Published in the IEEE International Conference on Communications, Networks, and Satellite (2024) proceedings.
The bistatic radar setup, with two ultra-wideband modules bisecting a glass Petri dish where liquids are held to be analyzed.
With a growing rise in urbanization, there is an inevitable need to create sustainable digital networks for management of natural resources in developing smart cities, due to the increased value of such resources with population growth and development. One such necessary resource is water, and management of water consumption is important to the welfare of society. Previous studies have focused on invasive methods of water quality monitoring, which can be inefficient, not continuous, or pose other hazards. Noninvasive (without contact) water quality monitoring is crucial for preserving sample integrity, enhancing safety, and enabling continuous, real-time assessments. This paper presents a method using a bistatic Ultra-Wideband (UWB) radar to noninvasively predict salinity and total dissolved solids (TDS), which are key water quality parameters, of a sample. By processing the UWB signal's time of flight (ToF) and received signal strength indicator (RSSI), and applying the Kalman filter for additional feature generation, then training a Random Forest Regressor model for prediction based on simulated UWB signal data, accurate calculations of water quality are achieved. With overall R-squared values of 0.906 for salinity and 0.907 for TDS, and low root mean square errors and mean absolute errors, this approach offers a reliable and efficient alternative to traditional invasive methods, with broad implications for environmental monitoring and public health.
Scatter plots showing the relationship between water quality parameters and signal metrics. Time-of-Flight increases with Salinity and Dissolved Solids, while Signal Strength decreases.
Signals are simulated from a Ridge Regression model that evaluates likely signal strength and time based on prexisting water quality conditions. A Kalman Filter is applied to the simulated signals to increase the quality of input data for water quality prediction. This graph shows a comparison between the raw and Kalman filtered signal attributes.
A comparison between model-predicted and measured values for salinity and dissolved solids across 25 data points. High accuracy of ~90% is achieved with the Random Forest Regressor used to predict the water parameters.