Implementation of gaussian process models for non-linear system identification

This thesis is concerned with investigating the use of Gaussian Process (GP) models for the identification of nonlinear dynamic systems. The Gaussian Process model is a non-parametric approach to system identification where the model of the underlying system is to be identified through the application of Bayesian analysis to empirical data. The GP modelling approach has been proposed as an alternative to more conventional methods of system identification due to a number of attractive features. In particular, the Bayesian probabilistic framework employed by the GP model has been shown to have potential in tackling the problems found in the optimisation of complex nonlinear models such as those based on multiple model or neural network structures. Furthermore, due to this probabilistic framework, the predictions made by the GP model are probability distributions composed of mean and variance components. This is in contrast to more conventional methods where a predictive point estimate is typically the output of the model. This additional variance component of the model output has been shown to be of potential use in model-predictive or adaptive control implementations. A further property that is of potential interest to those working on system identification problems is that the GP model has been shown to be particularly effective in identifying models from sparse datasets. Therefore, the GP model has been proposed for the identification of models in off-equilibrium regions of operating space, where more established methods might struggle due to a lack of data. The majority of the existing research into modelling with GPs has concentrated on detailing the mathematical methodology and theoretical possibilities of the approach. Furthermore, much of this research has focused on the application of the method toward statistics and machine learning problems. This thesis investigates the use of the GP model for identifying nonlinear dynamic systems from an engineering perspective. In particular, it is the implementation aspects of the GP model that are the main focus of this work. Due to its non-parametric nature, the GP model may also be considered a ‘black-box’ method as the identification process relies almost exclusively on empirical data, and not on prior knowledge of the system. As a result, the methods used to collect and process this data are of great importance, and the experimental design and data pre-processing aspects of the system identification procedure are investigated in detail. Therefore, in the research presented here the inclusion of prior system knowledge into the overall modelling procedure is shown to be an invaluable asset in improving the overall performance of the GP model. In previous research, the computational implementation of the GP modelling approach has been shown to become problematic for applications where the size of training dataset is large (i.e. one thousand or more points). This is due to the requirement in the GP modelling approach for repeated inversion of a covariance matrix whose size is dictated by the number of points included in the training dataset. Therefore, in order to maintain the computational viability of the approach, a number of different strategies have been proposed to lessen the computational burden. Many of these methods seek to make the covariance matrix sparse through the selection of a subset of existing training data. However, instead of operating on an existing training dataset, in this thesis an alternative approach is proposed where the training dataset is specifically designed to be as small as possible whilst still containing as much information. In order to achieve this goal of improving the ‘efficiency’ of the training dataset, the basis of the experimental design involves adopting a more deterministic approach to exciting the system, rather than the more common random excitation approach used for the identification of black-box models. This strategy is made possible through the active use of prior knowledge of the system. The implementation of the GP modelling approach has been demonstrated on a range of simulated and real-world examples. The simulated examples investigated include both static and dynamic systems. The GP model is then applied to two laboratory-scale nonlinear systems: a Coupled Tanks system where the volume of liquid in the second tank must be predicted, and a Heat Transfer system where the temperature of the airflow along a tube must be predicted. Further extensions to the GP model are also investigated including the propagation of uncertainty from one prediction to the next, the application of sparse matrix methods, and also the use of derivative observations. A feature of the application of GP modelling approach to nonlinear system identification problems is the reliance on the squared exponential covariance function. In this thesis the benefits and limitations of this particular covariance function are made clear, and the use of alternative covariance functions and ‘mixed-model’ implementations is also discussed