Unsupervised learning-based data cleaning framework for structural health monitoring using GoogLeNet and clustering models

Abnormal data in structural health monitoring (SHM) system can have a significant impact on the assessment of the true condition of a structure. Most existing studies are costly and need the manual screening of training data, which have significantly hindered the development of automated and efficient cleaning of monitoring data. To address this, an unsupervised learning-based data cleaning framework for bridge structural health monitoring is proposed. The continuous wavelet transform is used to convert monitoring data into wavelet time–frequency (WTF) images. A pre-trained GoogLeNet model is used as a feature extractor to capture the features from the WTF images. The extracted high-dimensional features are clustered using K-means algorithm. The results of cluster assignments serve as pseudo-labels, providing initial supervised labels for subsequent model refinement. The main-girder acceleration data from the SHM system of a large-span cable-stayed bridge is taken as an example to verify the effectiveness of the proposed framework. In the large-scale test, the precision of normal data is 98.4%. In addition, the proposed framework can also accurately clean all data types with an accuracy of 97.9%. The results show that all data types can be accurately detected. Meanwhile, the proposed framework does not require any manual prior screening of the dataset and accomplishes automatic cleaning of monitoring data at low cost.