Mohd Azlan Hussain

Review of the applications of neural networks in chemical process control—simulation and online implementation

As a result of good modeling capabilities, neural networks have been used extensively for a number of chemical engineering applications such as sensor data analysis, fault detection and nonlinear process identification. However, only in recent years, with the upsurge in the research on nonlinear control, has its use in process control been widespread. This paper intend to provide an extensive review of the various applications utilizing neural networks for chemical process control, both in simulation and online implementation. We have categorized the review under three major control schemes; predictive control, inverse-model-based control, and adaptive control methods, respectively. In each of these categories, we summarize the major applications as well as the objectives and results of the work. The review reveals the tremendous prospect of using neural networks in process control. It also shows the multilayered neural network as the most popular network for such process control applications and also shows the lack of actual successful online applications at the present time.

Adaptive sliding mode control with neural network based hybrid models

In the sliding mode control with a boundary layer approach, the thickness of the boundary layer required to completely eliminate the control input chattering depends on the magnitude of the switching gain used. A controller with higher switching gain produces higher amplitude of chattering and thus needs to use a thicker boundary layer. On the other hand, the value of the switching gain used depends on the bounds of system uncertainties. Hence, a system with large uncertainties needs to use a thicker boundary layer to eliminate chattering. However, the control system is actually changing to a system without sliding mode if we continuously increase the boundary layer thickness in order to cater for systems with large uncertainties. To solve this problem, it is proposed here to use neural networks to model the unknown parts of the system nonlinear functions such that we can obtain a better description of the plant, and hence enable a lower switching gain to be used. The network outputs were combined with the available knowledge, which formed the so-called hybrid models, to approximate the actual nonlinear functions. The controller performance is demonstrated through simulation studies on a two-tank level control system and a continuous stirred tank reactor system. The results showed that the incorporation of networks has enabled a lower switching gain to be used, and thus the chattering in the control inputs can be eliminated even though with a thin boundary layer.

Review and classification of recent observers applied in chemical process systems

Abstract Observers are computational algorithms designed to estimate unmeasured state variables due to the lack of appropriate estimating devices or to replace high-priced sensors in a plant. It is always important to estimate those states prior to developing state feedback laws for control and to prevent process disruptions, process shutdowns and even process failures. The diversity of state estimation techniques resulting from intrinsic differences in chemical process systems makes it difficult to select the proper technique from a theoretical or practical point of view for design and implementation in specific applications. Hence, in this paper, we review the applications of recent observers to chemical process systems and classify them into six classes, which differentiate them with respect to their features and assists in the design of observers. Furthermore, we provide guidelines in designing and choosing the observers for particular applications, and we discuss the future directions for these observers.

The Influence of Ziegler-Natta and Metallocene Catalysts on Polyolefin Structure, Properties, and Processing Ability

50 years ago, Karl Ziegler and Giulio Natta were awarded the Nobel Prize for their discovery of the catalytic polymerization of ethylene and propylene using titanium compounds and aluminum-alkyls as co-catalysts. Polyolefins have grown to become one of the biggest of all produced polymers. New metallocene/methylaluminoxane (MAO) catalysts open the possibility to synthesize polymers with highly defined microstructure, tacticity, and steroregularity, as well as long-chain branched, or blocky copolymers with excellent properties. This improvement in polymerization is possible due to the single active sites available on the metallocene catalysts in contrast to their traditional counterparts. Moreover, these catalysts, half titanocenes/MAO, zirconocenes, and other single site catalysts can control various important parameters, such as co-monomer distribution, molecular weight, molecular weight distribution, molecular architecture, stereo-specificity, degree of linearity, and branching of the polymer. However, in most cases research in this area has reduced academia as olefin polymerization has seen significant advancements in the industries. Therefore, this paper aims to further motivate interest in polyolefin research in academia by highlighting promising and open areas for the future.

Model structure selection for a discrete-time non-linear system using a genetic algorithm

Abstract In recent years, extensive works on genetic algorithms have been reported covering various applications. Genetic algorithms (GAs) have received significant interest from researchers and have been applied to various optimization problems. They offer many advantages such as global search characteristics, and this has led to the idea of using this programming method in modelling dynamic non-linear systems. In this paper, a methodology for model structure selection based on a genetic algorithm was developed and applied to non-linear discrete-time dynamic systems. First the effect of different combinations of GA operators on the performance of the model developed is studied. A proposed algorithm called modified GA, or MGA, is presented and a comparison between a simple GA and a modified GA is carried out. The performance of the proposed algorithm is also compared to the model developed using the orthogonal least squares (OLS) algorithm. The adequacy of the developed models is tested using one-step-ahead prediction and correlation-based model validation tests. The results show that the proposed algorithm can be employed as an algorithm to select the structure of the proposed model.

