Few-shot classification for sensor anomalies with limited samples

Abstract

Structural health monitoring (SHM) systems generate a large amount of sensing data. Data anomalies may occur due to sensor faults and extreme events. Sensor faults can result in low-fidelity measurement data, while data associated with extreme events are crucial for assessing the structural safety condition and should be given special attention. Accurate detection and classification of anomalies can improve the performance of SHM systems. However, most existing classification methods work well only when the number of a-single-class anomalies is sufficient. This study proposes an automatic few-shot classification method for sensor anomalies with limited labeled samples. The most discriminatory shapelet, a new representation of abnormal data, is learned from the standard normal class by maximizing the overall distance, which can locate the prominent abnormal features from 1-h acceleration data. The classification is then learned based on manual feature extraction and deep-learning-based feature extraction by measuring the similarity between the most discriminatory shapelets from the query and support sets. The proposed few-shot classification method is applied to datasets collected from two SHM systems of a long-span bridge and a campus footbridge. Results demonstrate that the proposed method can classify new anomalies with limited samples that differ from the defined anomalies.