Uncertainty quantification of damage localization based on a probabilistic convolutional neural network

Abstract

SHM is vital in quantitatively identifying engineered critical structural damage due to its potential economic and security interests. Convolutional Neural Network (CNN) is a popular method used for SHM on damage localization and classification. However, traditional CNN methods have limitations in predicting performance uncertainty and only provide point evaluations without indicating their accuracy. To address this issue, this paper introduces a PCNN framework, which combines a traditional CNN with a probabilistic layer to generate overall confidence intervals (CIs) for prediction results, as well as conditional probability distributions (CPDs) and likelihood for each prediction result. The PCNN method provides a manner to quantify the prediction uncertainty of neural networks and determine the confidence of each prediction. The paper also recommends using Leaky ReLU as the activation function, which retains negative value information. The effectiveness of the PCNN method is illustrated through case studies of carbon fiber-reinforced polymer beams with different layups. The results show that PCNN is effective in giving damage location prediction for CIs, CPDs and likelihood.