Neural networks in control engineering

Master Thesis


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title
The purpose of this thesis is to investigate the viability of integrating neural networks into control structures. These networks are an attempt to create artificial intelligent systems with the ability to learn and remember. They mathematically model the biological structure of the brain and consist of a large number of simple interconnected processing units emulating brain cells. Due to the highly parallel and consequently computationally expensive nature of these networks, intensive research in this field has only become feasible due to the availability of powerful personal computers in recent years. Consequently, attempts at exploiting the attractive learning and nonlinear optimization characteristics of neural networks have been made in most fields of science and engineering, including process control. The control structures suggested in the literature for the inclusion of neural networks in control applications can be divided into four major classes. The first class includes approaches in which the network forms part of an adaptive mechanism which modulates the structure or parameters of the controller. In the second class the network forms part of the control loop and replaces the conventional control block, thus leading to a pure neural network control law. The third class consists of topologies in which neural networks are used to produce models of the system which are then utilized in the control structure, whilst the fourth category includes suggestions which are specific to the problem or system structure and not suitable for a generic neural network-based-approach to control problems. Although several of these approaches show promising results, only model based structures are evaluated in this thesis. This is due to the fact that many of the topologies in other classes require system estimation to produce the desired network output during training, whereas the training data for network models is obtained directly by sampling the system input(s) and output(s). Furthermore, many suggested structures lack the mathematical motivation to consider them for a general structure, whilst the neural network model topologies form natural extensions of their linear model based origins. Since it is impractical and often impossible to collect sufficient training data prior to implementing the neural network based control structure, the network models have to be suited to on-line training during operation. This limits the choice of network topologies for models to those that can be trained on a sample by sample basis (pattern learning) and furthermore are capable of learning even when the variation in training data is relatively slow as is the case for most controlled dynamic systems. A study of feedforward topologies (one of the main classes of networks) shows that the multilayer perceptron network with its backpropagation training is well suited to model nonlinear mappings but fails to learn and generalize when subjected to slow varying training data. This is due to the global input interpretation of this structure, in which any input affects all hidden nodes such that no effective partitioning of the input space can be achieved. This problem is overcome in a less flexible feedforward structure, known as regular Gaussian network. In this network, the response of each hidden node is limited to a -sphere around its center and these centers are fixed in a uniform distribution over the entire input space. Each input to such a network is therefore interpreted locally and only effects nodes with their centers in close proximity. A deficiency common to all feedforward networks, when considered as models for dynamic systems, is their inability to conserve previous outputs and states for future predictions. Since this absence of dynamic capability requires the user to identify the order of the system prior to training and is therefore not entirely self-learning, more advanced network topologies are investigated. The most versatile of these structures, known as a fully recurrent network, re-uses the previous state of each of its nodes for subsequent outputs. However, despite its superior modelling capability, the tests performed using the Williams and Zipser training algorithm show that such structures often fail to converge and require excessive computing power and time, when increased in size. Despite its rigid structure and lack of dynamic capability, the regular Gaussian network produces the most reliable and robust models and was therefore selected for the evaluations in this study. To overcome the network initialization problem, found when using a pure neural network model, a combination structure· _in which the network operates in parallel with a mathematical model is suggested. This approach allows the controller to be implemented without any prior network training and initially relies purely on the mathematical model, much like conventional approaches. The network portion is then trained during on-line operation in order to improve the model. Once trained, the enhanced model can be used to improve the system response, since model exactness plays an important role in the control action achievable with model based structures. The applicability of control structures based on neural network models is evaluated by comparing the performance of two network approaches to that of a linear structure, using a simulation of a nonlinear tank system. The first network controller is developed from the internal model control (IMC) structure, which includes a forward and inverse model of the system to be controlled. Both models can be replaced by a combination of mathematical and neural topologies, the network portion of which is trained on-line to compensate for the discrepancies between the linear model _ and nonlinear system. Since the network has no dynamic ·capacity, .former system outputs are used as inputs to the forward and inverse model. Due to this direct feedback, the trained structure can be tuned to perform within limits not achievable using a conventional linear system. As mentioned previously the IMC structure uses both forward and inverse models. Since the control law requires that these models are exact inverses, an iterative inversion algorithm has to be used to improve the values produced by the inverse combination model. Due to deadtimes and right-half-plane zeroes, many systems are furthermore not directly invertible. Whilst such unstable elements can be removed from mathematical models, the inverse network is trained directly from the forward model and can not be compensated. These problems could be overcome by a control structure for which only a forward model is required. The neural predictive controller (NPC) presents such a topology. Based on the optimal control philosophy, this structure uses a model to predict several future outputs. The errors between these and the desired output are then collected to form the cost function, which may also include other factors such as the magnitude of the change in input. The input value that optimally fulfils all the objectives used to formulate the cost function, can then be found by locating its minimum. Since the model in this structure includes a neural network, the optimization can not be formulated in a closed mathematical form and has to be performed using a numerical method. For the NPC topology, as for the neural network IMC structure, former system outputs are fed back to the model and again the trained network approach produces results not achievable with a linear model. Due to the single network approach, the NPC topology furthermore overcomes the limitations described for the neural network IMC structure and can be extended to include multivariable systems. This study shows that the nonlinear modelling capability of neural networks can be exploited to produce learning control structures with improved responses for nonlinear systems. Many of the difficulties described are due to the computational burden of these networks and associated algorithms. These are likely to become less significant due to the rapid development in computer technology and advances in neural network hardware. Although neural network based control structures are unlikely to replace the well understood linear topologies, which are adequate for the majority of applications, they might present a practical alternative where (due to nonlinearity or modelling errors) the conventional controller can not achieve the required control action.