Binary Tree SVM Based Framework for Mining Fatigue Induced Damage Attributes in Complex Lug Joints Clyde K. Coelho *a , Santanu Das b , Aditi Chattopadhyay a , a Department of Mechanical and Aerospace Engineering, Arizona State University, Tempe, Arizona, USA 85287-6106 b NASA Ames Research Center, Building 269, Moffett Field, CA 94035 ABSTRACT Research is being conducted in damage diagnosis and prognosis to develop state awareness models and residual useful life estimates of aerospace structures. This work describes a methodology using Support Vector Machines (SVMs), organized in a binary tree structure to classify the extent of a growing crack in lug joints. A lug joint is a common aerospace ‘hotspot’ where fatigue damage is highly probable. The test specimen was instrumented with surface mounted piezoelectric sensor/actuators and then subjected to fatigue load until failure. A Matching Pursuit Decomposition (MPD) algorithm was used to preprocess the sensor data and extract the input vectors used in classification. The results of this classification scheme show that this type of architecture works well for categorizing fatigue induced damage (crack) in a computationally efficient manner. However, due to the nature of the overlap of the collected data patterns, a classifier at each node in the binary tree is limited by the performance of the classifier that is higher up in the tree. Keywords: Damage classification, support vector machines, matching pursuit decomposition, structural health monitoring, binary tree. INTRODUCTION Diagnosing faults that could lead to catastrophic failures in service structures is of utmost importance in structural health monitoring (SHM). An effective early warning system will allow the operators of these systems to do preventative maintenance and make decisions about the remaining useful life of the structure. Lug joints have been identified as structural hotspots in various aerospace systems due to their complex geometries that lead to multiple stress concentration points 1 . Subjecting these samples to fatigue loading showed different failure points depending on the applied load and surface finish. This highlights the need for robust inspection schemes on such components. The health monitoring scheme used in this research is based on the excitation and reception of guided Lamb waves in a structure by piezoelectric transducers. Support Vector Machines, an advanced machine learning based technique, has been utilized in this paper because it provides good generalization 2 when few samples are available for training purposes. Also, SVMs have a strong mathematical foundation 3 and has been widely used for classification in a number of fields. Although the original design of SVMs by Vapnik consisted of a two class structure, this approach has been expanded to do multiclass classification. Classification of multiclass problems has been conducted in the form of ‘one vs one’, ‘one vs rest’, hybrid 4, 5 and clustering 6 algorithms. In ‘one vs one’, the amount of training time required is very large since ௞ሺ௞ଵሻ ଶ classifiers need to be constructed for a k class problem. For a ‘one vs rest’ scheme, a problem involving k classes of data will require the construction of k classifiers. One problem with the latter case is that each classifier will require the use of the entire training set which becomes computationally intractable. For both methods, a voting scheme is used in which the classifier that scores the highest for a given data set assigns all the points in that set to a given class. Also, in such a case, it is very difficult to decide which class the test data belongs to if two classifiers have similar scores. Clustering schemes are able to learn signal characteristics well and can decide the uniqueness of different classes (or even classes within classes) based on the clustering of data points in hyperspace. While this approach is promising for damage detection scenarios where all possible damage types cannot be known, the computational expense involved with determining the cluster boundaries increases exponentially with the increase in training sets. *clyde.coelho@asu.edu; phone 1 480 965 2434; fax 1 480 965 1384