Integrating AI into the asset management of rural water supply schemes in Nepal
THE QUESTION
Can AI tools be used to improve the quality of data collected and stored within Nepal’s national management information system – NWASH - for asset management of WASH infrastructure and planning and delivery of WASH projects?
LOCATION: Nepal
SECTOR: WASH
TECH: AI
TIMELINE: September 2022 - Present
PIONEER: Kamala KC
PARTNER: Rara Labs & Prixa Consultancy
The Challenge
Currently, across Nepal, data on the nature and condition of WASH assets is collected and uploaded onto the government’s NWASH data portal. The data collected determines the budget allocated to WASH projects, informing the way WASH construction and maintenance projects are planned and delivered. Structured data collected and stored in NWASH is often inaccurate, due to inconsistent and inaccurate data collection and insufficient validation. Ultimately, if the data in the system is inaccurate, then the government cannot make good decisions on resource allocation which compromises the water supply infrastructure and therefore people’s access to water.
The Idea
The proposed solution is to use an AI image recognition algorithm to analyse the nearly 1-million photos in the NWASH database to identify assets accurately and to identify faults with those assets. The AI solution will validate, or invalidate, current NWASH data and improve the quality of data on NWASH, increasing the likelihood that faults with assets will be recorded on the system and eventually addressed on the ground. Partial data validation is currently being done manually by a small team at the Ministry of Water Supply, who are key stakeholders and will take ownership of the solution at the end of the project.
Over the course of 3 months in the pre-pilot phase, the team undertook a landscape review to identify and understand the way WASH data is managed and the key problems that exist, such as the impacts of poor-quality data and the needs of users for improving WASH data. At the same time, an AI algorithm minimum-viable-product was developed to identify assets and faults, such as broken pipes and corrosion, using WASH infrastructure images from NWASH. Throughout this phase, the team worked closely with partners in the Ministry of Water Supply (MoWS) and the Department of Water Supply and Sewerage Management.
Key metrics
Whilst significant improvements have been made, only 25% of the water supply is reported to be fully functioning and almost 40% requires major repairs.
Manual data validation only consists of 8 people.
For tap data alone there are almost a million assets and photographs collected in NWASH, meaning a sample approach must be used.
The AI was able to detect anomalies in 91% of instances, compared to 87% of a WASH expert, and 79-77% for non-WASH expert.
What we learned
Structured data collected and stored in NWASH is often inaccurate, due to inconsistent and inaccurate data collection and insufficient validation.
Inaccurate data directly impacts communities’ access to water supplies. Poor data negatively impacts the ability for stakeholders to deliver strategic policies and priorities and the ability to make appropriate budget allocations for projects. At a project level, inaccurate and incomplete data impacts effective decision making, project design and implementation.
It was technically possible to build an AI solution capable of processing the imagery in NWASH portal, to both identify assets, and identify those with faults (such as breakages, corrosion) - including faults not currently logged on the system.
Image recognition by AI can identify WASH assets and faults more accurately and more quickly than people can. When tested against existing manual validation processes, an AI tool was found to be more efficient and accurate at identifying anomalies (including corrosion and breakages to taps) in WASH asset photographs.
AI can be applied to identify a wider variety of WASH faults than first anticipated. The technical approach could be expanded to look beyond identifying faults relating to corrosion and breakages, identifying additional errors that we know are high priority for the Ministry, such as reservoir tanks being recorded as an incorrect size in NWASH.
AI appears to be the cheapest of all solutions to the WASH data validation problem. There are a range of technical ways of fixing these issues - but given current limitations (including financial limitations) AI does appear to be a favourable approach over options such as allocating improved technical data collection resources, implementation of smart sensors, or more dedicated data validation resources.
What happened next
Following the pre-pilot phase of work, the Hub has decided to continue supporting the team to build on the existing AI solution. Using the FT sprint methodology, we can test the feasibility and value-add of the different functional elements identified during the initial phase, and also ideate and identify additional possibilities for testing. By the end of the pilot, we will aim to identify an MVP solution that could be built and integrated into existing Ministry of Water Supply systems and processes and work with the government to take ownership.
See here for a report sharing initial findings from the pilot so far:
This pilot is ongoing and key learnings are forthcoming. Stay tuned!
Explore the NWASH Portal website here