Data Augmentation for a Virtual-Sensor-Based Nitrogen and Phosphorus Monitoring

Area of Science:

• Environmental Science
• Water Quality Monitoring
• Data Science

Background:

• Eutrophication control requires accurate phosphorus and nitrogen loading data.
• High-frequency monitoring is costly and impractical.
• Virtual sensing offers a cost-effective alternative for predicting nutrient loads.

Purpose of the Study:

• To develop a cost-effective method for generating sufficient training data for virtual sensor models.
• To improve the accuracy of predicting phosphorus and nitrogen loading.
• To address the limitations of acquiring adequate training samples due to sensor costs.

Main Methods:

• Utilized variational autoencoders (a generative model) to create synthetic water quality data.
• Applied data augmentation techniques, including experimental settings and hyperparameter optimization.
• Validated the model using water quality data from River Thames tributaries.

Main Results:

• The novel data augmentation approach improved root mean squared errors by 23-63% compared to the state of the art.
• Significant improvements were observed when using up to three predictors.
• K-nearest neighbors and extremely randomized trees demonstrated superior predictive accuracy and computational efficiency.

Conclusions:

• Variational autoencoders are effective for generating synthetic data to augment training sets for virtual sensor models.
• The proposed data augmentation strategy enhances the accuracy of nutrient load predictions.
• This method offers a practical solution for cost-effective, high-frequency water quality monitoring.