Time and sensor domain data decomposition and analysis for industrial process monitoring: a novel approach with applications to prognostic and health management of cutting tools

Industrial process monitoring aims to guarantee process safety and improve production efficiency by detecting the happening of faults or predicting the remaining useful life of key components. The data-driven process monitoring technique has been found quite useful across various fields, including chemical, semiconductor, and advanced manufacturing. However, the existing methods still face challenges including the lack of sufficient training data, the complicated time-varying change in process conditions, the redundancy of available signal features, etc. In this study, a novel framework referred to as time and sensor domain data decomposition and analysis, that has potential to be used to address these problems, is developed and validated in application to an essential topic in advanced manufacturing process monitoring, namely the prognostic and health management of cutting tools.

This dissertation contains seven chapters. In Chapter 1, the research background and objectives are described. A comprehensive review of the data-driven tool condition monitoring (TCM) methods are presented in Chapter 2. The novel time-sensor decomposition framework is established in Chapter 3, with a direct application to fault diagnosis of cutting tools in milling processes. Afterwards, three main topics including process anomaly detection, time-varying process monitoring, and process states prediction are studied, respectively. Chapter 4 proposes a time-sensor domain synthesis framework to resolve the problem of insufficient anomaly samples. Chapter 5 develops a recursive time-sensor decomposition approach to adapt to the time-varying trend of process variations. Chapter 6 introduces a local regularisation assisted split augmented Lagrangian shrinkage algorithm to deal with the feature selection problem in the existence of numerous redundant features. Both experiment and simulation studies are conducted to show the the performance of the proposed methods. In Chapter 7, some conclusions obtained during this project and future works are summarised. The prime contributions of this project are summarised as follows.

1. A novel time and sensor domain data decomposition and analysis framework is proposed, which compresses the multi-sensor signals into a lower number of time and sensor domain components (TDCs and SDCs) with dominant information remaining. Both time and sensor domain feature are used to build a tool fault diagnosis model. The results show that the obtained features grasp the dominant information and achieve a stable and optimal performance in different cases.

2. A novel time-sensor domain synthesis framework is designed to provide a solution to unsupervised detection of cutting tool breakage without the involvement of a training data set. The sensor domain information is exploited to determine the baseline cutting passes while the time domain information is used to generate tool condition-related features. The approach can create an unsupervised tool breakage detection model, and adaptively update the threshold to realise accurate anomaly detection.

3. A novel recursive time-sensor decomposition algorithm is proposed by exploiting the temporal dependency among different data snapshots. The obtained SDCs and TDCs can represent the general trend and the overall time-varying behaviours (including both normal and fault-induced variations) of an industrial process, respectively. This enables the creation of novel TDCs-based control chart statistics, and SDCs-based adaptive control limits, ensuring a good trade-off between fault detection rate and false alarm rate.

4. A local regularisation assisted split augmented Lagrangian shrinkage algorithmis proposed under the Bayesian evidence framework, introducing an individual penalty parameter for each model coefficient. During the optimisation process, redundant variables can be pruned accordingly, which significantly reduces the overall computing complexity. In feature selection of process prediction problems, this algorithm method exhibits impressive model sparsity, prediction accuracy, and higher computational efficiency.

5. Extensive milling experiments have been conducted over the period of this project. To validate the effectiveness of the proposed algorithms in real manufacturing processes, all the experimental conditions are designed according to production specifications. During the experiments, common signals used in TCM are collected from the numerical control system and external sensors. The analysis results show that the challenges existing in current process monitoring systems can be effectively addressed under the proposed frameworks.