Hierarchical outlier detection approach for online distributed structural identification

Abstract

In this paper, a hierarchical outlier detection approach is proposed for online distributed structural identification. In contrast to centralized identification, distributed identification extracts important features from the raw response data at the sensor nodes and transmits only them to the base station. Therefore, outlier detection is substantially more complicated than the traditional approach. In the proposed method, the local outliers in the raw data are detected directly at the corresponding sensor node, and they are excluded from further processing. However, if a sensor node is biased or exhibits other patterned outliers, these outliers will be undetectable at the sensor node level. It is necessary to conduct another level of outlier detection at the base station, namely, global outlier detection, before fusion. These two levels of outlier detection are of different nature. Local outlier detection concerns directly with the raw response data, whereas the targets of global outlier detection are the local estimation results of the stiffness parameters. Therefore, they require different mathematical tools. The proposed hierarchical outlier detection approach detects the local outliers according to the outlier probability of the data points at the sensor nodes, whereas it detects the global outliers according to the outlier probability of the local estimation results. By excluding both types of outliers, reliable online distributed structural identification can be achieved. Two examples are presented to demonstrate the proposed method.

CRND: An Unsupervised Learning Method to Detect Network Anomaly

“…At the same time, it does not rely on annotated datasets, which contributes to the promotion and production application of network anomaly detection model greatly. At the same time, CRND can be widely used as a feature extractor, which may be better used to solve the target problem when combined with the related research of outlier detection [38,39].…”

Switching Kalman filter for damage estimation in the presence of sensor faults

“…An efficient, stand-alone and autonomous SHM algorithm should therefore be designed for localizing and isolating a faulty sensor at its time of occurrence and still continues with the damage estimation, undeterred by the fault. Faulty sensor data may also arise due to the presence of outliers in the measured data, for which outlier probability for detection can be utilized [21,22]. However present study focuses on faulty measurement data obtained following before-mentioned reasons only.…”

Sensor fault detection, localization, and reconstruction for online structural identification

“…As a result, the algorithm shifted to level II for finer localization of the faulty sensors in the 2nd group. The possible combinations of the sensors in the 2nd sensor group were denoted as (4), ( 5), ( 6), (4,5), (4,6), (5,6), and (4,5,6), and they corresponded to the model classes M 2, j ð Þ , j ¼ 1, 2,…, 7 in level II, respectively. During the time period from t ¼ 51:015 s to t ¼ 52:29 s, it is seen that the model class M 2,6 ð Þ had the highest plausibility (Figure 7c), so the faulty sensors in the 2nd group could be located at the 5th and 6th sensors.…”

“…It involves the observation and analysis of a structural system over time using sampled response measurements, so that changes of the material and geometric properties for engineering structures can be monitored. SHM contains various research topics, such as optimal sensor placement, 1,2 data cleansing, 3,4 structural model updating, [5][6][7][8] structural damage evaluation, [9][10][11] input identification, [12][13][14] and uncertainty quantification. 15,16 Among various SHM techniques, online structural identification is regarded as one of the most desirable approaches to assess the health status of the structure, since it can track the structural states promptly.…”

“…Well-known methods include control chart techniques, 26,27 multivariate statistical analysis, 28,29 and model residual-based methods. 4,30 Besides, among various faulty sensor detection methods, statistical hypothesis testing is the most popular framework for dense sensor networks. 24,28,31 The null hypothesis considers the case without faulty sensor, while the alternative hypothesis considers the case with faulty sensors.…”