Neural-Based Data Processing in Intelligent Distributed Sensor Network V.Turchenko, V.Kochan, A.Sachenko Research Laboratory of Automated Systems and Networks, Ternopil Academy of National Economy, 3 Peremoga Square, Ternopil, 46004, UKRAINE, e-mail: vtu@tanet.edu.te.ua Abstract The experimental researched results of sensor drift prediction by proposed earlier neural-based methods of (i) additional approximating neural network and (ii) integration of a priori data within an Intelligent Distributed Sensor Network (IDSN) are presented in this paper. The three-level structure of IDSN hardware is grounded based on estimation of necessary productivity of computational devices. There are developed the recommendations of using uniprocessor or dual-processor hardware structure of the Intelligent Node as the middle IDSN level. 1: Introduction The modern systems of sensor data acquisition and processing have a scheme of distributed networks [1, 2] as a rule. Thus the dominant error, which strongly influences on reliability of physical quantities calculation, is the sensor error, in particularly, drift of sensor conversion characteristics [3]. The methods of reliability improvement of sensor data acquisition and processing, in particularly testing and calibration require interruption of data acquisition process. They have high laborious and cost that limits their wide usage. A prediction method [3] has not this lack and allows correcting sensor errors during system operation using testing results of similar-type sensors in the same exploitation conditions. However, specific features of sensors drift do not provide necessary accuracy of prediction at usage non-adaptive mathmodels. Therefore it is expediently to use these methods (calibration, testing and prediction) together with the purpose of increasing of time between failures, which corresponds to inter-testing interval [5]. In [6] it is proposed to use artificial neural networks for sensor drift prediction in order to improve both accuracy and reliability of sensor data acquisition and processing into IDSN. 2: Neural-based Methods of Sensor Drift Prediction Using neural network for sensor drift prediction causes the contradiction [7]: the high-quality neural network training allows sharply reduce prediction error and therefore increase the time between failures, that the data number of obtained calibration results will not provides high-quality training of neural network. It is proposed to use methods of (i) additional approximated neural network [8, 9] and method of a priori data integration [9-11]. The multiplayer perceptron (one input neuron, N hidden neurons with logistic activation function and one output neuron) has been used as approximating neural network (ANN) [8]. The algorithm of multiple propagation error was used for training [9]. The single – layer perceptron with Widrow-Hoff learning rule and multiplayer perceptron with back propagation error algorithm were used as integrating a priori data (IPD) neural network [10, 11]. The three-layer recurrent neural network (one hidden layer of nonlinear neurons and one linear output neuron) is used as predicting neural network (PNN) (Fig. 1). The output value of PNN is ∑ = - = h N i r i i r T h w Y 1 0 , where h N is the number of neurons of hidden layer, r i h are the output values of neurons of hidden layer in time moment r , T is the threshold of output neuron, 0 i w are weights from i - neurons to output neuron [9].