Anomaly Detection in Nuclear Power Production Based on Neural Normal Stochastic Process
To ensure the safety of nuclear power production, nuclear power plants deploy numerous sensors to monitor various physical indicators during production, enabling the early detection of anomalies. Efficient anomaly detection relies on complete sensor data. However, compared to conventional energy sou...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-07-01
|
| Series: | Sensors |
| Subjects: | |
| Online Access: | https://www.mdpi.com/1424-8220/25/14/4358 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | To ensure the safety of nuclear power production, nuclear power plants deploy numerous sensors to monitor various physical indicators during production, enabling the early detection of anomalies. Efficient anomaly detection relies on complete sensor data. However, compared to conventional energy sources, the extreme physical environment of nuclear power plants is more likely to negatively impact the normal operation of sensors, compromising the integrity of the collected data. To address this issue, we propose an anomaly detection method for nuclear power data: Neural Normal Stochastic Process (NNSP). This method does not require imputing missing sensor data. Instead, it directly reads incomplete monitoring data through a sequentialization structure and encodes it as continuous latent representations in a neural network. This approach avoids additional “processing” of the raw data. Moreover, the continuity of these representations allows the decoder to specify supervisory signals at time points where data is missing or at future time points, thereby training the model to learn latent anomaly patterns in incomplete nuclear power monitoring data. Experimental results demonstrate that our model outperforms five mainstream baseline methods—ARMA, Isolation Forest, LSTM-AD, VAE, and NeutraL AD—in anomaly detection tasks on incomplete time series. On the Power Generation System (PGS) dataset with a 15% missing rate, our model achieves an F1 score of 83.72%, surpassing all baseline methods and maintaining strong performance across multiple industrial subsystems. |
|---|---|
| ISSN: | 1424-8220 |