Field Testing of a Mixed Potential IoT Sensor Platform for Methane Quantification

Emissions of CH4 from natural gas infrastructure must urgently be addressed to mitigate its effect on global climate. With hundreds of thousands of miles of pipeline in the US used to transport natural gas, current methods of surveying for leaks are inadequate. Mixed potential sensors are a low cost, field deployable technology for remote and continuous monitoring of natural gas infrastructure. We demonstrate for the first time a field trial of a mixed potential sensor device coupled with machine learning and internet-of-things platform at Colorado State University’s Methane Emissions Technology Evaluation Center (METEC). Emissions were detected from a simulated buried underground pipeline source. Sensor data was acquired and transmitted from the field test site to a remote cloud server. Quantification of concentration as a function of vertical distance is consistent with previously reported transport modelling efforts and experimental surveys of methane emissions by more sophisticated CH4 analyzers.

Methane is a greenhouse gas (GHG) with a global warming potential of 72x that of CO 2 on the 20-year timescale and 25x of CO 2 on a 100-year timescale. 1The 1990-2020 Inventory of GHG emissions, estimates that the total CH 4 emissions associated with natural gas systems in 2020 is 6,596 kilotons of CH 4 or 164.9 millions of metric tons of CO 2 equivalent.Transmission systems account for 25% of these CH 4 emissions.CH 4 emissions from the entirety of the natural gas industryaccounts for 11% of total US GHG emissions. 2atural gas production in the US has increased from 20 trillion cubic feet in 2000 to 35 trillion cubic feet and is projected by the US Energy Information Administration to continue increasing over the next 30 years. 3 Transmission of natural gas within the US takes place over an estimated 300,000 miles of pipeline. 4Surveying these pipelines for leaks is technically and economically challenging.In the context of buried pipeline monitoring, the condition of the soil also affects the permeability of methane to the surface, 5 so infrequent monitoring may not pick leaks that are dependent on seasonal weather conditions.Evaluating the conditions at the scale of the natural gas transmission infrastructure will require low cost, field deployable, continuous monitoring technologies.
Current methods for monitoring of natural gas emissions include optical gas imaging, aerial or satellite surveys, and point sensors. 6ptical gas imaging (OGI) uses infrared cameras to image the emission of plumes from leaking methane sources, with 80% of emissions detectable from 10 meters away. 7OGI requires strong thermal contrast to detect less obvious leaks, 8,9 and a survey by Lyman et al. found that aerial surveys by OGI in winter had 2-6x worse limit of detection when surveying the oil and gas infrastructure in Utah's Uinta Basin. 10 Zimmerle et al. also found that accurately identifying leaks with handheld OGI tools was strongly dependent on the surveyor's experience and that a 90% probability of detection required on average a 3.29 standard liters per minute (SLPM) emission rate. 112][13] Satellite surveys of methane emissions can cover a wide area, but at high cost and spatial resolution of 50 km × 50 km as in the GOSAT satellite. 14These surveys can broadly identify which facility produces an elevated methane emission but cannot identify which specific components need to be repaired.3][14] Lavoie et al. found that aerial surveys of methane emissions over the Eagle Ford facility produced day-to-day variation as high as 50% deviation from mean in emissions rate and was also dependent on the type of survey aircraft and the sensing method used. 15These results illustrate the need for point sensors that their cost are low enough to deploy at high spatial resolution (10 s of meters) and can continuously measure methane emissions with the ability to identify the source.For buried pipelines, the sensors should have a limit of detection of 1 ppm for atmospheric detection, but closer to 100-1000 ppm within a meter above the surface. 