Bao Lin

A systematic approach for soft sensor development

This paper presents a systematic approach based on robust statistical techniques for development of a data-driven soft sensor, which is an important component of the process analytical technology (PAT) and is essential for effective quality control. The data quality is obviously of essential significance for a data-driven soft sensor. Therefore, preprocessing procedures for process measurements are described in detail. First, a template is defined based on one or more key process variables to handle missing data related to severe operation interruptions. Second, a univariate, followed by a multivariate principal component analysis (PCA) approach, is used to detect outlying observations. Then, robust regression techniques are employed to derive an inferential model. A dynamic partial least squares (DPLS) model is implemented to address the issue of auto-correlation in process data and thus to provide smoother estimation than using a static regression model. The proposed methodology is illustrated through applications to a cement kiln system for estimation of variables related to product quality, i.e., free lime, and to emission quality, i.e., nitrogen oxides (NOx) emission. The case studies reveal the effectiveness of the systematic framework in deriving data-driven soft sensors that provide reasonably reliable one-step-ahead predictions.

Observer and Data-Driven-Model-Based Fault Detection in Power Plant Coal Mills

This paper presents and compares model-based and data-driven fault detection approaches for coal mill systems. The first approach detects faults with an optimal unknown input observer developed from a simplified energy balance model. Due to the time-consuming effort in developing a first principles model with motor power as the controlled variable, data-driven methods for fault detection are also investigated. Regression models that represent normal operating conditions (NOCs) are developed with both static and dynamic principal component analysis and partial least squares methods. The residual between process measurement and the NOC model prediction is used for fault detection. A hybrid approach, where a data-driven model is employed to derive an optimal unknown input observer, is also implemented. The three methods are evaluated with case studies on coal mill data, which includes a fault caused by a blocked inlet pipe. All three approaches detect the fault as it emerges. The optimal unknown input observer approach is most robust, in that, it has no false positives. On the other hand, the data-driven approaches are more straightforward to implement, since they just require the selection of appropriate confidence limit to avoid false detection. The proposed hybrid approach is promising for systems where a first principles model is cumbersome to obtain.

Optimal component lumping : Problem formulation and solution techniques

This paper presents a systematic method for optimal lumping of a large number of components in order to minimize the loss of information. In principle, a rigorous composition-based model is preferable to describe a system accurately. However, computational intensity and numerical issues restrict such applications in process modeling, simulation and design. A pseudo-component approach that lumps a large number of components in a system into a much smaller number of hypothetical groups reduces the dimensionality at the cost of losing information. Moreover, empirical and heuristic approaches are commonly used to determine the lumping scheme. Given an objective function defined with a linear weighting rule, an optimal lumping problem is formulated as a mixed integer nonlinear programming (MINLP) problem both in discrete and in continuous settings. A reformulation of the original problem is also presented, which significantly reduces the number of independent variables. The application to a system with 144 components demonstrates that the optimal lumping problem can be efficiently solved with a stochastic optimization method, Tabu Search (TS) algorithm. The case study also reveals that the discrete formulation is preferable due to the reduced search space compared to a continuous model formulation.

Data-driven soft sensor design with multiple-rate sampled data: A comparative study

Multi-rate systems are common in industrial processes where quality measurements have slower sampling rate than other process variables. Since inter-sample information is desirable for effective quality control, different approaches have been reported to estimate the quality between samples, including numerical interpolation, polynomial transformation, data lifting and weighted partial least squares (WPLS). Two modifications to the original data lifting approach are proposed in this paper: reformulating the extraction of a fast model as an optimization problem and ensuring the desired model properties through Tikhonov Regularization. A comparative investigation of the four approaches is performed in this paper. Their applicability, accuracy and robustness to process noise are evaluated on a single-input single output (SISO) system. The regularized data lifting and WPLS approaches are implemented to design quality soft sensors for cement kiln processes using data collected from a plant log system. Preliminary results reveal that the WPLS approach is able to provide accurate one-step-ahead prediction. The regularized data lifting technique predicts the product quality of cement kiln systems reasonably well, demonstrating the potential to be used for effective quality control.

OBSERVER-BASED AND REGRESSION MODEL-BASED DETECTION OF EMERGING FAULTS IN COAL MILLS

Abstract In order to improve the reliability of power plants it is important to detect fault as fast as possible. Doing this it is interesting to find the most efficient method. Since modeling of large scale systems is time consuming it is interesting to compare a model-based method with data driven ones. In this paper three different fault detection approaches are compared using a example of a coal mill, where a fault emerges. The compared methods are based on: an optimal unknown input observer, static and dynamic regression model-based detections. The conclusion on the comparison is that observer-based scheme detects the fault 13 samples earlier than the dynamic regression model-based method, and that the static regression based method is not usable due to generation of far too many false detections.

Observer-Based and Regression Model-Based Detection of Emerging Faults in Coal Mills

In order to improve the reliability of power plants it is important to detect fault as fast as possible. Doing this it is interesting to find the most efficient method. Since modeling of large scale systems is time consuming it is interesting to compare a model-based method with data driven ones. In this paper three different fault detection approaches are compared using a example of a coal mill, where a fault emerges. The compared methods are based on: an optimal unknown input observer, static and dynamic regression model-based detections. The conclusion on the comparison is that observer-based scheme detects the fault 13 samples earlier than the dynamic regression model-based method, and that the static regression based method is not usable due to generation of far too many false detections. Copyright

Robust statistics for soft sensor development in cement kiln

Abstract This paper presents a systematic approach of developing data-driven soft sensor using robust statistical technique. Data preprocessing procedures are described in detail. First, a template defined with a key process variable is used to handle missing data. Second, a univariate, followed by a multivariate approach, principal component analysis (PCA), is used to detecting outlying observations. Then, regression technique is employed to derive an inferential model. The proposed methodology is applied to a cement kiln system for realtime estimation of free lime, demonstrating improved performance over a standard multivariate approach.

Product quality estimation using multi-rate sampled data

Abstract This paper investigates different approaches to develop soft sensors from multi-rate sampled data. The data lifting approach consists of two steps, identifying a model with a slow/lifted sampling period and extracting a fast model. Approaches based on direct extraction and linear regression are briefly reviewed, followed by reformulating the task as an unconstrained optimization problem. An illustrative example concerning design of a free lime soft sensor for cement kiln systems is presented. Using data collected from a simulation system based on first principles models, a weighted partial, least squares (WPLS) approach for soft sensor development is compared with data lifting techniques. Case studies reveal the superior performance of the WPLS approach. In addition the product quality for cement kiln systems can be estimated reasonably well, demonstrating the potential to be used for effective quality control.