Anomaly detection of sensor faults and extreme events by anomaly locating strategies and convolutional autoencoders

Deep learning has demonstrated advantages in the detection and classification of sensor faults. However, challenges exist in the domain of extreme event detection, including the scarcity of training data due to the rarity of extreme events and the difficulty of distinguishing between sensor anomalies and extreme events. This study proposes a novel two-stage anomaly detection method that can separately detect extreme events and sensor faults from the normal pattern without using any data related to extreme events in the training step. In the first stage, the potential segments are located using three defined indexes from raw normal acceleration data and then represented as peak and root mean-square envelopes. Then, these envelopes are normalized and grouped by the three indexes to train three convolutional autoencoders. The preprocessing strategies minimize the difference of the normalized sequence shapes between the normal pattern and extreme response. Hence, well-trained convolutional autoencoders can detect sensor fault patterns with different sequence shapes from the normal pattern. In the second stage, several thresholds are defined to separate the patterns sensitive only to absolute values from the normal pattern, including extreme events and partial Minor scenarios. A multilabel classification is followed to identify specific sensor fault patterns. Datasets from two real structural health monitoring systems are utilized to validate the proposed method. Results show that extreme events and sensor faults can be detected efficiently and accurately. Three extreme events, Typhoon Saola, Typhoon Higos, and a pedestrian test, are successfully detected using the developed method.