Xuefeng Yan

Performance-Driven Distributed PCA Process Monitoring Based on Fault-Relevant Variable Selection and Bayesian Inference

Multivariate statistical process monitoring involves dimension reduction and latent feature extraction in large-scale processes and typically incorporates all measured variables. However, involving variables without beneficial information may degrade monitoring performance. This study analyzes the effect of variable selection on principal component analysis (PCA) monitoring performance. Then, it proposes a fault-relevant variable selection and Bayesian inference-based distributed method for efficient fault detection and isolation. First, the optimal subset of variables is identified for each fault using an optimization algorithm. Second, a sub-PCA model is established in each subset. Finally, the monitoring results of all of the subsets are combined through Bayesian inference. The proposed method reduces redundancy and complexity, explores numerous local behaviors, and provides accurate description of faults, thus improving monitoring performance significantly. Case studies on a numerical example, the Tennessee Eastman benchmark process, and an industrial-scale plant demonstrate the efficiency.

Generalized Dice's coefficient‐based multi‐block principal component analysis with Bayesian inference for plant‐wide process monitoring

Plant‐wide process monitoring is challenging because of the complex relationships among numerous variables in modern industrial processes. The multi‐block process monitoring method is an efficient approach applied to plant‐wide processes. However, dividing the original space into subspaces remains an open issue. The loading matrix generated by principal component analysis (PCA) describes the correlation between original variables and extracted components and reveals the internal relations within the plant‐wide process. Thus, a multi‐block PCA method that constructs principal component (PC) sub‐blocks according to the generalized Dice coefficient of the loading matrix is proposed. The PCs corresponding to similar loading vectors are divided within the same sub‐block. Thus, the PCs in the same sub‐block share similar variational behavior for certain faults. This behavior improves the sensitivity of process monitoring in the sub‐block. A monitoring statistic T2 corresponding to each sub‐block is produced and is integrated into the final probability index based on Bayesian inference. A corresponding contribution plot is also developed to identify the root cause. The superiority of the proposed method is demonstrated by two case studies: a numerical example and the Tennessee Eastman benchmark. Comparisons with other PCA‐based methods are also provided. Copyright

Modified nonlinear generalized ridge regression and its application to develop naphtha cut point soft sensor

A novel modified nonlinear generalized ridge regression (MNGRR) is proposed to model highly nonlinear system. MNGRR applies nonlinear transformation for independent variables to expand independent variable space. Then, the generalized ridge regression (GRR), which employs a modified differential evolution (MDE) to obtain the optimal ridge parameters according to the predicting error, is applied to remove the multicollinearity among the expanded variables, and thus the model that can describe complex nonlinear system and has good predicting ability is obtained. In practice, MNGRR is applied to develop naphtha 95% cut point soft sensor due to the existence of highly nonlinear relationship between process variables and naphtha 95% cut point in atmosphere distillation unit and the fact that few on-line hardware sensors are available and these are also difficult to maintain. Satisfactory results were obtained. The comparison results show that the performance of MNGRR is better than line regressions, nonlinear ordinary least squares regression and nonlinear traditional ridge regression. Further, MDE uses an adaptive mutation operator to overcome the premature and enhance the probability of obtaining the global optimal solution. The comparison results demonstrated that MDEs on-line and off-line performances are all superior to those of traditional DE (TDE), the probability of obtaining the global optimal solution is larger than that of TDE, and that the parameter sensitivity degree of MDE is lower than that of TDE.

Quality Relevant and Independent Two Block Monitoring Based on Mutual Information and KPCA

Traditional process monitoring methods take all the measured variables into account, whereas it will be inappropriate for indicating quality-relevant faults. Some measured variables are independent from the quality variables and these redundancy variables will no doubt degrade the prediction performance of quality variables. This paper proposes a novel quality relevant and independent two block monitoring scheme based on mutual information (MI) and kernel principal component analysis (KPCA). First, all the process variables are divided into two subblocks according to their MI value with quality variables. Then, KPCA monitors the quality-relevant subblock and quality-independent subblock, respectively. When a fault is detected, kernel principal component regression is further utilized to obtain the predicted state of quality variables. Either of the information, whether the current fault disturbs quality-relevant variables or process quality, is necessary and important for engineers. The benefits of MI-KPCA are illustrated through a numerical simulation and the Tennessee Eastman process, and the results reveal the superiority of the proposal compared with some other monitoring methods.

Bayesian Fault Diagnosis With Asynchronous Measurements and Its Application in Networked Distributed Monitoring

Conventional Bayesian fault diagnosis assumes that all measurements are available synchronously; however, this condition does not always hold in practical industry because a process can be characterized by multiple sampling or transmitting rates. This paper introduces a Bayesian fault diagnosis system incorporating both historical and online information to address the asynchronous measurement problem. First, the Expectation Maximization approach is utilized to deal with the historical asynchronous measurements; second, the online incomplete measurements are handled through a Bayesian marginalization method within a moving horizon. Then, a Bayesian diagnosis system revealing both the underlying fault status of the whole plant and the unavailable statuses of the corresponding local units is established, which is more robust for practical application. The proposed scheme is tested on a numerical example, the distributed monitoring problem of Tennessee Eastman benchmark process, and the distributed monitoring problem of an industrial tail gas treatment plant. Monitoring results demonstrate the feasibility and efficiency of the proposed approach.

Batch process monitoring based on multiple-phase online sorting principal component analysis.

