Robustness of IoT-connected e-Taps for sustainable service delivery of rural water supply

‘e-Taps’ monitor flow at rural water points in sub-Saharan Africa and enhance revenue collection using pre-paid tags. Real-time, high temporal resolution e-Tap usage data are available to service providers. In this paper, the robustness of the e-Tap is evaluated in the laboratory regarding (1) accuracy of the flow meter and (2) the flow rate reduction caused by addition of a y-strainer and debris build-up. An average relative error of þ3.63% across varying flow rates is found. A general calibration will bring 95.45% of measurements within a ±4.54% error range. In the y-strainer, smaller gauze sizes, smaller debris sizes, and higher debris loads cause greater flow rate reductions. The maximum reduction observed was only approximately 68% of the baseline flow rate. These physical findings can be integrated into software solutions to management problems.


INTRODUCTION
Operating on the principle of Internet of Things (IoT) connectivity, eWATERpay taps (herein called e-Taps) are currently being deployed on communal standpipes in The Gambia, Tanzania, and Ghana. Pre-paid credit tag operation leads to the 100% revenue collection required for sustainable service delivery (Harvey & Reed ). Their usage data are reported in real-time. This innovation has a great potential for improving reliability and sustainability of water supply across rural sub-Saharan Africa (Eliamringi & Kazumba ; Ingram & Memon ). Greater understanding of e-Tap robustness will aid this deployment.
Such innovations for rural water supply in developing countries are of growing interest in the research and practitioner communities (Andres et al. ). Robustness of these technologies has not been well researched, and existing evaluations focus on impacts on water supply and management (e.g. Hope et al. ; Nagel et al. ). Remedial technical alterations have been concomitant with product development, and unreported.
Here, this study instead investigates potential limitations to robustness across longer time-scales of operation, which is a new contribution and especially important considering the expected increase in such IoT technologies. This research contributes new empirical understandings of flow measurement accuracy, and of flow rate reductions from debris build-up inside the e-Tap, which have not been well studied in such technologies. These novel contributions demonstrate that the added capacities from such IoT innovations not only allow for improved management, but also a new ability to remotely refine data collection measurements and create predictive maintenance alerts, both detailed here, and provide software solutions to hardware problems. Water delivery points tend to be imprecisely categorised as 'working' or 'broken' (Carter & Ross ).
These findings now further show how timely usage data can provide more detailed narratives of functionality.
Robustness is of fundamental importance for long-term effectiveness (Klug et al. ; ; Kelly et al. ).
Hardware that is more resilient to breakdowns shall: (1) reduce interruptions to water supply; (2) reduce operation and maintenance (O&M) costs; and (3)  Second, the study evaluates the impact on flow rate of the addition of the y-strainer, and of debris build-up inside of the y-strainer This evaluation considers flow rate reduction from (1) addition of y-strainer and interior gauzes and (2) built-up debris inside the y-strainer that could restrict water flow.
Accumulation of debris in drinking water distribution systems is observed globally (Neilands et al. ) and is especially relevant in rural sub-Saharan Africa where infiltration of grit, sand, organic matter and plastic waste can be high. Reduced flow rates could lead to longer water collection times, queuing, and community dissatisfaction.
Results reveal flow rate reductions from different debris and gauze variables, and novel flow rate threshold alerts for predictive y-strainer maintenance are proposed.  This procedure was repeated 180 times over a varying Q real of 4.3-38.6 l min À1 . Below 4.3 l min À1 a 'low flow' error is activated; 38.6 l min À1 is the hydrobench pump limit. This range suitably replicates the observed range in The Gambia (4-32 l min À1 ) and Tanzania (20-60 l min À1 ).
Three different flow meters of the same model were used to discount the effects of testing one faulty unit.
The volume and the flow rate measured by the e-Tap (V e-Tap and Q e-Tap ) were manually calculated to two decimal places using Flow Count and reported seconds (t e-Tap ), avoiding rounding errors (normalised mean of 1.54%).
The percentage differences (%Δ) between 'real' experimentally measured volumes and volumes measured by the e-Tap are calculated using Equation (1) (Criminisi et al.

