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

Summary

This study proposes an Attention-Gated Liquid Neural Network for water quality anomaly detection that achieves superior accuracy and interpretability through continuous-time modeling and attention-based gating, thereby enhancing the reliability of environmental monitoring and supporting sustainable water management.

Abstract

With the rapid expansion of water monitoring networks, time-series anomaly detection has become increasingly crucial for safeguarding aquatic environments and ensuring their sustainable management. However, conventional models often struggle with irregular sampling intervals, multivariate correlations, and interpretability in practical applications. To address these challenges, this study proposes an Attention Gated-Liquid Neural Network (AG-LNN) that integrates the dynamic modeling capability of the Liquid Neural Network (LNN) with attention-based gating mechanisms. The model introduces an input-attention gate to emphasize anomaly-relevant variables such as Dissolved Oxygen (DO) and the Permanganate Index (COD_Mn), and a time-constant gate that adaptively adjusts the model's temporal memory. Using data from the China National Environmental Monitoring Center (CNEMC) collected between 2019 and 2024 across 13 provinces, AG-LNN demonstrated superior performance over baseline models, including Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), Transformer, and Graph Neural Network (GNN) architectures. It achieves a Precision-Recall Area Under Curve (PR-AUC) of 0.95 and an F1-score of 0.90, while maintaining stability under cross-region and temporal evaluations. A compact version, AG-LNN-light, reduces parameters by 62% with minimal accuracy loss, enabling efficient edge deployment. The results confirmed that attention-gated continuous-time modeling provides a robust and interpretable approach for large-scale water quality anomaly detection.

Introduction

Water quality monitoring is essential for protecting public health, sustaining aquatic ecosystems, and ensuring compliance with environmental regulations. Traditional water quality models, such as the Soil and Water Assessment Tool (SWAT), the mechanistic QUAL model, and statistical forecasting approaches, often struggle to capture the highly nonlinear, multivariate, and temporally irregular nature of water systems influenced simultaneously by chemical, biological, hydrological, and meteorological factors^1,2,3,4,....

Protocol

NOTE: The overall workflow of this study is shown in Figure 1.

1. Data acquisition

1. Data source identification
   1. Obtain water quality time-series data from the China National Environmental Monitoring Center (CNEMC) public database (https://www.cnemc.cn/sssj/). Retrieve continuous national-level monitoring records covering major lakes and reservoirs in China. Key physicochemical indicators included pH, dissolved oxygen (DO), total phosphorus (TP), total nitrogen (TN), permanganate index (COD_Mn), turbidity (NTU), and electrical conductivity (EC). 

Results

Overall performance comparison

Table 2 presents the comparative results of the different baseline models and the proposed AG-LNN for the water quality anomaly detection task. Among the traditional unsupervised baselines, Isolation Forest²² and One-Class SVM²³ achieved only moderate precision and recall, reflecting their limited ability to capture the temporal dependencies inherent in multi-source water quality da.......

Discussion

This study introduces an Attention-Gated Liquid Neural Network (AG-LNN), a novel architecture for water quality anomaly detection that combines continuous-time liquid dynamics with attention-based gating mechanisms. The liquid neural component of an AG-LNN is inspired by Liquid Time-constant Networks (LTCs), which model time series with learnable, input-dependent time constants²⁰. However, AG-LNN extends beyond standard LTCs by integrating two additional gates: an input attention gate, which empha.......

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
256 GB DDR4 RAM	Samsung	M393A4K40DB3	High-capacity memory for handling large multivariate time-series datasets
CUDA 11.6	NVIDIA	N/A	GPU acceleration toolkit for PyTorch
Intel Xeon Gold 6330 CPU (2.0GHz, 28 cores)	Intel	BX80708-6330	Used as the main computation server for model training
Matplotlib 3.4, Seaborn 0.11	Open Source	N/A	Visualization of experimental results
NumPy 1.21, Pandas 1.3	Open Source	N/A	Data preprocessing and feature engineering
NVIDIA A100 GPU (40GB)	NVIDIA	900-21001-0000-001	Accelerated training of AG-LNN with CUDA support
Python 3.9	Python Software Foundation	N/A	Main programming language for implementation
PyTorch 1.12	Meta AI	N/A	Deep learning framework used for building AG-LNN
Scikit-learn 0.24	Open Source	N/A	Evaluation metrics and baseline models
Ubuntu 20.04 LTS OS	Canonical	N/A	Operating system for the computational environment