Classification-based Reliability Assessment with Semi-Supervised Learning

Abstract

In the proposed research, the accuracy of a classification-based surrogate model for reliability assessment is improved by augmenting the training data with a large number of unlabeled data. The implemented algorithm, Semi-Supervised Learning (SSL) has the ability to drastically reduce the computational cost in the reliability assessment process which often requires simulation tools. This feature is considerably important because obtaining class labels for a data point in reliability analysis requires the utilization of the simulation tool which is highly computationally intensive. A combination of Probabilistic Neural Network (PNN) and Expectation-Maximization (EM) algorithm is considered to use the labeled and unlabeled data simultaneously in order to improve the accuracy of the PNN classifier. The proposed algorithm first trains a PNN classifier using the labeled data. This classifier is then used to probabilistically label the unlabeled data. Then, a new PNN classifier is trained using both labeled and unlabeled data. This procedure is iterated until convergence. The example problem of a ten bar truss shows the efficacy of the proposed procedure in estimating the probability of failure (Pf). It clearly demonstrates considerable improvements in the classifier performance when both labeled and unlabeled data are used without requiring additional computational costs.