Alejandro García

Hybrid Differential Neural Network Identifier for Partially Uncertain Hybrid Systems

This chapter presents a hybrid differential neural network (DNN)-identifier has demonstrated excellent results even in the presence of perturbations. Convergence analysis is realized considering the practical stability of identification error for a general class of hybrid systems. As can be seen in the numerical examples this algorithm could be easily implemented. In this sense the artificial modeling strategy of the continuous subsystems could be used in the automatic control implementation.

Adaptive linearization for nonlinear systems using continuous Neural Networks

The adaptive linearization of dynamic nonlinear systems remains as an open problem due to the complexities associated with the methods required to obtain the linearized sections. This problem is even more difficult if the system is uncertain, it means, if only partial or null information about the mathematical model of the system is available. This paper presents a proposal of an adaptive linearization method for uncertain nonlinear systems affected by additive perturbations by the Aritificial Neural Networks approach. The stability of the indentification error is formally boarded and proved by the second Lyapunovs method. Such suggested structure preserves some inherited structural properties that allows this method to behave as the original model as is exposed. A comparison of the developed algorithm with a similar structure without adaptable linear term is carried out, considering a genetic regulation mathematical model. The results of the simulation show that this proposal presents a superior performance as is observed in the trajectories of each identifier and by comparing the performance index of each one.

Differential Neural Networks Observers: Development, Stability Analysis and Implementation

The control and possible optimization of a dynamic process usually requires the complete on-line availability of its state-vector and parameters. However, in the most of practical situations only the input and the output of a controlled system are accessible: all other variables cannot be obtained on-line due to technical difficulties, the absence of specific required sensors or cost (Radke & Gao, 2006). This situation restricts possibilities to design an effective automatic control strategy. To this matter many approaches have been proposed to obtain some numerical approximation of the entire set of variables, taking into account the current available information. Some of these algorithms assume a complete or partial knowledge of the system structure (mathematical model). It is worth mentioning that the influence of possible disturbances, uncertainties and nonlinearities are not always considered. The aforementioned researching topic is called state estimation, state observation or, more recently, software sensors design. There are some classical approaches dealing with same problem. Among others there are a few based on the Lie-algebraic method (Knobloch et. al., 1993), Lyapunov-like observers (Zak & Walcott, 1990), the high-gain observation (Tornambe 1989), optimization-based observer (Krener & Isidori 1983), the reduced-order nonlinear observers (Nicosia et. al.,1988), recent structures based on sliding mode technique (Wang & Gao, 2003), numerical approaches as the set-membership observers (Alamo et. al., 2005) and etc. If the description of a process is incomplete or partially known, one can take the advantage of the function approximation capacity of the Artificial Neural Networks (ANN) (Haykin, 1994) involving it in the observer structure designing (Abdollahi et. al., 2006), (Haddad, et. al. 2007), (Pilutla & Keyhani, 1999). There are known two types of ANN: static one, (Haykin, 1994) and dynamic neural networks (DNN). The first one deals with the class of global optimization problems trying to adjust the weights of such ANN to minimize an identification error. The second approach, exploiting the feedback properties of the applied Dynamic ANN, permits to avoid many problems related to global extremum searching. Last method transforms the learning process to an adequate feedback design (Poznyak et. al., 2001). Dynamic ANN’s provide an

The Bergman's insulin-glucose regulation model: DNN-state observer

The nonlinear Bergman regulation model of insulin-glucose in plasma is analyzed using modern observability theory. The observability conditions are derived. The glucose concentration measurements in plasma or any its combinations with the other components are shown to provide the observability property, that is, they contain the complete information about the considered state space dynamic model. Because this model has several a priori unknown parameters and the measurable data may have noises, the dynamic neuro state observer is suggested to obtain immeasurable state estimates. Such an approach brings significant advantages with respect to the insulin doses decision, traditionally used in the normal insulin infusion pumps: it presents the opportunity to avoid any active patient actions, the numerical simulations illustrate the effectiveness of the suggested approach.

Projectional differential neural network observer with stable adaptation weights

A class of dynamic neural network (DNN) observers involving a projection operator inside is considered. Such observers seem to be useful when an uncertain nonlinear system, affected by external perturbations, keeps its states in an a priori known compact set, defined by the given state constraints independently of the measurement noise effects. Since the projection method introduces discontinuities into the trajectory dynamics, the standard Lyapunov method is not applicable to describe the convergence property of this class of observers. This problem is suggested to be resolved using a Lyapunov-Krasovski functional including both the estimation error and the weights involved in the DNN description. The stable adaptive laws for the DNN-weights adjustment are derived. The upper bound for the estimation error is obtained based on linear matrix inequality (LMI) technique implementation. An illustrative example clearly shows the effectiveness of the suggested approach. It deals with an environment control problem, related to the soil contaminants degradation by ozonation.

Projectional dynamic neural network observer

Abstract The projectional method is applied to describe a novel class of dynamic neural network (DNN) observers, which tourn out to be useful when an uncertain nonlinear system, affected by external perturbations keeps its states in an a priori known compact set, defined by given state constraints (usually having physical meaning) independently of the measurement noise effects. The learning law for the weights associated adjustment with the DNN observation problem is derived. The illustrative example dealing with the soil contaminants degradation problem, proves the nice workability of the suggested approach.

Iron-zinc oxide films prepared by ultrasonic spray pyrolisis

Starting from an aqueous nitrate solution, iron-zinc oxide films were prepared by the ultrasonic spray pyrolisis method. Films with different contents of zinc (2, 3, 30 and 50 at. %) were prepared by adding the appropriate zinc acetate amount to the solution. Sample deposited at several temperatures (50, 150, 250, 350 and 450 °C) were studied by means of optical (UV-vis and mid-infrared ranges) and XRD measurements. Several phases, FeO, α-Fe2O3 and ZnFe2O4, depending on the deposition temperature and Zn concentration, were observed. The Zn content in the film controlled the microstructure. A very rough surface morphology was found from AFM observations (RMS roughness ≫ 560 nm), making the films potentially useful for gas sensor applications.

Electrical behavior of Au/IrO 2 /Si heterostructures

Au/IrO 2 /Si heterostructures were built. Their DC current versus temperature characteristics were experimentally obtained to get the corresponding Richardson plots. From these plots, the Richardson constant was estimated for these devices. Then, from the current-voltage plots at room temperature the series resistance, ideality factor and barrier height were obtained by applying the method proposed by Cheung for parameter extraction from the thermionic theory. The model is found to fit reasonably the electrical behavior of the heterostructures for voltages higher than O B .