Developed Hybrid Model for Propylene Polymerisation at Optimum Reaction Conditions

A statistical model combined with CFD (computational fluid dynamic) method was used to explain the detailed phenomena of the process parameters, and a series of experiments were carried out for propylene polymerisation by varying the feed gas composition, reaction initiation temperature, and system pressure, in a fluidised bed catalytic reactor. The propylene polymerisation rate per pass was considered the response to the analysis. Response surface methodology (RSM), with a full factorial central composite experimental design, was applied to develop the model. In this study, analysis of variance (ANOVA) indicated an acceptable value for the coefficient of determination and a suitable estimation of a second-order regression model. For better justification, results were also described through a three-dimensional (3D) response surface and a related two-dimensional (2D) contour plot. These 3D and 2D response analyses provided significant and easy to understand findings on the effect of all the considered process variables on expected findings. To diagnose the model adequacy, the mathematical relationship between the process variables and the extent of polymer conversion was established through the combination of CFD with statistical tools. All the tests showed that the model is an excellent fit with the experimental validation. The maximum extent of polymer conversion per pass was 5.98% at the set time period and with consistent catalyst and co-catalyst feed rates. The optimum conditions for maximum polymerisation was found at reaction temperature (RT) 75 °C, system pressure (SP) 25 bar, and 75% monomer concentration (MC). The hydrogen percentage was kept fixed at all times. The coefficient of correlation for reaction temperature, system pressure, and monomer concentration ratio, was found to be 0.932. Thus, the experimental results and model predicted values were a reliable fit at optimum process conditions. Detailed and adaptable CFD results were capable of giving a clear idea of the bed dynamics at optimum process conditions.

Two Phase Dynamic Model for Gas Phase Propylene Copolymerization in Fluidized Bed Reactor

A two-phase model is proposed for describing the dynamics of a fluidized bed reactor used for polypropylene production. In the proposed model, the fluidized bed is divided into an emulsion phase and bubble phase where the bubble phase flow pattern is assumed to be plug flow and the emulsion phase is considered to be perfectly mixed. Similar previous models consider the reaction in the emulsion phase only. In this work the contribution of reaction in the bubble phase is considered and its effect on the overall polypropylene production is investigated. The kinetic model combined with hydrodynamic model in order to develop a comprehensive model for gas-phase propylene copolymerization reactor. Simulation profiles of the proposed model were compared with those of well mixed model for the emulsion phase temperature. The simulated temperature profile showed a lower rate of change compared to the previously reported models due to lower polymerization rate. Model simulation showed that about 13% of the produced polymer comes from the bubble phase and this considerable amount of polymerization in the bubbles should not be neglected in any modeling attempt.

Optimization and control of polystyrene batch reactor using hybrid based model

Abstract The effects of operating conditions such as initiator and monomer concentration as well as reactor temperature of polymerization reactors have been studied in this work. A recently developed hybrid model for polystyrene batch reactor was utilized in simulation study. The simulation results revealed the sensitivity of polymer properties and conversion to variation of these operating conditions. Furthermore, the study deals with the optimization of batch polymerization reactors. The optimization problem involving minimum time optimal temperature policy has been formulated and solved. Different numerical techniques have been tested and compared. The online control works were performed to validate the optimal temperature profiles. The experimental studies reveal that the calculated optimal policies were able to reduce the batch time keeping the same polymer quality.

Optimal control of batch reactors using generic model control (GMC) and neural network

On-line implementation of the optimal reactor temperature profiles in batch reactors is considered here. The optimal reactor temperature profiles are obtained by solving dynamic optimisation problems off-line to achieve maximum conversion to the desired product. Generic Model Control (GMC) algorithm is used to design the controller to track the optimal temperature profiles (dynamic set points). Neural Network technique is used as the on-line estimator to estimate the amount of heat released by the chemical reaction. A complex reaction scheme is considered in this work to illustrate the ideas. The results clearly show that the GMC controller coupled with a Neural Network based estimator tracks the optimal temperature profiles very well to obtain the desired products on target.

Process Monitoring and Fault Detection in Non-Linear Chemical Process Based On Multi-Scale Kernel Fisher Discriminant Analysis

Abstract This paper presents a multi-scale kernel Fisher discriminant analysis (MSKFDA) algorithm combining Fisher discriminant analysis (FDA) and its nonlinear kernel variation with the wavelet analysis. This approach is proposed for investigating the potential integration of wavelets and multi-scale methods with discriminant analysis in nonlinear chemical process monitoring and fault detection system. In this paper, a discrete wavelet transform (DWT) is applied to extract the dynamics of the process at different scales. The wavelet coefficients obtained during the analysis are used as input for the algorithm. By decomposing the process data into multiple scales, MSKFDA analyse the dynamical data at different scales and then restructure scales that contained important information by inverse discrete wavelet transform (IDWT). A monitoring statistic based on Hoteling’s T 2 statistics is used in process monitoring and fault detection. The Tennessee Eastman benchmark process is used to demonstrate the performance of the proposed approach in comparison with conventional statistical monitoring and fault detection methods. A comparison in terms of false alarm rate, missed alarm rate and detection delay, indicate that the proposed approach outperform the others and enhanced the capabilities of this approach for the diagnosis of industrial applications.