16hree types of low-cost point sensors for methane detection are catalytic sensors, metal oxide semiconductor sensors, and mixed potential electrochemical sensors.8][19][20] Catalytic sensors are already deployed in the industry as safety sensors to indicate a release of flammable gas, but they are limited by their relatively high limit of detection and non-specificity to methane.Metal oxide semiconductor sensors contain a metal oxide such as ZnO or SnO 2 that decreases in resistance upon exposure to CH 4 . 21,225][26][27][28] Mixed potential electrochemical sensors (MPES) consist of two or more dissimilar electrodes embedded in a solid-state electrolyte. 29,30The electrochemical kinetics of the electrodes generates different mixed potentials, and this difference is measured as a sensing parameter.2][33][34][35][36] These previous studies demonstrated that Indium Tin Oxide (ITO) measured vs Pt was an effective sensor for CH 4 , La 0.87 Sr 0.13 CrO 3 vs Pt was an effective sensor for heavier hydrocarbons such as ethane and propane as well as NO x , and Au vs Pt was an effective sensor for species including CO and NH 3 .NH 3 in particular is expected to be used as a fingerprint for agriculturally generated methane.Efforts by our group and others have also demonstrated high stability, making these sensors suitable for long term deployments in continuous monitoring. 29,31,37E-mail: lktsui@unm.eduECS Sensors Plus, 2024 3 011402 The bulk of the literature on mixed potential sensors has been in laboratory conditions where the sensors were exposed to well controlled concentrations of test gases.Real world tests for MPES devices have been reported by Brosha et al. on H 2 sensors at filling stations in California, 34 and by Kreller et al. on emissions monitoring in engine dynamometers. 38However, no field tests have ever been reported for the MPES technology for natural gas emissions detection.In this work, we report for the first time on field testing of MPES devices for natural gas emissions detection in the field using a combined mixed potential sensor array system paired with an internet of things (IoT) platform.For natural gas monitoring the impact of local weather, soil conditions, and network connectivity in the field are aspects that cannot be evaluated in the laboratory.The field test was conducted at Colorado State University Methane Emissions Technology Evaluations Center (CSU METEC).METEC is a user facility with simulated natural gas infrastructure including well pads, storage tanks, and pipelines. 11METEC's main features are controlled emission rates and a suite of atmospheric sensors which are not available at real infrastructure.0][41] The METEC is typically used by industrial partners to accelerate the market-readiness of the technology or product being evaluated.
We have demonstrated that we can successfully detect natural gas emissions from an underground pipeline at concentrations of 5-5000 ppm using an MPES sensor array at a horizontal distance of 3 m and a vertical distance of 0.3 m.Quantification of CH 4 was performed using an artificial neural network machine learning model based on laboratory training data and was generally in agreement with literature reports of concentration values taken on more sophisticated methane analyzer technology.These results show the efficacy of a combined MPES, IoT, and machine learning framework as a low cost, field deployable natural gas emissions detection system.

Materials and Methods
Sensor fabrication.-Mixedpotential sensors were manufactured by syringe extrusion on a Hyrel System 30 M printer.The substrate was prepared by mixing 1.32 wt% dispersant (12.5 wt% polyvinyl alcohol in water), 2.36 wt% binder (16.6 wt% ascorbic acid in water), 0.37 wt% polyethylene glycol (MW380-420), 72.5 wt% magnesia stabilized zirconia (Goodfellow, MgO 2.8 wt% + ZrO 2 ), and balance water.The liquids were first mixed together in a Thinky ARV-310 centrifugal planetary mixer for 60 s at 2000 rpm.The solids were then added in 1/3 batches and mixed at 2000 rpm for 60 s.The 20 mm diameter x 2 mm thick substrate was then printed with a 5 mm s −1 table speed, layer height of 0.25 mm and