No description available.

No description available.

No description available.

This thesis describes hydrocarbon vapour sensing using Langmuir-Blodgett films prepared from: a co-ordination polymer; substituted phthalocyanines containing copper and zinc as the central metal ions; and a polysiloxane. The physical and chemical properties of the co-ordination polymer, 5,5'-methylenebis (N- hexadecylsalicylidenamine), at the air water interface were investigated using Brewster angle microscopy and surface pressure versus area measurements. Langmuir-Blodgett films were built-up on a variety of substrates. The addition of copper acetate to the subphase caused a change in both the physical and optical properties of the Langmuir- Blodgett layers. Film thickness data suggest that a true monolayer (thickness ca 2 nm) is only formed under these conditions. The multilayer films were studied using X-ray diffraction, UV/Visible spectroscopy, ellipsometry, surface plasmon resonance, surface profiling and electron spin resonance. The response of each film when exposed to, benzene, toluene, ethanol and water vapours were recorded. Two optical systems were used, both based on surface plasmon resonance. The first incorporated a silicon photodiode to record the intensity of the reflected light. The second was similar to that of surface plasmon microscopy, using a charge coupled device camera to monitor the reflected light intensity from the Langmuir-Blodgett film/metal interface. The co-ordination polymer was found to be most sensitive to benzene and could reliably detect concentrations of this vapour down to 100 vapour parts per million. Data obtained when the co-ordination polymer was exposed to benzene and water vapour (using the latter system) were presented to a neural network for recognition.

Evaluation of the environmental impacts of Intelligent Transport Systems and transport infrastructure management schemes relies heavily on the development of accurate and reliable environmental emissions models. Existing state-of-the-art models estimate pollutants based on a typical urban driving cycle using an aggregate modelling approach where a 'characteristic' vehicle is used to represent dissimilar vehicle populations. While this approach has been accepted by transport planners for strategic level studies, it can be argued that modelling individual vehicle emissions based on vehicle dynamics would result in more reliable evaluations of operational-level project impacts. The primary objective of this thesis is to develop vehicle emissions and fuel consumption models under hot stabilised settings and various traffic conditions using Australian fleet data collected from laboratory tests. The models use second-by-second vehicle real-time data to predict fuel consumption (FC) and pollutant emissions (HC, CO, NOx) at different levels of speed, acceleration, air-to-fuel ratio and torque. The data required for model development, calibration and validation was collated from laboratory tests conducted by the Second National In-Service Emissions (NISE 2) project. A total of 27 vehicles (including small, medium and large passenger vehicles; four-wheel drive (small and large); and light commercial vehicles were used in model development. The laboratory data, which comprised more than 48,500 second-by-second observations, was then pre-processed and randomly assigned to calibration and validation data sets for model development. The thesis then adopted a rigorous approach to develop and evaluate a large number of neural network architectures to determine the most suitable modelling framework. First, a pilot test was conducted to test different model development scenarios and establish some guidelines on the general framework for model development. The results were used to determine some of the crucial neural network parameters (eg learning rule or optimisation technique and most appropriate architecture) for use in subsequent modelling. Selected models were then further refined using test data from individual and aggregate vehicle types. This resulted in further refinement of modelling inputs where, for example, sensitivity analysis showed that speed and acceleration were the two most crucial inputs and that including other input parameters did not improve the accuracy of the results. The performance of selected neural network models was then compared to a number of sophisticated and complex statistical techniques based on multiple and non-linear regression models. The results generally showed that ANN models are effective and suitable for modelling emissions and that they perform as well or even better than the complex regression models tested in this study. Another general finding across all vehicles and for all models (neural and statistical) is that predictions are more accurate for fuel consumption and CO emissions than for other vehicular emissions. The models were also found to under-predict the emissions values at the peaks of graphs, but were generally consistent in their outputs across all other driving conditions. In this study, it was also found that one of the main advantages of the neural network approach over regression is the ease of developing one model to accurately predict multiple outputs. This is in contrast to the regression modelling approach, where it was found that accurate results matching neural network performance can only be achieved using one distinct model for predicting each output. This would clearly undermine the statistical approach because a large number of models would then need to be developed for a road network where second-by-second data is available for hundreds of vehicle types. Hence, the benefits of using neural networks immediately become clear and more appealing. This thesis also identified a number of issues for future research directions. To increase the accuracy and overall quality of the models, future research needs to include further classifications of vehicle types and other pertinent variables such as manufacture year, odometer reading and making use of a larger sample of modern vehicles representing current vehicle fleet compositions. There is also scope to improve the testing procedures by including road grade and air condition use, which are important factors that impact on vehicle performance and emissions. Future research can also benefit from testing other drive cycles and cross validation of models across different driving cycles. Model performance can also be enhanced by collecting instantaneous data using instrumented vehicles where emissions can be collected under real-life conditions rather than from controlled laboratory environments. Finally, the real benefit from development of these models is the ability to interface them to micro-simulation models where instantaneous speed and acceleration data can be provided to the emissions model on a second-by-second basis. The neural network emissions model would then be used to evaluate the impacts of ITS and other traffic management strategies with the aim of identifying the best environment-friendly traffic management approaches. This thesis has successfully achieved its objectives by demonstrating the feasibility of using neural networks for modelling vehicle emissions. The thesis further demonstrated the superior quality and advantages of the neural network approach over the more established statistical regression methods. Finally, the models developed in this study will allow researcher and practitioners alike to develop a better understanding and appreciation of the environmental impacts resulting from transport schemes aimed at reducing traffic congestion and enhancing environmental quality.

