Complex Industrial Process Monitoring under High-dimensional Heterogeneous Data

Abstract

The escalating scale of production, advancing level of automation, and the increasing complexity of equipment and systems have put forward higher requirements for the safety, stability, and reliability of industrial process operations. To ensure secure and efficient operations, it is imperative to investigate and devise effective process monitoring methods that can promptly detect process abnormalities and eliminate potential safety hazards.

The massive and diverse process data collected through smart sensors and the industrial Internet of Things have propelled tremendous development and wide application of data-driven process monitoring methods in industrial processes. However, industrial process data often exhibit heterogeneity, encompassing sample heterogeneity and variable heterogeneity. Sample heterogeneity, or nonstationarity, refers to the varying statistical characteristics of samples collected at different times, including mean and variance. Variable heterogeneity refers to the coexistence of continuous and categorical variables, termed as hybrid variables. Continuous variables can take any value within a range, while categorical variables can only adopt a few fixed values. The heterogeneity of process data poses significant challenges for existing data-driven methods, potentially leading to model mismatch issues. Consequently, new methods that can adapt to actual application scenarios are urgently required. This thesis investigates a series of complex industrial process monitoring methods under high-dimensional heterogeneous data, aiming to effectively capture correlations and nonstationary characteristics embedded in heterogeneous data, thereby enabling accurate identification of process anomalies. Chapter 2 focuses on sample heterogeneity, proposing an adaptive monitoring method for nonstationary processes. Chapters 3 to 5 concentrate on variable heterogeneity, aiming at simultaneous modeling and monitoring of hybrid variables. Chapter 6 addresses both sample and variable heterogeneity, developing a recursive model for hybrid variables. The specific research contents are as follows:

To uncover and track stationary sources obscured by nonstationary trends in industrial processes, Chapter 2 proposes an exponential analytic stationary subspace analysis (EASSA) algorithm and develops an adaptive strategy. The EASSA algorithm estimates the stationary sources more accurately and numerically stably by mapping matrices in the generalized eigenvalue problem to their exponential forms. The adaptive monitoring strategy updates the EASSA model to track normal slow changes in nonstationary processes. The proposed method reduces the likelihood of erroneously adapting to incipient faults by constructing monitoring statistics based on stationary sources and updating the model with a batch of samples. Experiments on a simulation process and a real thermal power plant process demonstrate that the proposed method can distinguish real faults from normal changes while maintaining robustness to the disturbances in the nonstationary process.

Chapter 3 proposes a variational augmented mixture discriminant analysis (VAMDA) method to characterize complex dependencies between hybrid variables and accommodate different data distributions for fault detection and diagnosis (FDD). VAMDA specifies a finite mixture model for each class to handle non-Gaussian and non-Bernoulli data. Variational inference is introduced to make VAMDA adaptive to varied distributions. A unified statistical index is designed for hybrid variables, giving VAMDA the capability to distinguish unknown faults. Two case studies demonstrate the effectiveness of VAMDA. The average F1 score of VAMDA is 3.4/3.9 percentage points larger than that of traditional mixture discriminant analysis in the single-mode/multimode Tennessee Eastman (TE) process. In the practical industrial case, when unknown faults are added, the average F1 score of our method is 91.1% and decreases only 1.2 percentage points.

Chapter 4 proposes a unified subspace analysis method to address the curse of dimensionality problem resulting from modeling hybrid variables in high-dimensional observation space. By introducing a common low-dimensional continuous latent variable (LV), the proposed model can capture the dependencies between hybrid variables while avoiding the curse of dimensionality. An analytical Gaussian distribution is derived to approximate the true posterior distribution of the LV, accelerating learning and inference processes. Three statistics are designed with unified probability interpretation for process monitoring, achieving a comprehensive evaluation of hybrid variables. Application to a numerical simulation case and a real-world industrial case shows that the proposed EPLVM successfully detects faults in both binary and continuous variables, achieving the highest FDR in both cases.

Given that industrial process data are usually temporally correlated, Chapter 5 proposes a hybrid probabilistic slow feature analysis (HPSFA) method for dimensionality reduction and dynamic analytics of hybrid variables. The extracted low-dimensional slow features (SFs) are constrained to exhibit slow variations while reconstructing hybrid variables. A variational recursive filter (VRF) is developed for efficiently inferring posterior distributions of SFs. Three statistics are designed based on prediction or reconstruction errors, which successfully separate the dynamic and static changes of the process to achieve fine-grained monitoring. The experimental results show that HPSFA timely detects both static and dynamic anomalies of the hybrid variables, and achieves the highest fault detection rate (85.89%) while maintaining a considerably low false alarm rate (2.67%) in the practical industrial case.

Chapter 6 proposes an input-output hybrid probabilistic slow feature analysis (IO-HPSFA) model and develops an adaptive monitoring method, considering that industrial systems are often adjusted by manipulated variables and their parameters might drift with time. The IO-HPSFA model achieves the simultaneous depiction of systems' internal dynamics and the effects of external inputs in the presence of hybrid input and output variables. An adaptive model updating and process monitoring strategy is designed to track the normal evolution of nonstationary processes, which can discern between real faults and operating condition changes by simultaneously monitoring dynamic and static variations. Case studies validate the effectiveness of the proposed method in distinguishing between real faults and operating condition changes and identifying real fault variables.

Date of Award	27 Aug 2024
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Chunhui Zhao (External Supervisor), Min XIE (Supervisor) & S Joe QIN (Supervisor)