platen temperature of 55 °C on an oleic acid coated silicone mat.The substrate undergoes a partial sintering at 1175 °C, followed by the addition of Pt ink (Ferro 5570) by syringe extrusion to form a Pt electrode and three conducting leads for sensing electrodes in a cross shaped configuration.The Pt and substrate are then co-sintered at 1450 °C for 8 h.ITO (90 wt% In 2 O 3 , 10 wt% SnO 2 ) and LSC (La 0.87 Sr 0.13 CrO 3 ) inks consist of 65 wt% solids in a 4:1 mixture of ESL473:ESL401 vehicle and thinner.The ITO and LSC inks were applied by hand to the ends of the Pt conducting leads and sintered at 1200 °C for 30 min.Next, Au ink (Nexceris, AU-I) was applied to the remaining conducting lead and annealed at 800 °C for 1 h.Finally, a porous layer of YSZ (Tosoh, 3 mol% Y 2 O 3 , TZ-3YS) was applied by hand bridging the four electrodes as a solid electrolyte layer and sintered at 1000 °C for 30 min.Upon completion of the sensor, contact to the electrodes were formed by using Ag paste (SPI Supplies, OK-SPI) with 99.9% silver wire.The sensor was then bonded to a 10 Ω ceramic disk heater custom built by Induceramic with a built-in k-type thermocouple.Details for the production and characterization of MSZ substrate mixed potential sensors are detailed in Halley et al. 42 The sensor was calibrated with simulated natural gas in 21% O 2 and balance N 2 .The simulant mixture had a CH 4 :C 2 H 6 ratio of 0.04 and the CH 4 concentration was varied between 40-5000 ppm.Sensor test calibration was performed after field testing at CSU METEC.A fully connected artificial neural network with 1 hidden layer using a hyperbolic tangent activation function was trained on this data for signal deconvolution following our previous work. 33The hidden layer size was varied from 1-12 neurons.A schematic of ANN with 5 hidden layer nurons is shown in Fig. 1.This ANN was constructed using the Tensorflow 2.1.0package for Python 3.7.7 on a custom-built PC with an AMD Ryzen 7 3700X 8-core processor with an NVIDIA GeForce RTX 2080 Ti GPU on the Ubuntu 22.04.2LTS operating system.The results of the ANN were also compared with a simple linear regression model implemented in Scikit-Learn 1.1.2.The performance of the ANN was also evaluated on a Raspberry Pi 4 Model B, Rev 1.5 with Tensorflow 2.12.0 and Python 3.9.2 which is representative of portable IoT processing hardware.
Test platform.-Thesensor package consists of a mixed potential sensor connected to a custom designed SensorComm Technologies Inc. (SCT) IoT circuit board which controls both sensor signal readout, temperature sensing, and heating of the sensor.The sensor was heated to 532 °C at an applied voltage of 10 V. Due to restrictions at METEC for operations of non-intrinsically safe electronics, a 30 ft. sample line of ¼″ polyurethane tubing was run from the sensor package to a tripod that could be moved around the test pad.An air pump (Parker Hannifin) and flowmeters were used fix the flow rate of sampling at 0.5 SLPM.A tank of zero air was connected to a flowmeter to the sample inlet.Prior to testing and after testing, air was delivered at 0.5 SLPM for 30 min to establish a baseline signal measurement.
Field testing.-Fieldtesting at METEC was carried out between October 24-27, 2022.Sensor testing was performed on the natural soil bed buried pipeline pad with the emission source 2.5 ft.below ground level.All times reported in this manuscript are in Mountain Standard Time (UTC-7).Sensor measurements were collected at 5 s intervals and transmitted remotely via cellular network to SCT's cloud server.Measurements were taken for a range of emission rates between 22-37 SLPM (15-26 g hr −1 , assuming natural gas density at 1 atm of 0.712 kg m −3 ). 42For the lateral positioning test, the measurements were taken between a lateral distance between 0-3.05 m east of the emission point with the sample inlet at ground level.For the vertical positioning test, the lateral distance of 1.22 m east of the emission point and a vertical distance ranging from 0 to 0.3 m.Wind data was collected at a frequency of 1 Hz using CSU METEC's weather sensors and included absolute wind speed and wind direction.

