Date of Award

Fall 1998

Document Type

Thesis - Open Access

Degree Name

Master of Science in Aerospace Engineering


Aerospace Engineering

Committee Chair

Eric v. K. Hill

Committee Member

Deborah M. Osborne

Committee Member

Christopher J. Raczkowski


Edge-welded metal bellows present an ongoing challenge: the prediction of an accurate cycle life. Current methods rely on physical leak detection to determine a bellow's cycle life to failure. It is known, however, that crack initiation begins many cycles before a leak path is present. Bellows manufacturers require a method for detection of fatigue cracks when they initiate but before they result in leak rates large enough to contaminate a process. Acoustic emission (AE) testing is one method which can meet this need and is a proven, reliable technique for detecting crack initiation and monitoring fatigue crack growth.

Four sets of metal bellows samples were fatigue tested and AE parameter data recorded. The data sets were analyzed and the determination made that amplitude, duration, and time of occurrence were the AE data variables required for separation of the various failure mechanisms.

For two of the four materials, an expanded set of tests were performed. Fourteen tests were used to train and test a back-propagation neural network for prediction of bellows cycle life. The input data consisted of a material identifier, AE parameter data consisting of the amplitude distribution (50-100 dB) of the first 250 hits, and the final cycle life. The network was structured with an input layer consisting of the identifier and amplitude data, two hidden layers for mapping failure mechanisms, and an output layer for predicting cycle life. The network required training on four samples for the Inconel 718 and five samples for the 350 stainless steel. Once trained the network was able to predict cycle life with a worst case error of-4.45 percent and 2.66 percent for the Inconel 718 and 350 stainless steel, respectively.

Finally, through the use of multiple linear regression, a statistical analysis was made to develop a model capable of accurate prediction. Applying a natural log transformation to the independent variables of amplitude and energy resulted in a model capable of explaining 95 percent of the variability in cycle life prediction.