Development of Fatigue Damage Model of Wide-Band Process by Artificial Neural Network

Abstract

For the frequency-domain spectral fatigue analysis, the probability mass function of stress range is essential for the assessment of the fatigue damage. The probability distribution of the stress range in the narrow-band process is known to follow the Rayleigh distribution, however the one in the wide-band process is difficult to define with clarity. In this paper, in order to assess the fatigue damage of a structure under wide band excitation, the probability mass function of the wide band spectrum was derived based on the artificial neural network, which is one of the most powerful universal function approximation schemes. To achieve the goal, the multi-layer perceptron model with a single hidden layer was introduced and the network parameters are determined using the least square method where the error propagates backward up to the weight parameters between input and hidden layer. To train the network under supervision, the varieties of different wide-band spectrums are assumed and the probability mass function of the stress range was derived using the rainflow counting method, and these artificially generated data sets are used as the training data. It turned out that the network trained using the given data set could reproduce the probability mass function of arbitrary wide-band spectrum with success.

Novel models for fatigue life prediction under wideband random loads based on machine learning

“…For FDFA, some researchers 2,5,6,8,18,26,27 have devised various frequency-domain models. The TB and Dirlik models are well known for their excellent performance.…”

“…At present, machine learning models have been applied extensively in the field of fatigue life prediction. 16,17 Kang et al 18 developed an ANN model to fit the complex relationship between the 0th-order spectral moment and shape parameters of a power spectrum and the rainflow amplitude probability function. Durodola et al 15 estimated wideband random fatigue life using an ANN model with the spectral moments of a power spectrum and material parameters as inputs and the fatigue damage as an output.…”

Fatigue modeling using neural networks: A comprehensive review

“…The relative error of the NN is within 10%. Other studies using FNN for fatigue detection can be found in previous works 171,172,179–200 …”