Wei-Ting Yang

IFAC-PapersOnLine 59(6) p. 127-132 Doi: https://doi.org/10.1016/j.ifacol.2025.07.133

Ensuring safety and efficiency in industrial systems requires effective fault detection and diagnosis, which becomes increasingly challenging in high-dimensional and complex environments. Traditional multivariate statistical process monitoring methods, such as those based on Principal Component Analysis and Partial Least Squares, often fall short in their ability to diagnose localized faults due to their lack of causal modeling. This paper introduces a Causal Network-based Decentralized Multivariate Statistical Process Control (CNd-MSPC) framework, which employs causal networks and community detection—specifically the Leiden algorithm—to segment large systems into functional communities and perform distributed monitoring. This structural partitioning preserves essential causal and topological information, enhancing the sensitivity for fault detection in high-dimensional systems by allowing focused analysis of specific sub-networks. Through extensive testing with a graph-based data simulator, we demonstrate that CNd-MSPC consistently outperforms centralized methods across various network sizes, achieving higher fault detection sensitivity for both process perturbations and sensor biases, especially in large networks. The decentralized approach retains high sensitivity, even when data from several communities are missing due to process disruptions.

Knowledge-Based Systems 300 p. 1-14 Doi: https://doi.org/10.1016/j.knosys.2024.112149

Learning the structure of Bayesian networks (BNs) from data is an NP-hard problem as the solution space grows super-exponentially with the number of nodes. Many algorithms have been developed to efficiently find the best structures, with score-based algorithms being those that use heuristics or metaheuristics to explore potential structures in the search space. This paper proposes a new score-based algorithm that relies on a well-known metaheuristic called scatter search, which, to the best of our knowledge, has not been used in learning BN structure. The core of scatter search is to maintain a reference set that stores both high-quality and diverse solutions, thereby continuously tracking and improving the high-quality solutions, while exploring different search directions indicated by the diverse solutions. By incorporating a distance metric in the learning process, the exploration can be more systematic than purely random, as is often the case in most existing algorithms. The effectiveness and efficiency of the proposed algorithm are evaluated through computational experiments. In addition to learning higher-score structures, the results show that scatter search provides a higher degree of robustness compared with benchmark algorithms

Control Engineering Practice 127 Doi: https://doi.org/10.1016/j.conengprac.2022.105304

Process monitoring is a critical activity in manufacturing industries. A wide variety of data-driven approaches have been developed and employed for fault detection and fault diagnosis. Analyzing the existing process monitoring schemes, prediction accuracy of the process status is usually the primary focus while the explanation (diagnosis) of a detected fault is relegated to a secondary role. In this paper, an interpretable unsupervised machine learning model based on Bayesian Networks (BN) is proposed to be the fundamental model supporting the process monitoring scheme. The proposed methodology is aligned with the recent efforts of eXplanatory Artificial Intelligence (XAI) for knowledge induction and decision making, now brought to the scope of advanced process monitoring. A BN is capable of combining data-driven induction with existing domain knowledge about the process and to display the underlying causal interactions of a process system in an easily interpretable graphical form. The proposed fault detection scheme consists of two levels of monitoring. In the first level, a global index is computed and monitored to detect any deviation from normal operation conditions. In the second level, two local indices are proposed to examine the fine structure of the fault, once it is signaled at the first level. These local indices support the diagnosis of the fault, and are based on the individual unconditional and conditional distributions of the monitored variables. A new labeling procedure is also proposed to narrow down the search and identify the fault type. Unlike many existing diagnosis methods that require access to faulty data (supervised diagnosis methods), the proposed diagnosis methodology belongs to the class that only requires data under normal conditions (unsupervised diagnosis methods). The effectiveness of the proposed monitoring scheme is demonstrated and validated through simulated datasets and an industrial dataset from semiconductor manufacturing.

Expert Systems With Applications 155(113424) Doi: https://doi.org/10.1016/j.eswa.2020.113424

For decades, Run-to-Run (R2R) controllers have been widely implemented in semiconductor manufacturing. They operate over key process parameters on the basis of the metrological measurements acquired from the process and their deviations from the target setpoints. Conventionally, R2R controllers have been implemented independently of the actual equipment condition, which is obviously affecting the process stability and performance. Therefore, both equipment signals and process states shall be considered to make the R2R controllers more robust to the equipment condition drifts. In this paper, we propose a novel physics-informed framework to integrate the real-time equipment condition, based on the Fault Detection and Classification (FDC) data, into the R2R controllers. By utilizing Dynamic Bayesian Networks (DBN), the implicit relationship structure between metrology measurements, FDC indicators, and R2R regulators can be learned and reviewed explicitly. The structure shall be further reviewed to valid with the existing relationships and expert knowledge. Infeasible causalities on the structure will be constrained via setting up the blacklist at the structure learning stage. The proposed framework consists of the offline modeling stage, which incorporates the process, equipment variables, and the expert knowledge in the structure learning, and the online control stage, which constructs the Structured R2R controller (SRC) based on the relationship structure. As a result, the model is consistent by design with empirically known relationships and fundamental physical laws. The proposed SRC not only optimizes the operation with respect to the target control values but also considers the equipment and process states simultaneously. The effectiveness of SRC and the derivative control strategy are validated through a real dataset of a Chemical-Mechanical Polishing (CMP) process, and two simulated studies.

IEEE Transactions on Automation Science and Engineering 17 p. 1297-1306 Doi: https://doi.org/10.1109/TASE.2019.2941047

Virtual metrology (VM) has been widely studied in the semiconductor industry with the purpose of decreasing the cycle time and reducing the expensive metrology measurements. Ideally, a VM model should not only be able to provide accurate predictions but also present an interpretable and rational structure to accommodate fundamental restrictions and relationships that are known to be present in the process. The last aspects have been missing in the VM models proposed hitherto. Therefore, in this article, we propose a novel framework by combining in a single VM model the capability to learn from data with the ability to incorporate the domain knowledge on the process. Thus, the new methodology can use the best of both information sources: data and the subject-matter expert (SME) knowledge. The framework consists of two phases. In the first phase, a Gaussian Bayesian network (GBN) is used to extract the implicit relationships between the metrology and production/process variables. In the second phase, the target response variable is defined, the predictors are selected through the associated Markov blanket, and finally, an empirical model is estimated to accurately predict the response. The proposed framework was tested and its effectiveness was confirmed through a real industrial data from a chemical-mechanical polishing (CMP) process in semiconductor fabrication. The physical meaning of the model obtained was also scrutinized by an SME.