Sensor Image, Anomaly Detection Method for Hydroelectric Dam Structure Using Sensors Measurements and Deep Learning
Most modern hydroelectric dams are equipped with state-of-the-art sensors to measure and detect irregularities as part of a comprehensive safety-monitoring system. To better prevent future disasters, machine-learning algorithms have been employed. Often, these algorithms are trained on historical se...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Tamkang University Press
2025-01-01
|
| Series: | Journal of Applied Science and Engineering |
| Subjects: | |
| Online Access: | http://jase.tku.edu.tw/articles/jase-202509-28-09-0001 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Most modern hydroelectric dams are equipped with state-of-the-art sensors to measure and detect irregularities as part of a comprehensive safety-monitoring system. To better prevent future disasters, machine-learning algorithms have been employed. Often, these algorithms are trained on historical sensor data to predict future events. However, owing to dams’ high safety standards, it is uncommon for significant anomalies to occur. Hence, most of the collected sensor measurements were skewed to the normal range. This often results in the collection of extremely imbalanced datasets. Furthermore, the sensor measurements are highly correlated with each other along the spatial and temporal axes. Thus, the sensor measurement input vector for machine learning algorithms should be a multidimensional vector instead of a one-dimensional vector, as suggested by the traditional tabular form of data. In this study, a novel method for addressing these problems is proposed.
By converting the collected sensor data into images, it is possible to include both the temporal and spatial relationships of the data in the input vector. Moreover, an autoencoder was built to detect anomalies in an imbalanced dataset. To validate the proposed method, nearly seven years of sensor data collected from a critical section of a dam structure were used. The data are then grouped along the temporal axis and converted into an image that represents a single day of the sensor measurements. Using these images along with the proposed autoencoder structure, the method achieves an AUC score of 0.97 in detecting anomalies given a single day of sensor image. |
|---|---|
| ISSN: | 2708-9967 2708-9975 |