Date of Award

Fall 2012

Document Type

Thesis - Open Access

Degree Name

Master of Science in Aerospace Engineering


Graduate Studies

Committee Chair

Dr. Fady F. Barsoum

First Committee Member

Dr. Yi Zhao

Second Committee Member

Dr. Daewon Kim


The purpose of this research is to develop a methodology and technique to determine the optimal number of clusters in acoustic emission (AE) data obtained from a ground test stand of a rotating H-60 helicopter tail gearbox by using mathematical algorithms and visual inspection. Signs of fatigue crack growth were observed from the AE signals acquired from the result of the optimal number of clusters in a data set. Previous researches have determined the number of clusters by visually inspecting the AE plots from number of iterations. This research is focused on finding the optimal number of clusters in the data set by using mathematical algorithms then using visual verification to confirm it. The AE data were acquired from the ground test stand that simulates the tail end of an H-60 Seahawk at Naval Air Station in Patuxant River, Maryland. The data acquired were filtered to eliminate durations that were greater than 100,000 ìs and 0 energy hit data to investigate the failure mechanisms occurring on the output bevel gear. From the filtered data, different AE signal parameters were chosen to perform iterations to see which clustering algorithms and number of outputs is the best. The clustering algorithms utilized are the Kohonen Self-organizing Map (SOM), k-mean and Gaussian Mixture Model (GMM). From the clustering iterations, the three cluster criterion algorithms were performed to observe the suggested optimal number of cluster by the criterions. The three criterion algorithms utilized are the Davies-Bouldin, Silhouette and Tou Criterions. After the criterions had suggested the optimal number of cluster for each data set, visual verification by observing the AE plots and statistical analysis of each cluster were performed. By observing the AE plots and the statistical analysis, the optimal number of cluster in the data set and effective clustering algorithms were determined. Along with the optimal number of clusters and effective clustering algorithm, the mechanisms of each cluster can be determined from the statistical analysis as well. From the results, the 5 cluster output using the Kohonen SOM clustering algorithm showed the distinct separation of clusters. Using the determined number of clusters and the effective clustering algorithms, the AE data sets were analyzed for the fatigue crack growth. Recorded data from the mid test and end test of the data acquisition period were utilized. After each set of clusters were associated with different mechanisms dependent on their AE characteristics. It was possible to detect the increase in the activities of the fatigue crack data points. This indicates that the fatigue crack is growing as the acquisition continued on the H-60 Seahawk ground test stand and that AE has a good potential for early crack detection in gearbox components.