Evaluation of the environmental impacts of Intelligent Transport Systems and transport infrastructure management schemes relies heavily on the development of accurate and reliable environmental emissions models. Existing state-of-the-art models estimate pollutants based on a typical urban driving cycle using an aggregate modelling approach where a 'characteristic' vehicle is used to represent dissimilar vehicle populations. While this approach has been accepted by transport planners for strategic level studies, it can be argued that modelling individual vehicle emissions based on vehicle dynamics would result in more reliable evaluations of operational-level project impacts. The primary objective of this thesis is to develop vehicle emissions and fuel consumption models under hot stabilised settings and various traffic conditions using Australian fleet data collected from laboratory tests. The models use second-by-second vehicle real-time data to predict fuel consumption (FC) and pollutant emissions (HC, CO, NOx) at different levels of speed, acceleration, air-to-fuel ratio and torque. The data required for model development, calibration and validation was collated from laboratory tests conducted by the Second National In-Service Emissions (NISE 2) project. A total of 27 vehicles (including small, medium and large passenger vehicles; four-wheel drive (small and large); and light commercial vehicles were used in model development. The laboratory data, which comprised more than 48,500 second-by-second observations, was then pre-processed and randomly assigned to calibration and validation data sets for model development. The thesis then adopted a rigorous approach to develop and evaluate a large number of neural network architectures to determine the most suitable modelling framework. First, a pilot test was conducted to test different model development scenarios and establish some guidelines on the general framework for model development. The results were used to determine some of the crucial neural network parameters (eg learning rule or optimisation technique and most appropriate architecture) for use in subsequent modelling. Selected models were then further refined using test data from individual and aggregate vehicle types. This resulted in further refinement of modelling inputs where, for example, sensitivity analysis showed that speed and acceleration were the two most crucial inputs and that including other input parameters did not improve the accuracy of the results. The performance of selected neural network models was then compared to a number of sophisticated and complex statistical techniques based on multiple and non-linear regression models. The results generally showed that ANN models are effective and suitable for modelling emissions and that they perform as well or even better than the complex regression models tested in this study. Another general finding across all vehicles and for all models (neural and statistical) is that predictions are more accurate for fuel consumption and CO emissions than for other vehicular emissions. The models were also found to under-predict the emissions values at the peaks of graphs, but were generally consistent in their outputs across all other driving conditions. In this study, it was also found that one of the main advantages of the neural network approach over regression is the ease of developing one model to accurately predict multiple outputs. This is in contrast to the regression modelling approach, where it was found that accurate results matching neural network performance can only be achieved using one distinct model for predicting each output. This would clearly undermine the statistical approach because a large number of models would then need to be developed for a road network where second-by-second data is available for hundreds of vehicle types. Hence, the benefits of using neural networks immediately become clear and more appealing. This thesis also identified a number of issues for future research directions. To increase the accuracy and overall quality of the models, future research needs to include further classifications of vehicle types and other pertinent variables such as manufacture year, odometer reading and making use of a larger sample of modern vehicles representing current vehicle fleet compositions. There is also scope to improve the testing procedures by including road grade and air condition use, which are important factors that impact on vehicle performance and emissions. Future research can also benefit from testing other drive cycles and cross validation of models across different driving cycles. Model performance can also be enhanced by collecting instantaneous data using instrumented vehicles where emissions can be collected under real-life conditions rather than from controlled laboratory environments. Finally, the real benefit from development of these models is the ability to interface them to micro-simulation models where instantaneous speed and acceleration data can be provided to the emissions model on a second-by-second basis. The neural network emissions model would then be used to evaluate the impacts of ITS and other traffic management strategies with the aim of identifying the best environment-friendly traffic management approaches. This thesis has successfully achieved its objectives by demonstrating the feasibility of using neural networks for modelling vehicle emissions. The thesis further demonstrated the superior quality and advantages of the neural network approach over the more established statistical regression methods. Finally, the models developed in this study will allow researcher and practitioners alike to develop a better understanding and appreciation of the environmental impacts resulting from transport schemes aimed at reducing traffic congestion and enhancing environmental quality.