):
The international standard ISO 4064 permissible error range for water meters is ±5% at flow rates lower than the transitional flow rate (Q 2 , specific for each meter), and ±2% at flow rates above Q 2 (ISO ; Walter et al. ).
Once operating in the field these are suggested to increase to ±8% and ±3.5% respectively (van Zyl ). Here, an error range of ±5% across flow rates is taken as a permissible range.
Flow rate reduction from y-strainer and debris Decrease in Q baseline was measured across varying gauze sizes, debris sizes and debris loading of the y-strainer. Individual components of Module 2 were set up as above, with flow meters before the three-way ball-valve, and after the Module 2 components (with digital reader, calibrated).
Percentage flow rate reductions from a maintained Q baseline between set-ups without the y-strainer (and associated elbow) and with the y-strainer were measured across a Q baseline range of 3.0-17.6 l min À1 . This was then repeated, keeping the y-strainer in place, with seven gauzes of varying pore size (0.05-3.24 (and 63) mm 2 ; see Supplementary Material), across a comparable Q baselines range.

RESULTS
Accuracy of flow meter V e-Tap is larger than V real by a mean of 3.63% (%ΔV) across all measurements. The standard deviation across these %ΔV is 2.26%. There is significant variation between the accuracies of each flow meter tested (shown in Figure 2(a)).
This variation was verified by an additional ten repeats across the Q baseline range on flow meters 1 and 2.
Flow meter 1 is most inaccurate, with a higher mean %ΔV and standard deviation values, and also an evident negative correlation of %ΔV with increasing Q real . Flow meter 3 is the least inaccurate. In general, volume inaccuracies seem not to be sensitive to variation in flow rate within this range. The inaccuracy revealed here can be addressed by a general recalibration of all e-Tap data by À3.63%, as shown in Figure 2(b) (see Discussion).
The percentage error between 'real' and measured flow rates (%ΔQ) was also calculated as above for each measurement, and is shown in Figure 3(a); %ΔQ against Q real shares the characteristics of %ΔV against Q real for the flow meters,  and the horizontal y-strainer interior also begins to fill.
After this point, Q baseline decrease is more significant for gauzes with smaller pore sizes. Smaller pore sizes in the range of ∼0.05-0.36 mm 2 (Gauzes 6 and 7) cause an average ∼15% decrease with 22 g of debris.
Smaller sand sizes cause greater reductions in flow rate (until the sand is small enough to pass through pores). time. Volume is also a more relevant metric of water collection than flow rate. An average of 3.63% too much water is being recorded by the e-Taps, shown in both %ΔV and %ΔQ.
This marginal inaccuracy may be because of incorrect calibration in the measurement of volume from Flow Count.
This average is within the ±5% acceptable inaccuracy, Therefore, it appears that very slightly excessive credit is currently being charged from users per use, and slightly excessive water collection is being reported, probably resulting from an incorrect flow calibration value.
High-precision flow meters are expensive and impractical. Instead, a general calibration of all reported data using the mean %ΔV reported of 3.63% will bring the mean %ΔV to 0%, as shown in Figure 2(b). Then 95.45% (two standard deviations) of the above measurements would fall within ±4.54% inaccuracy, within the acceptable ±5%.
This can be practically achieved with an adjustment of the flow calibration value from 330 to 318.
To this end, eWATERpay developed a calibration application (for smartphone or web browser) for use at individual e-Taps in the field (or remotely). The application remotely commands the e-Tap to dispense a given volume of water, which is measured and re-entered to recalibrate the e-Tap. While the precision methodology employed in this paper is not possible in remote locations, repeats of the same principle will provide an acceptable accuracy range (ensuring flow rate is not unusually low or high for each water point). This protocol allows for community engagement and transparency.
In general terms, this demonstrates that software solutions can be used to overcome physical limitations to the precision of the e-Tap components. When compared with expensive manual calibration in laboratories, this finding demonstrates significant novel potential of such IoT-enabled technologies.

