Interpretable real-time TN soft sensor via multi-periodicity and dimension-aware temporal representation learning:...

Area of Science:

• Environmental Science
• Water Quality Monitoring
• Sensor Technology

Background:

• Continuous stream total nitrogen (TN) monitoring is crucial for water quality management but faces challenges like biofouling and data gaps.
• Existing in-situ monitoring methods are often unreliable for long-term deployment.

Purpose of the Study:

• To develop an interpretable, real-time, in-situ soft sensing framework (IRIS-TN) for robust TN estimation in dynamic river environments.
• To enhance data quality and improve the accuracy and reliability of TN monitoring.

Main Methods:

• A data-quality enhancement pipeline using statistical screening and CatBoost regression to reconstruct corrupted TN records.
• Training an edge-deployable soft sensor with a frequency-aware temporal modeling architecture for non-stationary river dynamics.
• Gradient-based attribution analysis to interpret TN estimate drivers and assess physical consistency.

Main Results:

• IRIS-TN demonstrated strong agreement with analyzer measurements (MAE: 0.191, RMSE: 0.259, R²: 0.809), outperforming baseline models.
• The framework maintained robust performance across different flow regimes with low inference latency (7.69 ms).
• Attribution analysis revealed dominant nitrate signals and flow-dependent contributions from hydraulic and physicochemical factors.

Conclusions:

• IRIS-TN offers an interpretable and real-time soft-sensing solution for reliable TN monitoring in urban river systems.
• The developed framework addresses key limitations of traditional in-situ monitoring, enabling better nutrient dynamics tracking.