| Author: | Zhang, Yuxuan |
| Title: | Detection of sensor faults and extreme events for structural health monitoring |
| Advisors: | Xia, Yong (CEE) Wang, Xiaoyou (CEE) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Subject: | Structural health monitoring System failures (Engineering) -- Prevention -- Data processing Anomaly detection (Computer security) Machine learning Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Civil and Environmental Engineering |
| Pages: | xix, 188 pages : color illustrations |
| Language: | English |
| Abstract: | Structural health monitoring (SHM) systems incorporate multi-type sensors to measure the environmental conditions, loads, and structural responses. These multi-type sensors generate a large amount of sensing data. On the one hand, sensor faults may cause the measurement data to have low fidelity and thus reduce the precision of condition assessment. Automatic detection of sensor faults ensures the normal operation of the SHM systems. On the other hand, extreme events, such as typhoons or earthquakes, may cause the monitoring data to look “abnormal”. These abnormal data are closely related to the structural safety condition and should require special attention. Therefore, automatic detection of sensor faults and extreme events plays a pivotal role in the normal operation of SHM systems. Existing approaches focus on the detection of sensor faults only but do not consider the extreme events. Moreover, they apply to a single type of monitoring data, generally acceleration. This PhD thesis first develops three machine learning-based anomaly detection methods to separate the normal pattern, sensor faults, and extreme events using acceleration data. The first method employs support vector data description (SVDD) to generate a decision boundary to enclose the normal training data in the feature space. Both sensor faults and extreme events are identified as sensor anomalies. Extreme events are separated from sensor faults by examining the correlation of several sensors. An extreme case is detected when several sensors are found to be abnormal. However, the method may fail if several sensor faults occur simultaneously. The second method is developed to separately detect extreme events and sensor faults from the normal pattern. The proposed locating strategy, normalization method, and training trick minimize the differences between the normal pattern and data under extreme events. Therefore, the one-dimensional (1D) convolutional autoencoder only trained with the normal data can reconstruct extreme responses but fails to reconstruct sensor fault patterns that have different sequence shapes from the normal pattern. In the third method, several parameters are defined to explore why the well-trained autoencoder can reconstruct extreme responses. The defined parameters based on the proposed normalization unify features of the normal pattern and data under extreme events; thus, sensor faults can be separated from the normal pattern and extreme events. The above methods are effective when sufficient labeled training data are available. For scenarios with limited labeled training data, few-shot classification methods are developed based on manual and deep-learning feature extraction. The few-shot classification is performed by measuring the similarity of feature vectors between the query and support sets. The manual feature extraction method is advantageous when explainable results are required, whereas the deep learning-based feature extraction method is preferred when there is an adequate amount of labeled data to establish base classes. Last, the anomaly detection methods are extended to other types of SHM sensing data, including wind speed and strain data, which are also categorized into normal patterns, sensor fault patterns, and extreme events. The developed anomaly detection methods are initially applied to identify anomalies in wind speed, temperature, and strain data. The results of extreme event detection and classification for acceleration data assist in separating sensor faults and extreme events in other data types. Specifically, anomalies detected in wind speed and strain data are more likely attributed to extreme events if such events are concurrently observed in the acceleration data. Datasets from real SHM systems of a footbridge in The Hong Kong Polytechnic University, a long-span bridge, and the Tsing Ma Bridge are utilized to validate the proposed method. Results show that the proposed anomaly detection framework can efficiently and accurately detect sensor anomalies. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13419