Ensemble machine learning and forecasting can achieve 99% uptime for rural handpumps

Area of Science:

• Water resource management
• Machine learning applications
• Public health engineering

Background:

• Broken water pumps hinder clean water access for the global poor.
• High water infrastructure uptime (>99%) is crucial for realizing health benefits and customer willingness to pay.
• Current maintenance strategies often fall short of optimal performance levels.

Purpose of the Study:

• To demonstrate the use of sensor data and supervised ensemble machine learning to significantly increase water pump fleet uptime.
• To forecast water pump failures and rapidly identify existing ones to improve operational efficiency.
• To estimate the cost savings associated with implementing a predictive maintenance algorithm for water pumps.

Main Methods:

• Utilized sensor data collected over 14 months from 42 Afridev handpumps in western Kenya.
• Applied supervised ensemble machine learning algorithms to analyze sensor data and predict pump failures.
• Compared the levelized cost of water over a 10-year lifespan for the proposed system versus a sensor-less scheduled maintenance program.

Main Results:

• Increased total fleet uptime from a baseline of approximately 70% to over 99%.
• Estimated a 7% reduction in the levelized cost of water compared to sensor-less maintenance.
• Successfully forecasted pump failures and identified existing failures with high accuracy.

Conclusions:

• Integrating sensors and machine learning offers a viable solution to dramatically improve water pump reliability.
• Predictive maintenance can lead to substantial cost savings and enhanced water service delivery.
• The proposed system has the potential to significantly improve public health outcomes and customer satisfaction in water-scarce regions.