Vision Transformer–Based Anomaly Detection Method for Offshore Platform Monitoring Data
Quanhua Zhu, Qinjiao Wu, Yalin Yue, Yuequan Bao, Tao Zhang, Xueliang Wang, Zhentao Jiang, Haozheng Chen
Abstract
The structural health monitoring system for offshore platforms exhibits anomalies in the collected monitoring data due to its prolonged service in complex and harsh environments. These anomalies significantly impede data analysis and early warning capabilities. In order to realize efficient and intelligent anomaly detection for the monitoring data, a method based on the vision transformer (ViT) model is proposed. Firstly, the monitoring data are transformed into image files by segmentation and visualization. Subsequently, the image features are analyzed to identify the anomaly patterns and construct an image database, so that the data anomaly detection problem is transformed into a classification problem based on the image features. Lastly, the ViT model combined with convolutional neural network (CNN) is constructed. The local perception ability of CNN is utilized to extract the underlying image features and smooth the image features inputted into the ViT model, which improves the accuracy of the model. Validation using actual monitoring data shows that the proposed method can efficiently detect multiple types of anomaly patterns in the monitoring data with an accuracy rate of 93.1%.