D.S. Kozlov, Yu.V. Tyumentsev

In this paper basic steps of failure identification procedure are considered as an appropriate combination of an observation task for an aircraft simulation model together with a classification task for a failure event. The tasks are performed using various types of heterogeneous NARX networks. A complete nonlinear model of a fighter aircraft is =considered as an object of diagnosis. The failures under consideration can be divided into two groups: aircraft actuator failures and sensor failures. All neural network models are derived from a heterogeneous nonlinear autoregressive network with exogeneous inputs. The main steps of the identification procedure are failure detection and failure isolation. The failure detection can be interpreted as a realization of an observation task for the plant model. NARX network with series-parallel architecture was chosen for the plant approximation. This model is applied to implement failure detection for both types of the failures. A modified version of Levenberg-Marquardt algorithm is used for the neural network training. The algorithm takes into account additional parameters of the neuron activation function. The classification task tadel advantage of the analytical redundancy existing in the system. NARX networks with parallel architecture were chosen. A modified version of Forward Perturbation algorithm is used for the neural network training. Sensor failure isolation is carried out by means of virtual sensors i. e. neural network models which predict angular rates cross-correlation functions. The implementation of the failure identification procedure was performed using MATLAB Neural Network Toolbox 6.0. The results of simulationfor sensor and actuator failures with regard to a fighter aircraft are presented to demonstrate efficeitncy of the proposed approach.