Deep neural networks (DNNs) and deep learning approaches yield state-of-the-art performance in a range of machine learning tasks, including automatic speech recognition. The multi-layer transformations and activation functions in DNNs, or related network variations, allow complex and difficult data to be well modelled. However, the highly distributed representations associated with these models make it hard to interpret the parameters. The whole neural network is commonly treated a ``black box''. The behaviours of activation functions and the meanings of network parameters are rarely controlled in the standard DNN training. Though a sensible performance can be achieved, the lack of interpretations to network structures and parameters causes better regularisation and adaptation on DNN models challenging. In regularisation, parameters have to be regularised universally and indiscriminately. For instance, the widely used L2 regularisation encourages all parameters to be zeros. In adaptation, it requires to re-estimate a large number of independent parameters. Adaptation schemes in this framework cannot be effectively performed when there are limited adaptation data. This thesis investigates structured deep neural networks. Special structures are explicitly designed, and they are imposed with desired interpretation to improve DNN regularisation and adaptation. For regularisation, parameters can be separately regularised based on their functions. For adaptation, parameters can be adapted in groups or partially adapted according to their roles in the network topology. Three forms of structured DNNs are proposed in this thesis. The contributions of these models are presented as follows. The first contribution of this thesis is the multi-basis adaptive neural network. This form of structured DNN introduces a set of parallel sub-networks with restricted connections. The design of restricted connectivity allows different aspects of data to be explicitly learned. Sub-network outputs are then combined, and this combination module is used as the speaker-dependent structure that can be robustly estimated for adaptation. The second contribution of this thesis is the stimulated deep neural network. This form of structured DNN relates and smooths activation functions in regions of the network. It aids the visualisation and interpretation of DNN models but also has the potential to reduce over-fitting. Novel adaptation schemes can be performed on it, taking advantages of the smooth property that the stimulated DNN offer. The third contribution of this thesis is the deep activation mixture model. Also, this form of structured DNN encourages the outputs of activation functions to achieve a smooth surface. The output of one hidden layer is explicitly modelled as the sum of a mixture model and a residual model. The mixture model forms an activation contour, and the residual model depicts fluctuations around this contour. The smoothness yielded by a mixture model helps to regularise the overall model and allows novel adaptation schemes.

Neural machine translation has been lately established as the new state of the art in machine translation, especially with the Transformer model. This model emphasized the importance of self-attention mechanism and sug- gested that it could capture some linguistic phenomena. However, this claim has not been examined thoroughly, so we propose two main groups of meth- ods to examine the relation between these two. Our methods aim to im- prove the translation performance by directly manipulating the self-attention layer. The first group focuses on enriching the encoder with source-side syn- tax with tree-related position embeddings or our novel specialized attention heads. The second group is a joint translation and parsing model leveraging self-attention weight for the parsing task. It is clear from the results that enriching the Transformer with sentence structure can help. More impor- tantly, the Transformer model is in fact able to capture this type of linguistic information with guidance in the context of multi-task learning at nearly no increase in training costs. 1

Computer integrated manufacturing environments and competition among companies to meet customer requirements raise the need for the use of online methodologies in combination with traditional Statistical Process Control tools. This study focuses on detecting the change point, when a shift in mean occurs, in a normal bivariate process using two different approaches. First, Multivariate Cumulative Sum (MCUSUM) and Multivariate Exponentially Weighted Moving Average (MEWMA) statistical procedures were used in detecting the mean shift in the process. Then the step-change detection and neural network approaches were applied to the outputs of MCUSUM and MEWMA statistical procedures to identify the time of the change. The results show that the step-change and neural network approaches are capable of detecting the time of the change earlier than either the MCUSUM or MEWMA statistical procedure.

Video super-resolution (VSR) aims to give a satisfying estimation of a high-resolution (HR) image from multiple similar low-resolution (LR) images by exploiting their hidden redundancy. The rapid development of convolutional neural network (CNN) techniques provide numerous new possibilities to solve the VSR problem. Recent VSR methods combine CNN with motion compensation to cancel the inconsistencies among the LR images and merge them to an HR images. To compensate the motion, pixels in input frames are warped according to optical-flow-like information. In this procedure, trade-off has to be made between the distraction caused by spatio-temporal inconsistencies and the pixel-wise detail damage caused by the compensation. We proposed a novel VSR method with the name, Video Super-Resolution via Dynamic Local Filter Network, and its upgraded edition, Video Super-Resolution with Compensation in Feature Extraction. Both methods perform motion compensation via a dynamic local filter network, which processes the input images with dynamically generated filter kernels. These kernels are sample-specific and position-specific. Therefore, our proposed methods can eliminate the inter-frame differences during feature extractions without explicitly manipulating pixels. The experimental results demonstrate that our methods outperform the state-of-the-art VSR algorithms in terms of PSNR and SSIM and recover more details with superior visual quality.