Theses and dissertations (Engineering and Built Environment)
Permanent URI for this collectionhttp://ir-dev.dut.ac.za/handle/10321/10
Browse
10 results
Search Results
Item Learning rate optimisation of an image processing deep convolutional neural network(2021) Buthelezi, Sibusiso Blessing; Reddy, Seren; Twala, BhekisiphoThe major contribution of this dissertation is the proposal of the use of mathematical models to identify an optimal learning rate for an image processing deep convolutional neural network (DCNN). This model is derived from a nonlinear regression relationship between the learning rate and the accuracy of a test DCNN model. This relationship is meant to (A) resolve the problem of arbitrarily selecting the initial learning rate (B) reduce computational resource requirement and (C) reduce training instabilities. An algorithm is developed to analyse an inputted DCNN model and subsequently render output parameters that may be used to aid in the selection of an OLR. The benefit of an OLR includes improved training stability and reduced computational resources. The results rendered by the OLR algorithm proposes that an optimal learning rate improves model performance; this is described by the test model average accuracy of 91%. Furthermore, a model validation graph is also extrapolated. which will illustrate the mathematical model accuracy and the region of interest (ROI). The ROI defines the region in the learning rate spectrum with a positive effect on model performance.Item Modeling and recognition of faults in smart distribution grid using maching intelligence technique(2018) Onaolapo, Adeniyi Kehinde; Akindeji, Timothy Kayode; Adetiba, EmmanuelElectrical power systems experience unforeseen faults attributable to diverse arbitrary reasons. Unanticipated failures occurring in power systems are to be prevented from propagating to other parts of the protective system to enhance economic efficacy of electric utilities and provide better service to energy consumers. Since most consumers are directly connected to power distribution networks, there is an increasing research efforts in distribution network fault recognition and fault-types identifications to solve the problem of outages due to faults. This study focuses on fault recognition and fault-types identification in electrical power distribution system based on the Design Science Research (DSR) approach. Diverse simulations of fault types at different locations were applied to the IEEE 13 Node Test Feeder to produce three phase currents and voltages as data set for this study. This was realized by modelling the IEEE 13-node benchmark test feeder in MATLAB-Simulink R2017a. In order to achieve intelligent fault recognition and fault-type identification, different Multi-layer Perceptron Artificial Neural Networks (MLP-ANN) models were designed and subsequently trained using the generated dataset with the Neural Network toolbox in MATLAB R2017a. The fault recognition task verifies if a fault occurs or not while the fault-types identification task determines the fault class as well as the faulty phase(s). Results obtained from the various MLP-ANN models were recorded and statistically analyzed. Acceptable performances were obtained for fault recognition with the 6-25-20-15-1 MLP-ANN architecture, for fault-types identification with the 6-40-4 MLP-ANN architecture and for fault location with the 6-30-15-5-4 MLP-ANN architecture. Given the result obtained in this study, MLP-ANN is adjudged suitable for intelligent fault recognition and fault-types identification in power distribution systems. The trained MLP-ANNs in this study could ultimately be incorporated in power distribution networks within South Africa and beyond in order to enhance energy customers’ satisfaction.Item Predicting mass transfer in pilot scale external loop airlift reactors using neural networks(2018) Naidoo, Nirvana; Pauck, Walter James; Carsky, MilanAirlift reactors are a viable means for conducting large scale mass transfer operations. However, due to the difficulty experienced in understanding the complex behavioural characteristics of these reactors, the design of airlift reactors becomes very complicated and largely empirical. Existing correlations and traditional computational fluid dynamic modelling has proven to be mostly reactor dependent and not widely applicable thereby limiting their application. There is therefore a need to develop a model that does not require prior knowledge of relationships between parameters of these reactors but instead uses an alternate method to assist with the design of airlift reactors. An artificial neural network represents this method. The aim of this investigation was to build an artificial neural network using selected input data of Newtonian fluids in pilot scale external loop airlift reactors of varying designs in order to predict the mass transfer coefficient in other external loop airlift reactors with more general geometry. To achieve this, a large base of experimental data (663) was generated using glycerine-air and water-air systems in 5 configurations of external loop airlift reactors with 3 categories of sparger design. The data was modelled using the artificial neural network software, Predict (Version 3.30) by Neuralware. The Coefficient of Correlation for the neural network model was 0.98. The neural network model was tested with unseen external data from various sources of which the R values ranged from 0.91 to 0.99. Additional external data was evaluated with the superficial gas velocity out of the range of the experimental data from this investigation and with a very different design of sparger. The R values for this additional data were 0.85 and 0.67-0.85 respectively. To achieve good correlations it was found necessary to take into account the sparger design and pore size; the actual geometric dimensions of the reactors namely the riser and downcomer diameters and heights; the visual observations of the approximate bubble size and bubble flow patterns and static liquid height in addition to the more usual data of the area and aspect ratios; the fluid properties namely, surface tension, density and viscosity; the superficial gas velocity; the downcomer superficial liquid velocity; the riser gas holdup and the downcomer gas holdup. However, some parameters like the static liquid height although considered important appeared not to be. By considering these as important input variables into the network, the artificial neural network was able to give excellent approximations for both seen and unseen data for some of the reactor configurations. However, the network