Centre for Intelligent Systems Research

Process Modelling and Analysis Lab

Control of Polymerisation Batch Reactor

Polymer products, or plastics, are used in numerous industries including construction, manufacturing, electronics, transportation, food processing, and aerospace. Despite the wide range of polymer applications, polymerisation reactor control is still a very challenging task as the polymerisation reaction is complex and nonlinear in nature. Controlling polymerisation reaction variables relies on whether they can be measured, estimated, or measured with some time delay. One of the major difficulties encountered in polymerisation reactor control is the lack of reliable online real time analytical data. Typically, temperature is used as an intermediate variable to control polymer quality, as the quality and quantity of polymer are directly dependent on reactor temperature.

Figure 1: Lab scale polystyrene batch reactor
Figure 1: Lab scale polystyrene batch reactor

Figure 2: Polystyrene batch reactor model in Matlab Simulink
Figure 2: Polystyrene batch reactor model in Matlab Simulink

The aim of this research is to develop advanced nonlinear controllers for monitoring polymerisation batch reactors that provide a smooth, safe and waste free production line, as well as high quality polymer products. Three advanced nonlinear controllers have been designed and implemented in a real polystyrene plant. The three controllers are an artificial Neural Network-based MPC (NN-MPC), an artificial Fuzzy Logic Controller (FLC) and a Generic Model Controller (GMC). The performance criterion, Integral absolute error (IAE) is used as cost function in the optimisation process to tune the controller parameters. The proposed controllers are tested with various implementations including the optimal temperature batch recipe and process disturbance rejection. Experimental results reveal that the NN-MPC is superior at tracking the optimal reactor temperature profile without noticeable overshoot as observed in the case of a FLC or GMC.

Figure 3: Optimal setpoint tracking using three different advanced nonlinear controllers. Reactor temperature is the controlled variable whereas heater power is the manipulated variable.
Figure 3: Optimal setpoint tracking using three different advanced nonlinear controllers. Reactor temperature is the controlled variable whereas heater power is the manipulated variable.

back to top

 

Simulation-based Learning for Control of Complex Conveyor Networks

The demands placed on material handling systems continue to grow with increasing design complexity and higher throughput requirements. In complex engineered environments, such as a baggage handling system, thousands of bags must be tracked through the system and delivered to the appropriate location once the bags have cleared security requirements. These systems are highly dynamic, with time varying traffic demands and changing flight schedules. Furthermore, complexity is evident in the stochastic behaviour of these systems as the path traffic will take is dependant on the processing outcomes as traffic traverses the system.

The focus of this research is to address the problem of directing traffic flows within a complex conveyor network. Items entering the system have different processing requirements, priorities and exit points, combined with dynamic process flows and the stochastic nature of the system to create an interesting problem for analysis. A generic algorithm has been developed that is applicable to such a system and is able to learn the most appropriate method to manage the traffic flows, ensuring correct processing and delivery.

A generic flow chart of a BHS
A generic flow chart of a BHS

back to top

 

Development and Application of Hybrid Soft Computing Models

Soft Computing (SC) is an inter-disciplinary area this is well-suited for the design and development of computerized intelligent systems. The main SC models include artificial neural networks, fuzzy systems, and evolutionary algorithms (to name a few). Each SC model, however, has its own benefits and limitations. As a result, this research focuses on the design and development of hybrid SC models, e.g. neural-fuzzy, neural-evolutionary, fuzzy-evolutionary, neural-fuzzy-evolutionary paradigms, with the aim of capitalising on the strengths of each SC model and, at the same time, alleviating the associated shortcomings.

An example of an evolutionary-based neural-fuzzy model is shown in Figure 1. Other hybrid SC models with online learning capabilities have been researched and developed. We have also applied the resulting hybrid SC models to a number of complex real-world problems. These include content-based image retrieval (Figure 2), fault detection and diagnosis of motors and condition monitoring of industrial systems/processes (Figure 3), as well as typing biometrics (or keystroke dynamics) and medical decision support problems (Figure 4).

A hybrid model combining a neural-fuzzy network and a genetic algorithm
Figure 1 - A hybrid model combining a neural-fuzzy network and a genetic algorithm

Retrieval of skin disease images
(a) Retrieval of skin disease images
Retrieval of satellite images
(b) Retrieval of satellite images
Figure 2 - Hybrid soft computing models for content-based image retrieval

Fault detection and diagnosis of motors
(a) Fault detection and diagnosis of motors
Condition monitoring of complex systems
(b) Condition monitoring of complex systems
Figure 3 - Hybrid soft computing models for fault detection and diagnosis of motors and condition monitoring of complex industrial systems/processes

Typing biometrics (keystroke dynamics)
(a) Typing biometrics (keystroke dynamics)
Online medical decision support
(b) Online medical decision support
Figure 4 - Hybrid soft computing models for typing biometrics and medical decision support problems

back to top

 

Video driven traffic modelling

A video driven modelling technique is introduced for traffic systems, where video processing is employed to estimate metrics such as traffic volumes. These metrics are used to update the traffic system model, which is then simulated using the Paramics traffic simulation platform. Video driven model tuning has widespread potential application in traffic systems, due to the convenience and reduced costs of model development and maintenance.

