Development of machine learning based classifier for the pressure test result prediction of type IV composite overwrapped pressure vessels

Abstract

The stringent safety regulations of type IV composite overwrapped pressure vessels (COPVs) for commercial vehicles mandate a certification process involving pressurization up to 1050 bar, with the critical requirement of withstanding burst pressures of 1570 bar. Analyzing proof test data is crucial to enhance and ensure tank safety regarding burst pressure. In this study, we developed various machine learning classifiers for structure health monitoring and damage prediction of COPVs. The classifiers were trained using a substantial amount of acoustic emission data collected during burst and pressure cycling tests. The test results were employed as label inputs during the training process. Statistical features were extracted per time unit and trained using Naive Bayes, Logistic Regression, Decision Tree, XGBoost, and TabNet models. Upon training the data collected from the burst pressure test, TabNet, Decision Tree, and XGBoost achieved classification accuracies above 0.94. Notably, TabNet demonstrated also the best performance for the pressure cycling test with an accuracy of 0.98. Furthermore, TabNet provided visualizations of feature sensitivity in relation to classification results. This study marks the first development of a machine learning classifier for predicting the damage state of COPV tanks in commercial applications pertaining to required safety tests.