Data-Driven Incremental Learning Model for Failure Prediction in Smart Water Taps in Sub-Saharan Africa

This thesis aims to predict failures in smart water taps to support proactive maintenance, which helps sustain water supply to rural communities in Sub-Saharan Africa and similar contexts. Solar-powered smart water taps (SWT) deployed to rural areas in some parts of Africa are perceived as low-cost and reliable water supply sources for domestic use in the Sub-Saharan region. These SWTs, often called e-Taps, dispense water when a pre-paid token comes in contact with them. However, these taps occasionally develop technical faults that lead to delayed service delivery by water service providers, causing a water supply shortage for the service users. Therefore, the sustainability of these metered water withdrawal taps becomes one of the challenges affecting the smooth running of solar-powered water taps in some African regions. This research focuses on timely failure prediction in e-Taps to support system maintenance by developing an early warning system based on the historical dataset acquired remotely from the actual usage of the taps. Water withdrawal data from Jarreng village was used as a case study. The dataset from the case study contained over a million events with some data inconsistencies and missing data values. From the dataset, the FlowRate, Voltage and DateTime variables formed the major data variables for the research because they hold the relevant information for the study. The dataset was preprocessed using the generative adversarial network (GAN) to remove impurities. The exponential smoothing method was used to correct overfitting in the dataset before the data was passed to the proposed model. A hybrid CNN-biLSTM model for failure prediction was developed using the convolutional neural network (CNN) and the bidirectional long short-term memory (biLSTM) networks. In addition, the developed model validation was explored using data from similar case studies in Africa. The validation exercise provided an accuracy of 88%. Overall, this research harnessed the CNN model to convert the time series data into a one-dimensional image to enable it automatically extract salient data features. The extracted features are further passed to the model biLSTM component, which learns data dependencies from the water withdrawal taps to predict failures in the e-taps. The proposed model can adapt to similar use cases with minimal modification. Also, this research focuses on using a hybrid combination of the convolutional neural network (CNN) and the bidirectional long short-term memory (biLSTM) networks to build an online (real-time) failure prediction model for the e-taps deployed in some parts of Africa. Furthermore, the research considers the e-tap's continuous nature and the discrepancies associated with the dataset. Furthermore, the proposed model developed in this research can also adapt to changes in the statistics of the continuous data stream while predicting failure.