Artificial neural network for Gaussian and non-Gaussian random fatigue loading analysis

Abstract

There has been a lot of work done on the analysis of Gaussian loading analysis perhaps because its occurrence is more common than non-Gaussian loading problems. It is nevertheless known that non-Gaussian load occurs in many instances especially in various forms of transport, land, sea and space. Part of the challenge with non-Gaussian loading analysis is the increased number of variables that are needed to model the loading adequately. Artificial neural network approach provides a versatile means to develop models that may require many input variables in order to achieve applicable predictive generalisation capabilities. Artificial neural network has been shown to perform much better than existing frequency domain methods for random fatigue loading under stationary Gaussian load forms especially when mean stress effects are included. This paper presents an artificial neural network model with greater predictive capability than existing frequency domain methods for both Gaussian and non-Gaussian loading analysis. Both platykurtic and leptokurtic non-Gaussian loading cases were considered to demonstrate the scope of application. The model was also validated with available SAE experimental data, even though the skewness and kurtosis of the signal in this case were mild.