Data-driven Key Performance Indicator Related Fault Detection Approaches

Process monitoring is an important research area for ensuring the safety and reliability of industrial systems. Thanks to the well-developed modern control theory, the modelbased methods have been deeply studied and a lot of research results have been achieved in the past few decades. However, it is quite difficult to construct accurate mathematical models for the industrial systems such as chemical, metallurgy, bio pharmaceutical and etc., if not impossible, due to the complexity and unknown mechanism of the reactions.Therefore, it is a hard work to promote and apply the model-based methods in such systems effectively. It is worth noting that a large amount of record and measured data are usually available in these systems. Thus, making full use of the abundant data to monitor and control the systems has attracted continued attentions of both academia and industry,and it also contributed to the rapid development of data-driven process monitoring. In the last twenty years, the main task of the data-driven process monitoring method is to detect any fault in the process rapidly and accurately. Then, the stable operation of the system is ensured by identification, isolation and recovery of the fault. However, the latest evidences from academic research and industrial practice indicate that not all of the faults in the process will affect the final product quality. On the contrary, it can significantly reduce unnecessary downtime and maintenance by ignoring the alarms of the faults which have no effect on the final product quality, and then greatly improve the production efficiency.Motivated by this requirement, the so called key performance indicator(KPI) related fault detection method becomes an urgent demand for industrial systems. It is also a research hotspot in the process monitoring community in the recent five years. In view of this, this thesis will carry out the following research contents based on multivariate statistical analysis, to solve the problems of KPI-related fault detection for linear and nonlinear static industrial systems.First, this thesis summarizes the development history and research status of datadriven fault detection methods, and gives the mathematical descriptions of KPI-related fault detection, and points out the shortcomings of the existing research results in the field of KPI-related fault detection: 1) at present, most of the proposed methods for linear systems are based on post processing on partial least squares(PLS) which is very sensitive to outliers and missing values in the modeling data, therefore these methods usually have bad performances in KPI prediction due to the presence of the abnormal data, besides,their fault determination logic is also not perfect; 2) due to the decomposition characteristics of PLS, the performances of the PLS-post-processing based methods are usually not stable when the fault intensity increases, as a result, the false alarm rates of these methods increase significantly; 3) at present, the existing KPI-related fault detection methods for nonlinear systems have obvious defects in the aspects of fault detection performance and fault determination logic; 4) at present, the linear and nonlinear KPI-related fault detection methods implemented in the same decomposition structure are extremely rare and not yet attracted much attention, however, it is very important in simplifying the design procedures of the fault detection system. Considering the above problems, this thesis will propose potentially viable solutions to solve them.Second, to solve the problems of the PLS-post-processing based methods in KPI prediction and fault determination logic, this thesis first introduces a robust PLS algorithm and the expectation maximization algorithm to reduce the influences of outliers and missing values on the PLS model. The singular value decomposition(SVD) is then performed to decompose the process variables space into two orthogonal subspaces that related and unrelated to KPI, respectively. Based on it, a robust KPI prediction and linear KPI-related fault detection approach is designed. Simulation results show that the proposed method is better than the existing PLS-post-processing based methods in the aspects of fault detection performance and fault determination logic, and it has strong inhibition capability on outliers and missing data, thus it is obviously superior to the existing methods in the aspect of KPI prediction performance.Third, to solve the problem of the PLS-post-processing based methods in high false alarm rates for high intensity fault, this thesis proposes two enhanced linear approaches by combining data preprocessing and PLS-post-processing, which first remove the irrelevant components from the process variables space by data preprocessing and then decompose the remaining part of the process variables space into two orthogonal parts,such that the process variables space is finally completely decomposed into KPI-related and KPI-unrelated subspaces. Simulation results show that the proposed methods significantly enhance the performances of the PLS-post-processing based methods, which have extremely low false alarm rates even though the fault intensity is very large.Fourth, to solve the problems of the existing nonlinear methods in fault detection performance and fault determination logic, this thesis proposes two kinds of nonlinear KPI-related fault detection approaches. Firstly, the first method constructs the linear relationship between kernel space and KPI by a kernel partial least squares(KPLS) model. Then, the SVD is performed to decompose the kernel space into KPI-related and KPI-unrelated subspaces, such that a nonlinear KPI-related fault detection approach is designed. The second method fully draws the fruits of the statistical learning methods in the approximation of nonlinear model. By statistical learning modeling, a nonlinear process can be equivalent to a number of local linear models. Therefore, a global nonlinear KPI-related fault detection method can be realized by implementing linear KPI-related fault detection method in each locally model. Simulation results show that the proposed two nonlinear methods are much better than the existing methods in the aspects of fault detection performance and fault determination logic.Fifth, to solve the problem of designing linear and nonlinear KPI-related fault detection methods in the same decomposition structure, this thesis proposes a novel solution for such purpose. Firstly, based on principal component analysis(PCA), a new algorithm that can completely decompose the process variables space into KPI-related and KPI-unrelated subspaces is proposed by using the idea of principal component regression(PCR). Then, a linear KPI-related approach is designed based on the new algorithm.According to the same decomposition structure, a new algorithm for completely decomposing the feature space into KPI-related and KPI-unrelated subspaces is proposed based on kernel principle component analysis(KPCA) by using the idea of kernel principle regression component(KPCR). Then, a nonlinear KPI-related fault detection approach is realized based on this new algorithm. Simulation results show that the proposed linear method is superior to the previous proposed linear methods as well as the existing PLSpost-processing based methods in the aspects of fault detection performance and fault determination logic, while the proposed nonlinear method has the same performance in KPI-related fault detection as the previous proposed nonlinear methods, but it is better than the existing nonlinear methods. More importantly, the same decomposition structure greatly simplifies the design procedures of the KPI-related fault detection system, and it also facilitates the understanding of the algorithms for the users. Therefore, it is more conducive to the promotion and application of the proposed methods in the practice.