Neural Remodelling of Objects with Variable Structures

The paper proposes the approach to uniformly present complicated objects and processes with a variable structure. The approach is based on the approximation of specific models, describing the object or the process by a uniformed remodeling class. The application of this approach is useful while solving optimization and control problems. Neural network models, which proved their high approximating capability, are suggested as a remodelling class. Applying the given approach is considered on the example of modelling of inertial torque transformer (ITT) workflow. This process has a cyclical pattern, where each phase of the cycle is divided into four segments, described by various systems of nonlinear differential equations with the same parameters. Moreover, the solution to the system describing the next segment depends on the solution to the system obtained by the previous segment. It noticeably makes it difficult to determine the optimum ITT parameters, as the fit function has the solution to the system of equations, describing the last, fourth segment of the cycle. The neural network model allows simplifying the solution to the given problem for each segment of the cycle. The input layer of the remodelling neural network was supplemented by taking into account the real output values from the previous moments of time. The neural network with such kind of structure demonstrated high level of accuracy.