Data-driven structural health monitoring techniques for damage detection have been widely studied. Most data-driven approaches analyze baseline signals recorded prior to deploying the structure in service and test signals acquired during service to detect damage in structures. Training such systems typically requires supervised learning algorithms and a large amount of data from healthy and damaged structures. While recording data from healthy structures is relatively easy and practical, collecting data from damaged structures is costly and in many situations, impractical. Moreover, the environmental and operating conditions (EOC) of the tests may be time varying. Consequently, testing structures based on a system that was trained on data recorded prior to deployment may not provide reliable damage detection performance. This speaks to the need for online learning algorithms. Furthermore, supervised online learning where the system knows the state of the structure is not practical in real-time structural health monitoring applications.&nbsp;In this paper, we propose an algorithm that can learn the state of the structure in time-varying environments and uses only signals from healthy structures to perform real time structural health monitoring. This approach is based on the support vector data description (SVDD) algorithm which uses signals from damage-free structures acquired in different EOCs to estimate a nonuniform hyper-sphere that describes the feature space of sensor signals recorded from healthy structures. Signals collected from damaged structures contain signatures that are typically different from those resulting from environmental changes. Features extracted from such signals will lie outside the feature space determined during training on healthy signals and will be classified as anomalies. A disadvantage of the SVDD algorithm is its inability to handle large streaming data common in real-time structural health monitoring applications.&nbsp;Fast incremental support vector data description (FISVDD) is a relatively new algorithm that has been shown to perform well in standard classification tasks on publicly available datasets. FISVDD is an iterative implementation of the traditional SVDD algorithm. Unlike the SVDD algorithm, the FISVDD algorithm analyzes only the incoming data samples during each iteration, along with the support vectors computed in the previous iteration. The computational cost of the algorithm arises primarily from incremental and decremental updates of the inverse of a similarity matrix between the new data samples and the existing set of support vectors. Because the FISVDD continuously updates the set of support vectors that describe the feature space of healthy signals, it computes a continuously evolving boundary that adapts to the changes in the test environment and the operating conditions. Similar to the SVDD algorithm, damage detection is treated as an anomaly detection problem using the adaptive boundary that is updated during each iteration.&nbsp;We evaluate the performance of the algorithm on experimental sensor signals acquired from an aluminum plate under different temperatures. We also compare the performance of the algorithm with the traditional SVDD algorithm on the same experimental data. Experimental results suggests that the FISVDD algorithm provides substantial gains in computational efficiency with almost no loss in damage detection accuracy when compared with the SVDD algorithm.&nbsp;

A Fast Unsupervised Online Learning Algorithm to Detect Structural Damage in Time-Varying Environments

Paper Proceedings: https://asmedigitalcollection.asme.org/qnde/QNDE2021/volume/85529

ASME 2021 48th Annual Review of Progress in Quantitative Nondestructive Evaluation (QNDE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.