A Soft Sensor Based Inference Engine for Water Quality Assessment and Prediction

Toxic contaminants with potential exposure to health risks now appear in urban drainage systems. These contaminants may be measured in specific concentrations or gross terms with impact on potential of hydrogen (pH) and dissolve oxygen (DO). Often, urban slum settlements are at the effluents of these drainage systems. These communities mostly lack functional water treatment facilities. They drink and sanitize from run-off, and shallow wells in alluvial aquifers. This is responsible for the poor health and low quality living conditions of urban slum dwellers. It becomes imperative to implement reliable, real-time, water quality assessment and prediction systems to resolve known and emerging water quality issues in order to improve the quality of life in such communities. Thus, this work developed an inference engine for robust water quality assessment and prediction using some statistical and intelligent discriminant functions. Results show that machine learning algorithms including the Logistic Regression, Decision Trees, Random Forest, XGBoost, and Neural Networks schemes reliably predicted water potability in the absence of two missing instrumentation parameters namely: pH and DO. Typically, the inference engine estimated pH in the absence of DO. Subsequently, the Dissolved Oxygen was estimated using the predicted pH. In outcomes, computational metrics of the pH prediction model returns Mean Squared Error (MSE) of 0.0163 for training and 0.1024 for validation. Similarly, the Dissolved Oxygen was estimated by the Random Forest model with a mean squared error (MSE) of 1.0335 for training and 0.7150 for validation. Decision engines built on statistical discriminant functions reliably predicted water potability with validation accuracy and F1 score of 95.19% and 96.94% for Logistic Regression, 100% and 100% for Decision Trees, 98.97% and 99.35% for Random Forest, and 100% and 100% for XGBoost. Also, the Neural Networks scheme predicted water potability with an MSE of 98.97% and 99.35% for training and validation respectively. The developed soft sensor solution is applicable in situations with complete or incomplete instrumentation data. Summarily, the developed inference engines provide a reliable framework for fault diagnosis and serve as backup sensors in case of failure. This enhanced the validation and reliability of sensor data for water quality assessment and prediction.