A general framework for supervised structural health monitoring and sensor output validation mitigating data imbalance with generative adversarial networks-generated high-dimensional features

Abstract

This study proposes a novelty-classification framework that applies to structural health monitoring (SHM) and sensor output validation (SOV) problems. The proposed framework has simple high-dimensional features with several advantages. First, the feature extraction method is extensively applicable to instrumented structures. Second, the high-dimensional features’ utilization alleviates one of the main issues of supervised novelty classifications, namely, imbalanced datasets and low-sampled data classes. Recurrent Neural Networks are employed for the classification of high-dimensional features. Furthermore, generative adversarial networks (GAN) are trained with low-sampled data classes’ high-dimensional features for generating new data objects. The generated data objects are combined with the initial training set for improving classification results. The proposed framework is studied on two SHM and SOV datasets. The SHM dataset has twenty-one data classes, with a total test accuracy of 99.60% compared to another study with 88.13% accuracy. The SOV classification shows improved results with a mean accuracy of 96.5% compared to three other studies with mean accuracy values of 93.5%, 92.97%, and 71.1%. Furthermore, the integration of GAN’s generated data objects with low-sampled classes improved those classes’ mean F1 score from 44.77% to 64.58% and from 73.39% to 90.84% on SOV and SHM case studies, respectively. The integration of GAN-generated data objects with the initial low-sampled data classes for accuracy improvement shows more potential in the SHM dataset than the SOV case, which can be due to the signal pattern-based labeling logic of SOV datasets.

System-reliability based multi-ensemble of GAN and one-class joint Gaussian distributions for unsupervised real-time structural health monitoring

“…There is always the possibility that the Generator is not trained well for all parts of that SGD, resulting in misleading generations. This phenomenon is shown in Soleimani et al [9]. In that study, generated data objects were "Capped" to avoid misleading generations to enter the analysis.…”

“…The first feature (F I) is high-dimensional (N × D L /2), made of magnitudes of half-spectrum FFTs of input signals with maximum values suppressed to ten. As discussed in Soleimani et al [9], the features to vary in a fixed-range is beneficial for the GAN's training. The second feature (F II) is a reduced F I representation, made of quartiles of vibrational energy in each time-series data object, which is tried on prior studies [23].…”

“…Gathering enough data objects of different structural conditions from facilities in operation is costly and impractical in most cases. Some efforts have been made to alleviate this issue, such as having only two classes of an undamaged and fully damaged system for performing the damage detection [7], or generating data objects from low-sampled damage classes via 1-D Generative adversarial networks (GAN) to improve the robustness of classifiers [9]. Unfortunately, obtaining fully-damaged states is a big challenge in itself.…”

“…In SHM, on the other hand, 1-D data are of more interest (i.e., frequency or time-domain data series). The 1-D GAN idea for constructing lost sensor data [13], and the classification improvement of low-sampled damage scenarios [9] are examples of 1-D GAN implementations in SHM.…”

Toward a general unsupervised novelty detection framework in structural health monitoring

“…High dimensional, domain‐specific features obtained through simple feature extraction methods (e.g., fast Fourier transform of a signal) can address the structure‐specific feature extraction challenges if paired with the right detection architecture. In supervised settings, high‐dimensional features have been used with 1D convolutional neural networks (CNNs) (Abdeljaber et al., 2018; Azimi & Pekcan, 2020), 2D CNN (Yu et al., 2019), 2D CNN (Yu et al., 2019), or with long short‐term memory (LSTM) units, a variation of recurrent neural networks (RNN) (Soleimani‐Babakamali, Soleimani‐Babakamali, & Sarlo, 2021). Specific target functions are required to employ such deep architectures in unsupervised settings.…”