Water Quality Anomaly Detection Method Based on Attention-Gated Liquid Neural Network.

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Water monitoring networks are expanding, increasing the need for effective time-series anomaly detection.
• Conventional models face challenges with irregular data, complex correlations, and interpretability in water quality monitoring.

Purpose of the Study:

• To develop a robust and interpretable model for large-scale water quality anomaly detection.
• To address limitations of existing models in handling real-world water quality data.

Main Methods:

• Proposed an Attention Gated-Liquid Neural Network (AG-LNN) integrating Liquid Neural Networks (LNN) with attention mechanisms.
• Implemented input-attention and time-constant gates to focus on relevant variables and adapt temporal memory.
• Utilized data from China National Environmental Monitoring Center (2019-2024) across 13 provinces.

Main Results:

• AG-LNN outperformed LSTM, TCN, Transformer, and GNN models.
• Achieved a Precision-Recall Area Under Curve (PR-AUC) of 0.95 and an F1-score of 0.90.
• Demonstrated stability across cross-region and temporal evaluations; AG-LNN-light offered efficient edge deployment.

Conclusions:

• Attention-gated continuous-time modeling offers a powerful approach for water quality anomaly detection.
• The AG-LNN provides a robust, interpretable, and efficient solution for safeguarding aquatic environments.
• The model's adaptability and performance highlight its potential for practical environmental monitoring applications.