Use this URL to cite or link to this record in EThOS:
Title: Neuro-fuzzy based intelligent approaches to nonlinear system identification and forecasting
Author: Alshejari, Abeer
ISNI:       0000 0004 7431 7879
Awarding Body: University of Westminster
Current Institution: University of Westminster
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
Nearly three decades back nonlinear system identification consisted of several ad-hoc approaches, which were restricted to a very limited class of systems. However, with the advent of the various soft computing methodologies like neural networks and the fuzzy logic combined with optimization techniques, a wider class of systems can be handled at present. Complex systems may be of diverse characteristics and nature. These systems may be linear or nonlinear, continuous or discrete, time varying or time invariant, static or dynamic, short term or long term, central or distributed, predictable or unpredictable, ill or well defined. Neurofuzzy hybrid modelling approaches have been developed as an ideal technique for utilising linguistic values and numerical data. This Thesis is focused on the development of advanced neurofuzzy modelling architectures and their application to real case studies. Three potential requirements have been identified as desirable characteristics for such design: A model needs to have minimum number of rules; a model needs to be generic acting either as Multi-Input-Single-Output (MISO) or Multi-Input-Multi-Output (MIMO) identification model; a model needs to have a versatile nonlinear membership function. Initially, a MIMO Adaptive Fuzzy Logic System (AFLS) model which incorporates a prototype defuzzification scheme, while utilising an efficient, compared to the Takagi–Sugeno–Kang (TSK) based systems, fuzzification layer has been developed for the detection of meat spoilage using Fourier transform infrared (FTIR) spectroscopy. The identification strategy involved not only the classification of beef fillet samples in their respective quality class (i.e. fresh, semi-fresh and spoiled), but also the simultaneous prediction of their associated microbiological population directly from FTIR spectra. In the case of AFLS, the number of memberships for each input variable was directly associated to the number of rules, hence, the “curse of dimensionality” problem was significantly reduced. Results confirmed the advantage of the proposed scheme against Adaptive Neurofuzzy Inference System (ANFIS), Multilayer Perceptron (MLP) and Partial Least Squares (PLS) techniques used in the same case study. In the case of MISO systems, the TSK based structure, has been utilized in many neurofuzzy systems, like ANFIS. At the next stage of research, an Adaptive Fuzzy Inference Neural Network (AFINN) has been developed for the monitoring the spoilage of minced beef utilising multispectral imaging information. This model, which follows the TSK structure, incorporates a clustering pre-processing stage for the definition of fuzzy rules, while its final fuzzy rule base is determined by competitive learning. In this specific case study, AFINN model was also able to predict for the first time in the literature, the beef’s temperature directly from imaging information. Results again proved the superiority of the adopted model. By extending the line of research and adopting specific design concepts from the previous case studies, the Asymmetric Gaussian Fuzzy Inference Neural Network (AGFINN) architecture has been developed. This architecture has been designed based on the above design principles. A clustering preprocessing scheme has been applied to minimise the number of fuzzy rules. AGFINN incorporates features from the AFLS concept, by having the same number of rules as well as fuzzy memberships. In spite of the extensive use of the standard symmetric Gaussian membership functions, AGFINN utilizes an asymmetric function acting as input linguistic node. Since the asymmetric Gaussian membership function’s variability and flexibility are higher than the traditional one, it can partition the input space more effectively. AGFINN can be built either as an MISO or as an MIMO system. In the MISO case, a TSK defuzzification scheme has been implemented, while two different learning algorithms have been implemented. AGFINN has been tested on real datasets related to electricity price forecasting for the ISO New England Power Distribution System. Its performance was compared against a number of alternative models, including ANFIS, AFLS, MLP and Wavelet Neural Network (WNN), and proved to be superior. The concept of asymmetric functions proved to be a valid hypothesis and certainly it can find application to other architectures, such as in Fuzzy Wavelet Neural Network models, by designing a suitable flexible wavelet membership function. AGFINN’s MIMO characteristics also make the proposed architecture suitable for a larger range of applications/problems.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available