This brief deals with nonlinear model predictive control designed for a tank unit. The predictive controller is realized by means of a recurrent neural network, which acts as a one-step ahead predictor. Then, based on the neural predictor, the control law is derived solving an optimization problem. An important issue in control theory is stability of the control system. In this brief, this problem is investigated by showing that a cost function is monotonically decreasing with respect to time. The derived stability conditions are then used to redefine a constrained optimization problem in order to calculate a control signal. As the automatic control system can prevent faults from being observed, the control system is equipped with a fault diagnosis block. It is realized by means of a multivalued diagnostic matrix, which is determined on the basis of residuals calculated using a set of partial models. Each partial model is designed in the form of a recurrent neural network. This brief proposes also a methodology of compensating sensor, actuator, as well as process faults. When a sensor fault is isolated, the system estimates its size and, based on this information, the controller is fed with a determined, close to real, tank level value. Actuator and process faults can be compensated due to application of an unmeasured disturbance model.
Recent approaches to fault detection and isolation for dynamic systems using methods of integrating quantitative and qualitative model information, based upon soft computing (SC) methods are surveyed and studied in some detail. SC methods are considered an important extension to the quantitative model-based approach for residual generation in fault detection and isolation (FDI). When quantitative models are not readily available, a correctly trained neural network (NN) can be used as a non-linear dynamic model of the system. The paper describes some powerful NN methods, taking into account the dynamic as well as non-linear system behaviour. Sometimes, further insight is required as to the explicit behaviour of the model-involved and it is then that fuzzy and even neurofuzzy methods come to their own in data-driven FDI applications. The paper also discusses the use of evolutionary programming tools for observer and NN design. The paper provides many powerful examples of the use of SC methods for achieving good detection and isolation of faults in the presence of uncertain plant behaviour, together with their practical value for fault diagnosis of real process systems.
Towards Robustness in Neural Network Based Fault Diagnosis: Challenging design problems arise regularly in modern fault diagnosis systems. Unfortunately, classical analytical techniques often cannot provide acceptable solutions to such difficult tasks. This explains why soft computing techniques such as neural networks become more and more popular in industrial applications of fault diagnosis. Taking into account the two crucial aspects, i.e., the nonlinear behaviour of the system being diagnosed as well as the robustness of a fault diagnosis scheme with respect to modelling uncertainty, two different neural network based schemes are described and carefully discussed. The final part of the paper presents an illustrative example regarding the modelling and fault diagnosis of a DC motor, which shows the performance of the proposed strategy.
7 May 2007 in IEEE Transactions on Neural Networks
This paper deals with problems of stability and the stabilization of discrete-time neural networks. Neural structures under consideration belong to the class of the so-called locally recurrent globally feedforward networks. The single processing unit possesses dynamic behavior. It is realized by introducing into the neuron structure a linear dynamic system in the form of an infinite impulse response filter. In this way, a dynamic neural network is obtained. It is well known that the crucial problem with neural networks of the dynamic type is stability as well as stabilization in learning problems. The paper formulates stability conditions for the analyzed class of neural networks. Moreover, a stabilization problem is defined and solved as a constrained optimization task. In order to tackle this problem two methods are proposed. The first one is based on a gradient projection (GP) and the second one on a minimum distance projection (MDP). It is worth noting that these methods can be easily introduced into the existing learning algorithm as an additional step, and suitable convergence conditions can be developed for them. The efficiency and usefulness of the proposed approaches are justified by using a number of experiments including numerical complexity analysis, stabilization effectiveness, and the identification of an industrial process