Access to master’s theses is restricted to current ERAU students, faculty, and staff.

Date of Award


Document Type

Thesis - Open Access

Degree Name

Master of Aerospace Engineering


Graduate Studies

Committee Chair

Dr. Eric v. K. Hill

Committee Member

Dr. Frank Radosta

Committee Member

Dr. Habib Eslami


This research demonstrates how acoustic emission (AE) data from flaw growth activity can be used to predict burst pressures in filament wound composite pressure vessels. Acoustic emission data were taken during hydroproof testing for a set of eleven ASTM standard 5.75 inch diameter fiberglass/epoxy bottles. Amplitude distributions were created using only the AE data up to 25% of the expected burst pressure to simulate low level proof loadings (thereby avoiding damage to the bottles). The bottles were tested at three different temperatures -- 32°F, 70°F, and 110°F - and hydroproofed using two different pressurization schemes and two transducer configurations. Moreover, two of the bottles contained simulated manufacturing defects which lowered their burst pressures significantly.

Both multivariate statistical analysis and artificial neural networks were used to generate burst pressure prediction models from the AE amplitude distribution data. For the multivariate statistical analysis, fixed failure mechanism bands were applied to the amplitude distributions for all eleven fiberglass/epoxy bottles. The optimum failure mechanism bands resulted in a prediction equation that had a worst case prediction error of-14% and a correlation coefficient of 49.9%. When the defective bottles were left out of the analysis, the results improved to a +10% worst case error and a 65.0% correlation coefficient.

The amplitude distribution frequencies, temperatures, pressurization schemes, and transducer configurations were all used as inputs to the artificial neural networks. To begin with, the two pressurization schemes and the two transducer configuration schemes were found to have no significant effect on prediction accuracies. When one of the defective bottles was included in the training set and the other in the test set, the errors were +15.2% and +14.7%, depending upon which bottle was used for training and which for testing. Including both defective bottles in the fraining set decreased the worst case prediction error to -7.8%. Finally, when the two defective bottles were removed from consideration, the worst case prediction errors dropped to -0.8% and -1.5%, the former being obtained with temperature included as an independent variable and the latter without.

The neural networks predicted burst pressures to a greater accuracy than multivariate statistical analysis. This could be explained by the fact that the statistical analysis generates a linear burst pressure equation, whereas the neural network is not limited to linear modeling. The neural network results suggest that the addition of one defective bottle in the training set would probably allow the neural network to predict burst pressures with a worst case error within the desired goal of ±5.0%. It can also be seen that, while the inclusion of temperature as one of the inputs to the neural network improves the prediction accuracy, the improvement does not appear to be significant.