Fuzzy, Neural and Expert Systems
Objectives and Syllabus
Provides a working knowledge of the principles of these techniques and an awareness of the evolving technology. Topics covered are - fuzzy systems; fuzzy vs. crisp signals, fuzzification, membership functions, design of rule base, Zadeh's rules of inference, cross membership, defuzzification, mean of maxima and centroid, comparison of fuzzy & PID controllers, non- linearity & time delays: self-adaptive fuzzy controllers; initiation & learning, genetic algorithms: artificial neural networks (ANN); the neuron, multi-layer perceptrons, activation functions, training methods, back propagation, steepest descent, evaluation of Jacobians, data encoding & pre-processing: radial basis function networks; centres, spreads and weights: use of ANNs for inferential estimation, dynamic modelling & optimisation: novel architectures & developments: knowledge based systems, history & examples, data & knowledge bases, inference engine, encoding facts, attributes & properties, inheritance, descriptive rules, identification of the expert, knowledge elicitation, rule based inferencing, user interface, real-time issues, project management.
Demonstrations and case studies of industrial applications are provided which are based on G2 and MATLAB toolboxes. Practicals use fuzzy logic and ANN toolboxes to give hands-on experience of estimation and control techniques.
|Code:||CME 8386 (formerly ACS 686)|
|Time Allocation:||Lectures||40 hours|
|Private Study||70 hours|
|Prerequisites:||Mathematics and Matlab (CME 8360)|
|Assessment:||By report on assignment
By 1 x 2 hour examination
To provide a thorough grounding in the techniques of fuzzy logic control, artificial neural networks and knowledge based systems, to appreciate their potential, and to understand how and when to apply them in the context of process automation.
- To introduce the terminology and technology associated with fuzzy logic control, knowledge based systems and artificial neural networks.
- To provide an understanding of the principles of fuzzy logic and of the functionality of fuzzy controllers.
To develop a quantitative understanding of the structure and training mechanisms of multi-layer perceptron and radial basis function neural networks.
- To become familiar with frame and object based structures and with the common knowledge based processes such as inferencing and elicitation.
- To appreciate the potential benefits and limitations of these techniques, and to be aware of how and when to apply them in the context of process automation.
Prerequisite to this module is the Mathematics and Matlab (CME 8360) module.
Note that there is, by intent, some overlap with the content of module Advanced Process Automation (CME 8368).
This module is of one week's full-time intensive study consisting of a variety of lectures, tutorials, demonstrations, case studies and structured computer-based laboratory work. It is followed by an assignment to be carried out in the student’s own time.
The time allocation for tutorials and practical work provides for the use of Matlab toolboxes for fuzzy logic controller design and the use of ANNs for dynamic modelling for estimation and control. G2 is used for expert system purposes.
- Driankov D, Hellendoorn H and Reinfrank M, An Introduction to Fuzzy Control, Second Edition, Springer, 1996.
- Jackson P, Introduction to Expert Systems, Addison Wesley, Third Edition, 1999.
- Love J, Process Automation Handbook, Springer, 2007.
- Irwin G W, Warwick K and Hunt K J, Neural Network Applications in Control, IEE, 1995.
Fuzzy control: Fuzzy versus crisp signals. Mamdami structure. Fuzzification. Partitioning into sub sets. Biasing. Membership functions. Multiple inputs and cross membership. Production rules and rule base. Operability jacket. Nested rules. Decision logic. Zadeh’s rules of inference. Introduction to fuzzy set theory. Defuzzification. Output membership profiles and weighting. Centre of gravity and mean of maxima methods. I/O scaling factors and tuning. Analogy between fuzzy and PID controllers. Handling non-linearity. Alternative applications. Self adaptive fuzzy controllers.
Artificial neural networks (ANN): Concept of neurons, synapses and weightings. Multi-layer perceptrons (MLP) and hidden layers. Bias. Sigmoid and hyperbolic activation functions and squashing. Operation of MLP. Back propagation training algorithms. Steepest descent. Learning rate and momentum constants. Network size and generalisation. Evaluation of Jacobians. Data encoding, eg scaling and spread encoding. Pre-processing of data. Radial basis function (RBF) networks. Gaussian activation function. Centres and k-means clustering. Spread and nearest neighbours. Weights and multiple linear regression. Use of ANNs for dynamic modelling: time series, globally and locally recurrent models. Use of ANNs for identification, estimation, classification, control and optimisation. Novel architectures and developments.
Knowledge based systems: History and examples of expert systems in engineering, eg advisory, diagnostic and real-time. Definitions and structure. Data and knowledge bases and inference engine. Methods of encoding facts. Frames versus objects. Attributes and properties. Inheritance. Instantiation. Data acquisition. Relationships between facts. Production rule based heuristics. Descriptive rules. Identification of the expert. Methodologies for knowledge elicitation: manual and commercial tools. Rule based inferencing. Forward and backward chaining versus breadth search. Pattern matching. User interface implications. Real-time issues: processing speed, time-stamping, etc. Pitfalls of AI projects. Project time scales and management.