Toward Data-Driven Structural Health Monitoring: Application of Machine Learning and Signal Processing to Damage Detection

Abstract

A multilayer data-driven framework for robust structural health monitoring based on a comprehensive application of machine learning and signal processing techniques is introduced. This paper focuses on demonstrating the effectiveness of the framework for damage detection in a steel pipe under environmental and operational variations. The pipe was instrumented with piezoelectric wafers that can generate and sense ultrasonic waves. Damage was simulated physically by a mass scatterer grease-coupled to the surface of the pipe. Benign variations included variable internal air pressure and ambient temperature over time. Ultrasonic measurements were taken on three different days with the scatterer placed at different locations on the pipe. The wave patterns are complex and difficult to interpret, and it is even more difficult to differentiate the changes produced by the scatterer from the changes produced by benign variations. The sensed data were characterized by 365 features extracted from a variety of signal-processing techniques. Automated feature selection methods were then developed using an adaptive boosting algorithm to identify the most effective features for damage detection. With the selected features, five machine-learning classifiers were formulated based on adaptive boosting and support vector machines and achieved 98.5–99.8% average accuracy during random testing and 84.2–89% average accuracy during systematic testing. In addition, other metrics for classifier evaluation generated from a confusion matrix and from a receiver operating characteristic curve are reported.