RESUMO
There is a growing need to sustain solar-powered water taps in most parts of the sub-Saharan Africa. The frequent failure of the water taps gives rise to intermittent water supply and poor service delivery by the water service providers. The challenge is to foresee and predict the failure of these water systems before they occur. This study develops a scalable machine-learning model for failure prediction in electronic water taps to ensure timely maintenance of the taps. Specifically, we develop a model based on long short-term memory (LSTM) to efficiently make failure predictions with noisy heterogeneous time-series data from rural water taps. Results from the experiment prove that the proposed model can effectively classify activities and patterns in various time-series datasets. With the proposed model, the failures of the solar-powered taps due to abnormal events can be successfully predicted well in advance, with an accuracy of 78.54%. Based on the data analyses, common causes of failures are presented.
