Data Anomaly Detection for Bridge SHM Based on CNN Combined with Statistic Features

Structural health monitoring of long-span bridge has received increasing attention in recent years. In order to achieve accurate monitoring, the integrity of data collection should be guaranteed. Unfortunately, these data inevitably contain a variety of types of anomalies due to sensor faults, harsh environments, and other issues. Identifying anomalies from the data is essential to ensure credible monitoring results. Machine learning methods have the potential to detect data anomaly automatically. However, a well-performed convolutional neural network (CNN) model requires a large number of balanced training samples and frequent tuning. When similar anomalous patterns exist in complex systems, the tuning process might be time-consuming. In this paper, a data anomaly detection method is proposed based on CNN combined with statistic features. Firstly, acceleration data are downsampled, stacked, and input into CNN as the training set. A CNN model is designed and trained. Intermediate results are obtained through the model. Subsequently, the statistic features are applied to analyze and classify the confusable patterns. The novelty of this framework is that it combines the advantages of CNN and statistical features, which can realize data anomaly detection faster and more accurately than using CNN alone. The results of the acceleration data from a bridge demonstrate the effectiveness of the proposed approach to identify the anomalous data.