelligence algorithms (e.g., deep learning algorithms) have been increasingly applied to manufacturing applications for mining valuable information from massive amounts of data, thus making complex manufacturing processes easier to be understood and improved. Therefore, in this thesis, based on multiple sensors and deep learning algorithms, a series of novel and systematic methods for tooling condition monitoring and prognosis for a machining system are proposed. The aim of this research is to enhance the overall performance of tooling condition monitoring and prognosis in terms of computational efficiency, prediction accuracy and system robustness.

This research covers the two primary indicators for tool condition monitoring and prognosis, that is, tool wear identification and tool remaining useful life (RUL) prediction. For tool wear identification, in this research, based on multi-sensor signals, a new method, i.e., recursive feature elimination and cross-validation (RFECV), is designed to optimise feature selection and fusion. In the process of RFECV, a support vector machine (SVM)-based classifier is used to recursively evaluate the contributions of different feature subsets to the classification of the tool wear for optimal feature selection. Dimensionality reduction on the selected features is further implemented via the isometric mapping (Isomap) method. Finally, a concise and robust 1D convolutional neural network (CNN) model is devised to perform tool wear identification using the newly generated features from the Isomap.

To improve accuracy and expedite computational efficiency for predicting the RUL of cutting tools for a machining system, a novel methodology is designed. The methodology integrates strategies of signal partition and deep learning algorithms for effectively processing and analysing multi-sourced sensor signals collected throughout the lifecycle of a cutting tool. In more detail, the methodology consists of two subsystems: (i) a Hurst exponent-based method is developed to effectively partition multi-sourced signals along with the tool wear evolution; (ii) a hybrid CNN¬LSTM (convolutional neural networks-long short-term memory) algorithm is developed to combine feature extraction, fusion and regression in a systematic means to facilitate the prediction based on segmented signals.

In order to meet the needs of small and medium-sized enterprises (SMEs), in this work, an economical, three-level, and multi-sensor-based monitoring and prognosis system is designed. In the system, a wireless acquisition platform collects multi-sensor signals. To expedite computational efficiency and minimise data traffic, an edge computing (EC)-based gateway is embedded as an extended Kalman filter for signal denoising, performing the conversion of the time-series data format to reduce the high latency of transmission. A hybrid CNN-RF (random forest) model is devised to achieve real-time tool wear identifications. Furthermore, based on the received signal image, the system can conduct RUL prediction based on an integrated multi¬channel CNN-LSTM model at the cloud computing end.

Two open-source experimental datasets and a workshop deployment case were used to verify the practicability and reliability of the methods and systems proposed in this research. Regard to the method for tool wear identification, it was verified that the prediction of the selected feature subsets achieved the best prediction accuracy. Moreover, the proposed RFECV-SVM approach was proved to be superior to other machine learning models on feature selection. For the tool RUL prediction, the data partition method was developed for assigning the signal data corresponding to each tool wear stage, and case studies were used to compare the designed hybrid deep learning algorithm with some other main-stream algorithms, such as CNN, LSTM, DNN (deep artificial neural network) and PCA (principal component analysis), under the conditions of partitioned and un-partitioned signals. The conducted performance comparison showed that, the methods proposed in this research is essentially better than those of the comparative algorithms. In addition, the developed EC-enabled tool prognosis system was validated that it can provide a comprehensive and satisfactory diagnosis result, in terms of tool wear identification and tool RUL prediction. The EC effectively reduced the amount of data, improved transmission efficiency and enhance the data privacy of the conventional IoT system. The proposed system was proved to be effective, flexible and affordable.

Date of Award	Dec 2020
Original language	English
Awarding Institution	Coventry University
Supervisor	Weidong Li (Supervisor), Xin Lu (Supervisor) & Sheng Wang (Supervisor)