Algorithme intelligent d'optimisation d'un design structurel de grande envergure

The implementation of an automated decision support system in the field of design and structural optimisation can give a significant advantage to any industry working on mechanical designs. Indeed, by providing solution ideas to a designer or by upgrading existing design solutions while the designer is not at work, the system may reduce the project cycle time, or allow more time to produce a better design.This approach was developed in order to reduce the operating cost of the process. However, as the system implementation cost is quite expensive, the approach is better suited for large scale design problem, and particularly for design problems that the designer plans to solve for many different specification sets.First, the CBR process uses a databank filled with every known solution to similar design problems. Then, the closest solutions to the current problem in term of specifications are selected. After this, during the adaptation phase, an artificial neural network (ANN) interpolates amongst known solutions to produce an additional solution to the current problem using the current specifications as inputs.Each solution produced and selected by the CBR is then used to initialize the population of an island of the genetic algorithm. The algorithm will optimise the solution further during the refinement phase.This thesis presents a new approach to automate a design process based on Case-Based Reasoning (CBR), in combination with a new genetic algorithm named Genetic Algorithm with Territorial core Evolution (GATE).Using progressive refinement, the algorithm starts using only the most important variables for the problem. Then, as the optimisation progress, the remaining variables are gradually introduced, layer by layer.The genetic algorithm that is used is a new algorithm specifically created during this thesis to solve optimisation problems from the field of mechanical device structural design. The algorithm is named GATE, and is essentially a real number genetic algorithm that prevents new individuals to be born too close to previously evaluated solutions. The restricted area becomes smaller or larger during the optimisation to allow global or local search when necessary.Also, a new search operator named Substitution Operator is incorporated in GATE. This operator allows an ANN surrogate model to guide the algorithm toward the most promising areas of the design space.The suggested CBR approach and GATE were tested on several simple test problems, as well as on the industrial problem of designing a gas turbine engine rotor's disc.These results are compared to other results obtained for the same problems by many other popular optimisation algorithms, such as (depending of the problem) gradient algorithms, binary genetic algorithm, real number genetic algorithm, genetic algorithm using multiple parents crossovers, differential evolution genetic algorithm, Hookes & Jeeves generalized pattern search method and POINTER from the software I-SIGHT 3.5.Results show that GATE is quite competitive, giving the best results for 5 of the 6 constrained optimisation problem. GATE also provided the best results of all on problem produced by a Maximum Set Gaussian landscape generator. Finally, GATE provided a disc 4.3% lighter than the best other tested algorithm (POINTER) for the gas turbine engine rotor's disc problem.One drawback of GATE is a lesser efficiency for highly multimodal unconstrained problems, for which he gave quite poor results with respect to its implementation cost.To conclude, according to the preliminary results obtained during this thesis, the suggested CBR process, combined with GATE, seems to be a very good candidate to automate and accelerate the structural design of mechanical devices, potentially reducing significantly the cost of industrial preliminary design processes.