Traffic simulation
Traffic simulation

back to top

 

Constructing Optimal Prediction Intervals for Load Forecasting Problem

Short Term Load Forecasting (STLF) is fundamental for the reliable and efficient operation of power systems. Application of artificial intelligent-based techniques and in particular Neural Networks (NNs) has proliferated for STLF within the last two decades. Despite this, one should notice that NN models are deterministic, and by that, their application for predicting future of stochastic systems, such as loads, is always in doubt and questionable.

The objective of this research is to construct Prediction Intervals (PIs) for future loads instead of forecasting their exact values. Different techniques are applied for constructing reliable PIs for outcomes of NNs. Statistical measures are developed and applied for quantitative and comprehensive evaluation of PIs. According to these measures, a new cost function is designed for shortening width of PIs without compromising their coverage probability. Evolutionary optimisation techniques are used for minimisation of this cost function and adjustment of NN parameters. Demonstrated results clearly show that the proposed method for constructing PIs outperforms the traditional techniques.

PIs constructed for test samples using the delta method (top) and optimized delta method (bottom)
PIs constructed for test samples using the delta method (top) and optimized delta method (bottom)

back to top

 

Simulation Modelling of Pedestrian Way-finding Behaviour under Normal Situation

Our aim is to develop a comprehensive conceptual model of pedestrian way-finding behaviour under normal and non-panic conditions. To gain a deeper insight, it is necessary to investigate the requirements for inclusion of behavioural theories and knowledge from human dynamics, cognitive science and psychology. Research is investigating how to merge the strength of the two most plausible pedestrian modelling paradigms, Social Force Model and Discrete Choice Model with the computational efficiency of discrete event simulation. Our model consists of three main elements: environment representation, agent characteristics and behavioural rules. Initially, a 2D space with walls, one obstacle, an attraction, entrance and exit has been designed. Two entities with similar characteristics have been introduced. Currently, speed of the entities is constant and the position is manipulated by attractive and repulsive forces. Attractive force is applied around attractive spot in the environment and also another attractive force is used to motivate the entities to move to reach the target as well as repulsive force, which is for obstacle avoidance. In this study, so far the early stages of modelling have been developed and trajectories of entities are captured.

Movement trajectory of two pedestrians with different motivations while moving towards a destination
Movement trajectory of two pedestrians with different motivations while moving towards a destination

back to top

 

Uncertainty Quantification through Construction of Prediction Intervals

As universal approximators, Neural Networks (NNs) have achieved great success in many regression problems in the last two decades. No matter how NNs are trained or used, they suffer from two basic deficiencies: susceptible to the uncertainty, and, no indication of their accuracy. To cope with these deficiencies, construction of Prediction Intervals (PIs) for NN outputs has been proposed in literature. Although NN-based PIs have been investigated in literature for almost two decades, there are many issues left unarticulated related to them: (i) PI assessment, (ii) PI optimization, and (iii) reducing the computational requirements.

Motivated by these gaps in literature, this research attempts to develop practically useful measures for quantitative evaluation of PIs. Secondly, a new cost function is developed based on these measures and is minimised in order to find the optimum values of some critical parameters of NN models. The optimisation algorithm attempts to develop NNs leading to narrower PIs without compromising their coverage probability. Finally, a new method is proposed for PI construction, which does not require calculation of the large matrices, as they are required by other methods such as the delta and Bayesian techniques.

Terminology and concept of prediction interval
Terminology and concept of prediction intervalPrediction intervals for targets
Prediction intervals for targets

back to top

 

Developing Artificial Intelligent-based Metamodels for Complex Systems

Motivations for this research are the recent trends in both academia and industry towards integration of simulation modelling techniques in system design and operation. Development of detailed 3D discrete event simulation models has become the common practice for modelling and analysis of complex systems. These models are, however, expert-intensive throughout their lifecycle. Besides, their computational requirement is huge, hindering their application for real time operational planning and optimization.

This research aims at developing abstract metamodels for modelling operations within complex man-made systems. A metamodel is a tool for analysis of a detailed simulation model, which provides insight into some aspect of the underlying system. Artificial intelligent based methods are used in this research for predicting the performance measures for manufacturing enterprises. Feedforward neural networks and adaptive neuro-fuzzy inference system metamodels are compared based on their performance for finding highly nonlinear relationships between independent and dependent variables. Demonstrated results indicate that both methods are capable of generating accurate point predictions. While neural network point predictions are more accurate, the neuro fuzzy models are more transparent.

real systems
arrow
neural network metamodels
From real systems to neural network metamodels

back to top

 

Deakin University acknowledges the traditional land owners of present campus sites.

30th September 2013