Transfer learning-based data anomaly detection for structural health monitoring

Abstract

The structural health monitoring (SHM) data of civil infrastructure are inevitably contaminated due to sensor faults, environmental noise interference, and data transmission failures. Anomalous data severely disturb the subsequent structural modal identification, damage identification, and condition assessment. Therefore, it is critical to detect and clean SHM data before data analysis. This paper proposes a novel approach for data anomaly detection based on transfer learning, that makes full use of the similarity of the anomalous patterns across different bridges and shares the knowledge incorporated in a deep neural network to achieve high-accuracy data anomaly identification for bridge groups. In the proposed approach, first, a multivariate database for a source bridge is built by plotting and labeling the raw sequential data. Then, a convolutional neural network (CNN) for data anomaly classification is designed and trained with the database in different conditions. The original CNN with the highest accuracy is transferred to a new bridge with enhancement training using a small part of the target bridge data. To validate the performance of the proposed method, the multivariate SHM data for two real long-span bridges are employed, including the acceleration, strain, displacement, humidity, and temperature data. The results demonstrate that transfer learning leads to a better classification capacity for the case of scarce labeled training data compared with the original network.