Minimizing transmission error with neural networks and genetic algorithm with quadratic mutation rates

Abstract

Transmission error quantifies the rotation delay between driving and driven gear, caused by the disturbances of inevitable factors such as elastic deformation, manufacturing error, and mesh misalignment in assembly. A major consequence of high transmission error is the distinguishable whining noise, which indicates wear and efficiency loss. The transmission error (TE) becomes a further serious concern in case of electric vehicles where there is no background noise of internal combustion engine to diminish the gear noise and thus, gears trains become a major, if not only, source of noise in electric vehicles. Hence minimizing TE is a major concern for modern powertrain systems. This research work puts forth a novel approach for finding the optimal mutually independent microgeometry parameters which contribute to minimal transmission error. A neural network learns from the model specific generated data and gives out an approximating function which is then fed to a genetic algorithm which finds the global minima of the output function and therefore the values of the corresponding microgeometry parameters. The neural network and genetic algorithm were programmed from scratch in Python 3.7. The genetic algorithm was further tested with uniform mutation rates and variable mutation rates to find the degree of possibility of premature convergence. The former was found more vulnerable to premature convergence.