Characterization of Impact Damage Uncertainty of Composite Materials Based on Bayesian Neural Network

To solve the problem of multi-source parameter coupling and damage dispersion quantification in the uncertainty characterization of impact damage in composite materials, this paper proposes a confidence interval quantification method based on a Bayesian neural network to realize the propagation analysis of material parameters and impact energy uncertainty to damage state. Firstly, based on Bayesian inference, the probability distribution of weight and threshold parameters was constructed, and the probability distribution of the predicted value was directly output through the weight posterior distribution sampling, and the traditional point estimation was transformed into probabilistic interval prediction. Secondly, a hierarchical Monte Carlo sampling strategy was designed, and the input material parameters were dispersed to inject the uncertainty of the damage state to generate a probability envelope covering 95% of the measured data points. Finally, the confidence interval curve of pit depth and damage area is predicted, including the confidence interval curve of pit depth and damage area, and the covariant relationship between the two under energy change is revealed, and the results show that the pit depth and damage area increase nonlinearly with the increase of impact energy, but show different sensitivities, the pit depth increases slowly in the low-energy segment (<50J), and the growth rate of the high-energy segment increases significantly, while the damage area expands exponentially before 50J.