A predictive and scalable health-monitoring system for handpumps

The handpump remains the most reliable and low-cost method to access groundwater, making it a critical component of rural water supply for around 200 million people in sub-Saharan Africa to meet their daily water needs. Despite this, an estimated one in four handpumps does not work at any given time. This is primarily due to a lack of low-cost and reliable remote monitoring methods for predicting the condition of the rural infrastructure. Existing handpump maintenance models rely on manual reporting from the water users or use flow meter data without further analysis or insights into unique usage patterns. The investigations in this thesis describe the development and evaluation of a dynamic distribute inference (DDI) system, a novel method whereby a novelty score is produced in-situ and “intelligent” data summaries are transmitted to a cloud-based station for further assessment to improve overall fidelity. This work involved developing, manufacturing, and installing the bespoke sensors for collecting the high-rate vibration data from handpumps to create three unique open-access data sets that were used to conduct this work. An in-depth literature review combined with extensive first-hand field experience and expert interviews informed the proposed standardised functionality categories. Five lightweight classifiers are considered to operate within the limited-resources available onboard the embedded sensor: logistic regression (LR), naïve Bayes classifier (NBC), decision trees (DT), k-nearest neighbours (kNN) and neural networks (NN). All of these were deemed good predictors of handpump failure; however, the probabilistic framework of the LR allows dynamic adjustment of the novelty threshold. As such, LR was used to produce in-situ novelty scores and implemented in the field to collect novelty filter scores produced on the embedded system along with raw waveforms from 20 different handpumps during multiple field visits as well as ten handpumps over 18 months. Based on the novelty score, the cloud may request additional data from the embedded sensor (such as features or raw waveforms over novelty scores) to perform a further assessment in order increase the fidelity of the health-monitoring system. The data summaries from the sensors were retrospectively assessed on the cloud using two supervised machine learning methods: support vector classifier (SVC) and random forest (RF) and compared to conventional machine learning methods: multilayer perceptron (MLP) and 1-D convolutional neural network (CNN). The novelty filter threshold is adjusted based on the desired false positive rate (FPR). Transferability of the DDI platform was shown by evaluating the predictive performance on both vibration data from handpumps as well as more complex physiological patient data. In both cases, the implementation of the DDI improved classifier accuracy by more than 10%, while significantly reducing the cost of data transmission. These results represent a significant advancement in the possibility of low-cost sensors combined machine learning to deliver scalable and reliable quantitative insights that can inform priority setting for improved rural services toward achieving the sustainable development goals (SDGs).