Fatigue Life Prediction Method of Ceramic Matrix Composites Based on Artificial Neural Network

Ceramic matrix composites have been widely applied in the aerospace field due to the excellent mechanical properties. However, the complex microstructure and failure mechanism bring great difficulties to the fatigue life analysis. Aiming at the problems that mesoscopic mechanical model and macroscopic phenomenological model requires a large amount of experimental data and the model parameters are difficult to obtain, a fatigue life analysis method for composites materials based on Artificial Neural Network (ANN) is proposed. In the neural network, material parameters and loading parameters are used as inputs, and fatigue life is used as output to build a model of the relationship between input parameters and fatigue life. Comparative investigation of different neural networks in the fatigue life prediction of two-dimensional braided ceramic matrix composites under the condition of small samples are investigated. Results show that Elman Network (ENN) and Convolutional Neural Network (CNN) can obtain high-precision prediction results under different training data volumes when using simulation data sets. The prediction accuracy decreases with the reduction of data volume, while the prediction accuracy of Generalized Regression Neural Network (GRNN) is difficult to meet the requirements. Taking the experimental data from literature as the data set, ENN and CNN are used for training, and good prediction are obtained under the condition of using only 4 S-N curves as the training set.

Automated summary

Ceramic matrix composites are widely used in the aerospace field due to their excellent mechanical properties. However, analyzing their fatigue life is challenging due to the complex microstructure and failure mechanism. To address this issue, a fatigue life analysis method based on Artificial Neural Network (ANN) is proposed, which utilizes material parameters and loading parameters to predict fatigue life. The study compares different neural networks and finds that Elman Network (ENN) and Convolutional Neural Network (CNN) provide high-precision predictions using simulation data sets. The Generalized Regression Neural Network (GRNN) fails to meet the requirements. Experimental data from literature are used to train ENN and CNN, and good predictions are achieved using only 4 S-N curves as the training set.