also had the ability to pick up differences in the reactor configurations were it did not predict well, especially with respect to sparger design. An important conclusion arrived at in this investigation was the significant influence of the sparger and its design on the mass transfer. The sensitivity of the network to the sparger design means that a greater quantification of the influence of the sparger design is required.Item Assessment of UV index using artificial neural networks(2002) Human, Sep; Bajic, Vladimir B.; Duffy, Kevin JanItem Statistical pattern recognition based on LVQ artificial neural networks : application to TATA box motif(2000) Wang, Haiyan; Bajic, Vladimir B.The computational analysis of eukaryotic promoters are among the most important and complex research domains that may contribute to complete gene identification. The current methods for promoter recognition are not sufficiently developed. Eukaryotic promoters contain a number of short motifs that may be used in promoter recognition. Having good computational models for these motifs can be crucial for increased efficiency of promoter recognition programs. This study proposes a combined statistical and LVQ neural network system as a computational model of the TAT A box motif of eukaryotic promoters. The methodology used is universal and applicable to any short functional motif in DNA. The statistical analysis of the core TAT A motif hexamer and its neighboring haxamers show strong regularities that can be used in motif recognition. Moreover, the positional distribution of the TAT A motif in terms of its distance from the transcription start site is very regular and is used in the statistical modeling. Furthermore, the matching score of the position weight matrix for the motif was used as a part of the model. Based on these statistical properties. a novel LV Q classifier for TAT A motif recognition is developed. The characteristics of the method are that the genetic algorithm was used for finding good initial weights of the LV Q system, while fine tuning of two LVQ networks was done by the lvq? algorithm. The final computational model is developed for a recognition level of 67.8o/c correct recognition on the test set with less than 1% false recognition. This model is evaluated in the task of promoter recognition on an independent test set. The results in promoter recognition outperform three other promoter recognition programs. It is shown that the recognition of promoters based on the recognition of the TAT A motifs using this new model is superior to the recognition based on the currently used position weight matrix description of this motif.Item Neural networks approach to process control : the case of processes with long dead times(1999) McLeod, Charles Meredith; Bajic, Vladimir B.This study relates to applications of static artificial neural networks (ANNs) to two basic problems of process control: (a) process model identification, and (b) optimal controller tuning. The emphasis is on model identification, where several novel techniques are introduced. A review of the use of ANNs for determining optimal controller settings is included as a logical adjunct which would make the complete system suitable for realisation as a portable or networked system. Three methods for obtaining good approximations for the parameters of first-order processes with long dead time using artificial neural networks (ANNs) are proposed and described. These are termed in this study: time-domain, frequency-domain and model-based methods. In each case the aim was to develop a brief one-shot test that could be applied with minimal disturbance to a closed loop control system. These methods build on existing techniques, but introduce the following novel aspects: 2. The frequency-domain method makes use of the first 81 components of the FFT without further selection as input to a static ANN to yield process parameter estimates. 3. The model-based method uses a simple single-neuron implementation of an ARX model and uses a static ANN to relate process parameter values to the weights of this neuron. In making the analysis, the process input and output are applied repetitively to the neuron model with delays getting progressively larger. Useful effects arising from this are explored. A technique in which ANN training sets are slightly distorted in a random way during training of a radial basis function is developed as part of the time- and frequencydomain methods. The benefits arising from this technique are demonstrated. These experimental ANN-based control methods are evaluated by means of simulations in which accuracy in the presence of measurement noise and performance with higher order processes is measured and analysed. Although the main theme of this study is first-order-plus-dead-time (FOPDT) processes, the full autotuning scheme is tested with some representative higher order processes. Finally, the composition of a complete autotuning scheme is proposed which includes the automatic generation of controller parameters by means of ANN s.Item Design and development of a process control valve diagnostic system based on artificial neural network ensembles(2016) Sewdass, Sugith; Govender, PoobalanThis research discusses the design and development of a computational intelligent based diagnostic system to assess the operating state of a process control valve. Process control valves react to a controller signal and are the main source of faults in a control loop. The elasticity inherent within a valve’s mechanical construction makes it prone to nonlinearities such as backlash, hysteresis and stiction. These nonlinearities negatively affect the performance of a process control loop during a control session. The diagnostic system proposed in this research utilises artificial neural network systems configured as ensembles to classify common control valve faults. Each ensemble functions as a ‘specialist’ trained to identify a specific loop fault. The team of specialized artificial neural networks are configured into a single comprehensive system to detect common control loops problems such as valve hysteresis, backlash, stiction and low air supply. The detection of a specific type of fault is achieved by comparing the mean square error output from each network. The ensemble having the lowest mean square error is the network that has been trained to identify a specific type of fault. Two practical methods to simulate control valve stiction and hysteresis are also presented in this study. These methods make it possible for researchers to investigate dynamics of nonlinear behaviour when these nonlinear effects