Data anomaly detection in structural health monitoring using modified transformer encoders with 1D-CNN layers

Structural health monitoring (SHM) has been widely used in civil infrastructure in recent decades. In SHM, vast amounts of data are collected using diverse sensors to monitor the health of civil structures. During this process, various types of anomalies may occur, which hindering an accurate assessment of the structure's condition. Anomalies mainly occur due to the influence of the harsh environment, sensor faults, or actual damage to the monitored structure. Therefore, early detection of anomalies is essential for monitoring the condition of structures. Conventional anomaly detection algorithms used in SHM systems, such as statistical thresholding, distance-based, rule-based, and clustering methods, have become ineffective today with growing data flow. These traditional algorithms face several limitations, including scalability issues, lack of adaptability to changing conditions, sensitivity to noise, and extensive feature engineering requirements. To address these issues, this paper proposes a modified transformer-based multiclass anomaly detection method for SHM systems. In our approach, we replace the feed-forward layers in the transformer encoder with two 1D-CNN layers and opt not to use positional encoding, as the occurrence of anomalies in SHM systems is not strongly related to specific positions within the sequence. Initially, the statistic and frequency domain features are extracted from the labeled time-series raw data. Then the modified transformer-based anomaly detection model is trained with extracted features and validated with acceleration data measured from a long-span cable-stayed bridge. The results confirm that the modified transformer encoders with 1D-CNN layers, without positional encoding, provide improved performance in detecting and classifying multiple types of anomalies with high accuracy. This demonstrates the potential of our method for enhancing the effectiveness of SHM systems.