Multisource Data Fusion for Classification of Surface Cracks in Steel Pipes

This paper focuses on the development and validation of a robust framework for surface crack detection and assessment in steel pipes based on measured vibration responses collected using a network of piezoelectric (PZT) wafers. The pipe structure considered in this study contained multiple progressive cracks occurring at different locations and with various orientations (along the circumference or length). The fusion of data collected from multiple PZT wafers was investigated based on two approaches: (a) combining the raw data from all sensors before establishing a statistical model for damage classification and (b) combining the features from each sensor after applying a multiclass support vector machine recursive feature elimination (MCSVM-RFE), for dimensionality reduction, and taking the union of discriminative features among the different sources of data. A MCSVM learning algorithm was employed to train the data and generate a statistical classifier. The dataset consisted of ten classes, consisting of nine damage cases and the healthy state. The accuracy of the prediction based on the two fusion approaches resulted in a high accuracy, exceeding 95%, but the number of features needed to enrich the accuracy (95%) differed between the two approaches. Furthermore, the performance and the precision in the prediction of the classifier were evaluated when the data from only a single sensor was used compared with the combined data from all the sensors within the network. Very promising results in the classification of damage were obtained, based on the case study that included multiple damage scenarios with different lengths and orientations.