Efficient vision-based data anomaly detection for structural health monitoring

Structural health monitoring (SHM) has been widely-Applied to manage and maintain civil infrastructure, which produces huge amounts of data. A prerequisite task for all SHM systems is data cleansing due to systems malfunctions that are mostly induced by harsh operating conditions. A challenge within is to handle the "treasure or trash" of anomalous SHM data, which means that these data could be generated either by emergencies, such as earthquake and ship collision, or just system malfunctions. Human experts are experienced with domain knowledge, but manual inspection for data cleansing is inefficient to meet the demand for online processing. Furthermore, traditional automated signal processing techniques are generally single-Task for one certain kind of anomaly; also it is difficult for numerical parameters tuning of multiple algorithms with limited utilized prior knowledge. In this paper, we summarize the vision-based data anomaly detection methods which are inspired by the manual data inspection. Two steps are included in each of the series method: A) data visualization and conversion; b) neural network training for anomaly classification. This process imitates human biological vision and logical thinking. Real-world acceleration data of a long-span cable-stayed bridge are used for method validation, showing that the method provides a novel alternative perspective for SHM data preprocessing.