Existing phase-based batch or fed-batch process monitoring strategies generally have two problems: (1) phase number, which is difficult to determine, and (2) uneven length feature of data. In this study, a multiple-phase online sorting principal component analysis modeling strategy (MPOSPCA) is proposed to monitor multiple-phase batch processes online. Based on all batches of off-line normal data, a new multiple-phase partition algorithm is proposed, where k-means and a defined average Euclidean radius are employed to determine the multiple-phase data set and phase number. Principal component analysis is then applied to build the model in each phase, and all the components are retained. In online monitoring, the Euclidean distance is used to select the monitoring model. All the components undergo online sorting through a parameter defined by Bayesian inference (BI). The first several components are retained to calculate the T(2) statistics. Finally, the respective probability indices of [Formula: see text] is obtained using BI as the moving average strategy. The feasibility and effectiveness of MPOSPCA are demonstrated through a simple numerical example and the fed-batch penicillin fermentation process.

Data-Driven Distributed Local Fault Detection for Large-Scale Processes Based on the GA-Regularized Canonical Correlation Analysis

Large-scale processes have become common, and fault detection for such processes is imperative. This work studies the data-driven distributed local fault detection problem for large-scale processes with interconnected subsystems and develops a genetic algorithm (GA)-regularized canonical correlation analysis (CCA)-based distributed local fault detection scheme. For each subsystem, the GA-regularized CCA is first performed with its all coupled systems, which aims to preserve the maximum correlation with the minimal communication cost. A CCA-based residual is then generated, and corresponding statistic is constructed to achieve optimal fault detection for the subsystem. The distributed fault detector performs local fault detection for each subsystem using its own measurements and the information provided by its coupled subsystems and therefore exhibits a superior monitoring performance. The regularized CCA-based distributed fault detection approach is tested on a numerical example and the Tennessee Eastman benchmark process. Monitoring results indicate the efficiency and feasibility of the proposed approach.

Independent component analysis model utilizing de-mixing information for improved non-Gaussian process monitoring

We focus on the de-mixing matrix, which is rarely studied in ICA model, to extract data information for fault detection.Multi-block strategy is employed to deal with big data in a novel way.The numerous data is divided through a similarity index Generalized Dices coefficient.Bayesian inference is also employed to combine the results with noise weakened.The way of fault diagnosis is modified with selected variables checked. The de-mixing matrix generated from independent component analysis (ICA) can reveal information about the relations between variables and independent components, but the traditional ICA model does not preserve the whole de-mixing information for the purpose of feature extraction and dimensionality reduction, so that some important information may be abandoned. Multi-block strategy has been improved to be an efficient method to deal with numerous data. However, the manner of dividing original data is still subject for discussion and the priori knowledge is necessary for process division. This paper proposes a totally data-driven ICA model that divides de-mixing matrix based on the Generalized Dices coefficient and combines the results from sub-blocks using Bayesian inference. All information in de-mixing matrix is fully utilized and the ability of monitoring non-Gaussian process is improved. Meanwhile, a corresponding contribution plot is developed for fault diagnosis to find the root causes. The performance of the proposed method is illustrated through a numerical example and the Tennessee Eastman benchmark case study.

A novel select-best and prepotency evolution algorithm and its application to develop industrial oxidation reaction macrokinetic model

A novel evolution algorithm with the select-best and prepotency operator (SPO), named select-best and prepotency evolution algorithm (SPEA), is proposed. The main genetic operators of SPEA are the proposed SPO and the uniform mutation operator. The SPO is defined as follows. Every individual in population has the same chance to select the best individual within its neighborhood range (the select-best range.) and produce new individuals through the crossover with the selected individual. Then the best one of two new individuals is selected as one individual of the next generation. With the SPO that preserves the diversity of individuals to avoid premature and can make excellent individuals be selected many a time, SPEA possesses advantages over conventional genetic algorithms. To compare the performances of SPEA with those of the real-coded genetic algorithm (GA), SPEA and the real-coded GA were applied to search the global optimal solution of a benchmark function. The comparison results demonstrated that SPEA spends less CPU time than the real-coded GA, the on-line, off-line, and local searching performances of SPEA are superior to those of the real-coded GA, and the probability of obtaining the global optimal solution for SPEA is larger than that for the real-coded GA. In addition, the relationship between the select-best neighborhood range and the CPU time consumed by SPEA was analyzed as well as the relationship between the select-best neighborhood range and the ratio of obtaining the global optimal solution. The results demonstrated that the proper ratio of the individual number in the select-best neighborhood to that in the population was 6%. Finally, SPEA was applied to develop a macrokinetic model of the industrial oxidation reaction of p-xylene to terephthalic acid (OXTA) in an Amoco reactor. The macrokinetic model based on the intrinsic kinetics model introduced the correction coefficients into the rate constants of the intrinsic kinetics model to indicate the mass transfer effects presented in the Amoco reactor. Then, with the data of the industrial OXTA, SPEA was employed to obtain the optimal correction coefficients, and the macrokinetic model with high precision for the industrial OXTA was developed.

Optimizing the echo state network based on mutual information for modeling fed-batch bioprocesses

Echo state networks (ESNs) have become one of the most effective dynamic neural networks because of its excellent fitting performance in real-valued time series modeling tasks and simple training processes. The original ESN concept used randomly fixed created reservoirs, and this concept is considered to be one of its main advantages. However, ESNs have been criticized for its randomly created connectivity and weight parameters. Determining the appropriate weight parameters for a given task is an important problem. An optimization method based on mutual information (MI) is proposed in this study to optimize the input scaling parameters and the structure of ESN to address the aforementioned problem and improve the performance of ESN. The MI optimization method mainly consists of two parts: First, the scaling parameters of multiple inputs are adjusted based on the MI between the network inputs and outputs. Second, some output weight connections are pruned for optimization based on the MI between reservoir states. Furthermore, three MI-ESN models are proposed for a fed-batch penicillin fermentation process. Our experimental outcomes reveal that the obtained MI-ESN models outperform the ESN models without optimization and other traditional neural networks.