Structural damage severity classification from time-frequency acceleration data using convolutional neural networks

Abstract

Structural damage identification involves interpretation of vibration measurements to identify patterns that are indicative of changes in structural characteristics. Identification of such patterns using machine learning (ML) algorithms is a primary task in data-driven structural health monitoring. However, providing ML algorithms with features that are sensitive to damage has always been a challenge for accurate damage detection. In recent years, deep Convolutional Neural Networks (CNNs) proved to be versatile and powerful tools to identify the underlying patterns directly from low-level raw data. This study explores the applicability of CNNs to damage severity classification. A new format of input data for CNN damage identifiers is proposed based on time–frequency representations of acceleration data from multiple sensors. First, Continuous Wavelet Transformation (CWT) of acceleration response is employed to derive the time–frequency data, then, the results of transformations from multiple accelerometers are stored in separate channels creating a three-dimensional block of data. Such input data provide a CNN with structural information that expands into time, frequency, and spatial domains. To evaluate the accuracy of such a CNN-based classifier, a dataset comprised of free vibration response of a concrete beam under impact hammer tests is utilized. It is shown that a CNN trained with CWT data can successfully classify various severity of the damage, from minor to severe, with 100% accuracy. Also, this study examines the sensitivity of the CNN classifier to the number of training samples. It is demonstrated that the accuracy of such a network has a relatively lower sensitivity to the number of training samples and/or the number of measurement points to the degree that it maintains an F1-score above 0.813 when trained with data from one impact point and one sensor. On the other hand, for the case of a CNN classifier trained directly with time-series acceleration data the value of the F1-score drops to 0.590. Moreover, when compared with a multilayer perceptron trained with the extracted modal properties, it is illustrated that CNN classifiers have a better capability in capturing key features from the measurement, and thus produce more accurate predictions. Therefore, the combination of CNN classifiers with time–frequency data from multiple sensors results in damage identifiers that are less sensitive to sample size and more robust to changes in excitations and measurement points.