Flow rate reduction from y-strainer and debris
Any decreased flow rate from y-strainer addition is negligible (maximum 4.8% with the smallest gauze size). Build-up of the 13 g of debris in the y-strainer required before flow rate starts to reduce is likely to take some time in operating e-Taps, depending on debris levels in the water distribution system. Risk of 'non-functionality' because of the y-strainer is low; even with 22 g of debris and the smallest gauze size, flow rate only reduces by one-third. The benefit to robustness of the e-Tap by limiting flow meter blockages is therefore highly preferable.
If characteristics of the debris are previously known for specific water distribution systems, these results can inform the choice of gauze size using the findings reported in An effective method for e-Taps is to create alerts to unacceptable flow reduction thresholds, from a baseline derived from a longitudinal average of recent flow rates.
This can be tailored to specific e-Taps and gauze sizes.
Once a flow rate threshold is consistently passed over a long enough period of time to discount underlying flow rate fluctuations, an alert can be sent to the service provider. Q baseline reduction alert thresholds are proposed in Table 1, based on the debris build-up measurements reported in Figure 4(c)-4(h) after the 13 g y-strainer filling threshold has been crossed. As seen, smaller gauze pore sizes would necessitate higher Q baseline reduction alert thresholds. Thresholds should also decrease as gauze pore sizes get larger, however the error range of the flow meter revealed above would mean that thresholds lower than 5% would give higher likelihoods of 'false alerts'.
This technique would utilise the newly available highquality data for accurate predictions, and would use empirical measurements that reflect real flow rates, rather than relying on probabilistic machine learning. Higher resolution and accuracy flow rate data from e-Taps now allow predictive maintenance to move beyond assessment of either 'working' or 'broken' water points (Carter & Ross ).
Application of real-time flow rate data for predictive maintenance would be a tangible 'use' of data, so far overshadowed by 'collection', 'transfer' and 'analysis' with this kind of monitoring data (Ingram & Memon ).
It is important to note that benefits from predictive maintenance would be lost without a responsive maintenance team, financial sustainability, a robust water distribution system, and other requirements of resilient and

CONCLUSIONS
In this paper, an evaluation of e-Tap operation has been outlined focusing on the accuracy of the flow meter reading and on flow rate reduction caused by y-strainer addition.
Findings suggest that the e-Tap measurement of flow is marginally inaccurate with an average relative error of þ3.63%.
Varying baseline flow rate does not significantly impact this.
This means that users have been collecting a slightly lower volume of water for the credit that they are paying for, however this is insignificant compared with broader benefits to rural water supply sustainability and access. A general calibration across e-Taps by À3.63% will fine-tune accuracy.
The benefits of blocking debris reaching the e-Tap flow meter with the addition of the y-strainer outweigh any minor flow rate reductions observed. These reductions are negligible for all gauze pore sizes measured until roughly 13 g of debris builds up inside the y-strainer gauze. The results here can inform decisions of approximate gauze size to install, refined when the nature of the debris is known.
These findings have direct application in planning and management. A calibration software application for staff use in the field has been developed. The flow rate reductions relative to debris build-up for specific gauze sizes can lead to accurate predictive maintenance using alerts based on thresholds proposed here. This benefits from high-accuracy flow meters on e-Taps (compared with proxy measurements used elsewhere). Both of these use software solutions for hardware problems. This ability for service providers to improve service delivery remotely has significant and original benefits. The study has demonstrated the enhanced capacity available from a combination of high-resolution sensing data and remote analytics, and shows that potential benefits of such IoT innovations go beyond those currently established, and can accelerate progress towards the Sustainable Development Goals.

SUPPLEMENTARY MATERIAL
The Supplementary Material for this paper is available online at https://dx.doi.org/10.2166/ws.2020.128.