occur in the control channel.Item Object motion detection, extraction and filtering using ANN ensembles(2009) Moorgas, Kevin Emanuel; Govender, PoobalanThis research is devoted to the development of an intelligent image motion detection system based on artificial neural networks (ANN’s). Object motion detection, non-stationary image isolation and extraction, and image filtering is investigated, with the intention of developing a system that will overcome some of the shortcomings associated with the performance of conventional motion detection systems. Motion detection and image extraction finds popular application in medical imagery and engineering based diagnostics systems. Conventional image processing systems utilise Digital Signal Processing (DSP) to perform the non-stationary image motion detection function. Aliasing and filtering are problematic processes in DSP based image processing systems. The proposed ANN motion detection system overcomes some of these shortcomings. The study compares the performance of conventional DSP systems to that of the proposed ANN based system. The excellent noise immunity, ability to generalise and robustness of the ANN system is exploited in the design of the motion detection system. The ANN’s are arranged as ensembles in order to improve the computation time of the proposed motion detection system. A hybrid system comprising DSP and ANN ensembles is also proposed in the study. The hybrid system exploits the positive characteristics of DSP and ANN’s within a single system. The performance of the pure ANN system and the hybrid system is compared to that of DSP systems, using the image’s signal-to-noise ratio and computation times as a basis for comparison.Item Design and implementation of an intelligent vision and sorting system(2009) Li, Zhi; Govender, PoobalanThis research focuses on the design and implementation of an intelligent machine vision and sorting system that can be used to sort objects in an industrial environment. Machine vision systems used for sorting are either geometry driven or are based on the textural components of an object’s image. The vision system proposed in this research is based on the textural analysis of pixel content and uses an artificial neural network to perform the recognition task. The neural network has been chosen over other methods such as fuzzy logic and support vector machines because of its relative simplicity. A Bluetooth communication link facilitates the communication between the main computer housing the intelligent recognition system and the remote robot control computer located in a plant environment. Digital images of the workpiece are first compressed before the feature vectors are extracted using principal component analysis. The compressed data containing the feature vectors is transmitted via the Bluetooth channel to the remote control computer for recognition by the neural network. The network performs the recognition function and transmits a control signal to the robot control computer which guides the robot arm to place the object in an allocated position. The performance of the proposed intelligent vision and sorting system is tested under different conditions and the most attractive aspect of the design is its simplicity. The ability of the system to remain relatively immune to noise, its capacity to generalize and its fault tolerance when faced with missing data made the neural network an attractive option over fuzzy logic and support vector machines.Item Diesel engine performance modelling using neural networks(2005) Rawlins, Mark SteveThe aim of this study is to develop, using neural networks, a model to aid the performance monitoring of operational diesel engines in industrial settings. Feed-forward and modular neural network-based models are created for the prediction of the specific fuel consumption on any normally aspirated direct injection four-stroke diesel engine. The predictive capability of each model is compared to that of a published quadratic method. Since engine performance maps are difficult and time consuming to develop, there is a general scarcity of these maps, thereby limiting the effectiveness of any engine monitoring program that aims to manage the fuel consumption of an operational engine. Current methods applied for engine consumption prediction are either too complex or fail to account for specific engine characteristics that could make engine fuel consumption monitoring simple and general in application. This study addresses these issues by providing a neural network-based predictive model that requires two measured operational parameters: the engine speed and torque, and five known engine parameters. The five parameters are: rated power, rated and minimum specific fuel consumption bore and stroke. The neural networks are trained using the performance maps of eight commercially available diesel engines, with one entire map being held out of sample for assessment of model generalisation performance and application validation. The model inputs are defined using the domain expertise approach to neural network input specification. This approach requires a thorough review of the operational and design parameters affecting engine fuel consumption performance and the development of specific parameters that both scale and normalize engine performance for comparative purposes. Network architecture and learning rate parameters are optimized using a genetic algorithm-based global search method together with a locally adaptive learning algorithm for weight optimization. Network training errors are statistically verified and the neural network test responses are validation tested using both white and black box validation principles. The validation tests are constructed to enable assessment of the confidence that can be associated with the model for its intended purpose. Comparison of the modular network with the feed-forward network indicates that they learn the underlying function differently, with the modular network displaying improved generalisation on the test data set. Both networks demonstrate improved predictive performance over the published quadratic method. The modular network is the only model accepted as verified and validated for application implementation. The significance of this work is that fuel consumption monitoring can be effectively applied to operational diesel engines using a neural network-based model, the consequence of which is improved long term energy efficiency. Further, a methodology is demonstrated for the development and validation testing of modular neural networks for diesel engine performance prediction.