Industrial Process Monitoring Based On Feature Extraction And Information Fusion

As the increasing demand for profitability and competitiveness in the global marketplace, modern industrial processes are facing a main challenge to maintain process safety and to improve product quality. Therefore, the investigation of process monitoring has its own significant practical and theoretical importance in both industrial and academic circles. In recent years, with the rapid development of computer technologies and the growing size of industrial processes, large amounts of process data can be easily measured and stored while the first-principal model describing the process accurately can be difficult to develop. How to extract meaningfull information that can be used for monitoring the operating conditions from the collected data is an unavoidable issue. This is the result of accelerating the development of data-driven process monitoring technioques in past decades. Particularly, the multivariate statistical process monitoring (MSPM) methods have gained greatly attentions and have become a rich research spot in the filed of process monitoring.Although MSPM has been widely investiaged and plenty of approaches has been proposed, there are some issues realted to feature extraction, dimensionality reduction and multimode process monitoring remain unsolved. In addition, the combination of the model-based and data-based methods to overcome the shotcomings of individual methods is rarely attempted. On the basis of the existing research, this paper proposes several novel monitoring schemes for these different problems, which are mainly focus on feature extraction and information fusion.1) To tackle the problem of information loss in the traditional principal component analysis (PCA) based monitoring approaches, a novel distributed statistical process monitoring method based on four subspaces construction and Bayesian inference is proposed. According to their relevance and irrelevance to principal component subspace and residual subspace, the process variables are divided into four distinct variable subspaces. Each subspace has its own relationship with the PCA model of the original dataset and serves as a low-dimensional representation of the original data space. A monitoring index based on squared Mahalanobis distance is built directly for each subspace and the Bayesian inference is adopted here to obtain a final and unique monitoring index by combining the results of the monitoring statistic in each subspace in a probabilitical manner. From the aspect of induvidal variable subspace, the dimensionality rduction is achieved in a way similar to key variables selection. As a whole, all the process variables are saved without any projection or weighting, to some extent, the information of the original data space remains to maximum. The feasibility and superiority of the proposed method is demonstrated by conducting case studies of a numerical example and the well-known Tennessee Eastman (TE) benchmark process.2) In general, either the global or the local structure of the normal data is changed after an abnormal event occurs, the loss of either global or local structure can deteriorate the performance of fault detetion given that the global structure defines the outer shape of the process data and the local structure presents inner organization. Considering the equal importance of the global and local structure contained in the original data space from the fault detection point of view, we formulate a multi-manifold projection (MMP) algorithm that can extract both global and local structure. After analyizing the objective functions of PCA, locality preserving projections (LPP), global-local structure analysis (GLSA) and local and global PCA (LGPCA) algorithms comprehensively, we point out that those methods all ignore the neighborhood manifold in the objective of global structure preserving. A new global structure preserving function is designed with the neighborhood information embedded, and the MMP algorithm considers neighborhood manifold in both global and local graphs. Based on the MMP algorithm, the developed monitoring model presents satisfied monitoring performance compared to other methods.3) A new independent component analysis (ICA) based non-Gaussian process monitoring scheme is proposed. Firstly, some widely used dimensionlity reduction criteria are detailly described, and it should be noted that it is difficult to determine an optimal criterion without enough prior knowledge. If the criterion is wrongly choosed for modeling, some kinds of faults might not be successfully detected. Secondly, a flexible and feasible dimensionality reduction strategy is developed for the traditional ICA-based monitoring technique through using the idea of ensemble learning in pattern recognition research area. Instead of using only one criterion, the proposed ensemble ICA approach takes advantages of all possible criteria to build several different statistical monitoring models. Finally, the Bayesian inference is utilized to ensemble all the monitoring results together into an unqiue one, and the monitoring performance is illustrated through case studies on a numerical example and the TE benchmark process.4) With respect to the fact that many processes are usually operated under several conditions and processes also present time-vairant dynamic behavior, a systematic solution is proposed for monitoring multimode processes, it brings together mode clustering, adaptive mode identification and adaptive model updation. During the offline modeling phase, a combined strategy of LPI method and ensemble c-cluster algorithm is developed to identify the multimodality of the taining process data with the consideration of no eough prior knowledge for mode clustering and dimension disaster. Subsequently, mode unfolding strategy, which is motivated by batch unfolding technique, is presented for preprocessing the multimode data in order to eliminate the nonlinear behavior resulting from the change of mean and variance. This strategy provides possibability and feasibility for building one global monitoring model for multimode processes. Within the online monitoring phase, mode identification is conducted in an adaptive way by making different assumptions, and monitoring model is also uopdated to meet the time-variant behavior.5) This paper also discusses the application issue of combined data-driven and observer-design methodology in monitoring hybrid process systems. Within the process control circle, a system switching between finite operating modes is often modeled as hybrid process system. An examination of existing literature shows that the model-based methods and the data-driven techniques are usually investigated separately. Firstly, model-based methods have problems in dealing with mode identification, the constriants of necessary consdition or model requirement limit its scope of application. This is replaced by data-drriven techniques for their simplicity given that only recorded process data is required. If the first-principal model is not available, subspace model identification methods can then be utilized to obtain an approximated linear state-space model. Secondly, by taking the advantage of the superiority of model-based methods in fault detection and diagnosis, observer-design approach, which is well investigated for model-based methods, is introduced to build a specific monitoring model for each operating mode. The problems of ambiguous results inheriated in data-driven techniques can be avoided to some extent. The proposed combined methodology is still feasible even if it is applied to monitoring large-sacle processes, like TE process.Finally, conclusions and future research are discussed on the basis of current work.