Automated Preliminary Structural Rotor Design Using Genetic Algorithms and Neural Networks

Abstract

It is important for any design process to have a good starting point in order to reduce the cycle time and the number of design iterations required. This paper presents an automated preliminary structural design system for a gas turbine rotor, using only preliminary aerodynamic data and a simplified structural analysis, with the objective of producing a good, feasible starting solution for the blades and the disc. The process starts with a CBR (case-based reasoning) algorithm coupled with a databank of existing solutions. The algorithm uses a neural network to choose from among the closest existing rotor solutions and interpolate between them. These designs, along with the interpolated solution, will constitute the initial set of possible designs. An adaptation algorithm then processes each possible design using simplified analysis to compute the estimated sensitivity of the design function with respect to each parameter in the neighbourhood of these design solutions. The algorithm uses those sensitivities to separate the design parameters into several layers according to their relative importance. In a following phase, these design solutions are used to train a surrogate neural network model for the function, and also as the starting population for a genetic algorithm (GA). The GA is then run, with the objective of minimizing the weight of the rotor while respecting stress and aerodynamics constraints. More parameter sets (beginning with the most important) are gradually added as an input for each GA run. Although this process would not be capable of replacing a detailed design system, as it currently uses only simplified analysis, it can provide a concept designer with a very good starting solution within a relatively short computing time.