Results and Discussion
Machine learning model.-Figure 2 shows the sensor response to between 40-5000 ppm of CH 4 in a simulated natural gas mixture ECS Sensors Plus, 2024 3 011402 at a temperature of 532 °C as read on the heater's integrated thermocouple.strongest sensor response was recorded on the ITO vs Pt electrode, which is consistent with our previous findings. 33 linear response to the logarithm of the CH 4 concentration is also consistent with the expected behavior of mixed potential sensors. 29,42At the lowest end of the concentration, the signal is −20 mV, and when linearly extrapolated to 10 ppm, would still have a signal in the mV range and is easily resolved on our IoT hardware.Based on the required concentration sensitivity in the range of 10-1000 ppm CH 4 as described in Ref. 16, our sensor is expected to be able to detect near surface emissions from buried natural gas pipelines.A 3-layer artificial neural network with 5 hidden layers was evaluated as a quantification method.First, the dataset was split into 566 training data points and 142 test data points, the ANN was trained on the training dataset and evaluated on the test dataset.The training and test accuracy were recorded using the R 2 metric across 20 iterations of random segmentation of the dataset into test and training subsets.An R 2 = 0.997 (Standard Deviation = 0.000515 over 20 iterations) was measured on the training dataset and an R 2 = 0.997 (Standard Deviation = 0.000800 over 20 iterations) was measured on the test dataset.Test accuracy as a function of hidden layer size is shown in Supplementary Information Fig. S1.No gain in accuracy is observed beyond a size of 3 hidden-layer neurons.We chose a size of 5 to be sure the process could work in real time on low cost computing hardware as describe below in case more complex mixture analysis was needed in future applications.When compared with linear regression, we observed a small improvement of using an ANN with 5 hidden layers vs Linear Regression where an average test 2 = 0.990 (SD = 0.001132 over 100 iterations) was recorded.When fit across only the most sensitive electrode pair of ITO vs Pt, a test accuracy R2 of 0.985 (SD = 0.001915 over 100 iterations).Locally, the error is higher near ∼300 ppm CH 4 due to the switchover between two mass flow controllers in the gas mixer operating near their upper and lower limitations.This introduces a +/− 15% increase in error near the crossover point.While this is clearly undesirable, a possible solution would be to use a larger number of CH 4 cylinders and avoid operating near the upper and lower end limitations of the MFCs.
Table I shows the training and inference times for the artificial neural network on both a desktop PC and a Raspberry Pi 4. While the inference time increases by a factor of approximately 1.5x, an inference speed of 0.015 s per data point means assuming, no other overhead, the artificial neural network can operate at a data processing frequency of 6.6 data points per second.Because this data was collected at 5 s per point, the ANN will not be a limiting factor in real time processing of collected sensor data.
Field testing results.-Figure 3 shows the sensor data collected during the measurements on 10/25/2022, with the timeseries data shown in Supplementary Information, Fig. S2.Emission rates between 37 and 26 SLPM were measured directly above the source.The concentration for these settings were on average 3000 ppm.There is no observable change in concentration between these conditions because we hypothesize the CH 4 concentration is limited by the permeation of natural gas through the soil bed.The subsequent measurements taken at emission rates of 22-26 SLPM during that day show a linear decrease in log(CH 4 ) concentration as a function of lateral distance away from the source to a distance of 3 m.Riddick et al. observed and modeled an exponential drop-off of methane concentration from a buried pipeline situated 0.5 m below the surface.Both modeled and measured emissions in their work of an 80 g/hr leak started in the range of 8000-16000 ppm directly above the emission source and decays to <100 ppm at a distance of 3 m.This generally agrees with the measurements of CH 4 concentration that we have collected on our mixed potential sensor. 43][46] Supplementary Information, Figs.S3 and S4 show the sensor measurements as a function of vertical distance at a fixed horizontal distance of 1.2 m during two days of testing on 10/26/22 and 10/27/ 22.The flow rate for these experiments was set to 20 SLPM. Figure 4 shows that the large transient signals collected on 10/26 resulted in wide signal variance and an order of magnitude variance when concentration is calculated.Wind data collected at METEC (Supplementary Information, Fig. S5) shows that from 15:40 to 17:30 on 10/26/22, a sustained east wind would have delivered additional methane towards the collection point.In contrast, over this same time period on the following day, the wind was primarily in the south or north direction which would transport methane away from the collection point.Turbulence associated with wind driving methane towards the collection point would explain the large fluctuations in signal on 10/26/22 but a relatively stable signal on 10/27.The accuracy of the quantification of emission rates is also dependent on the wind speed, as a higher wind speed results in faster dispersion of methane, while a lower wind speed results in the accumulation of methane at the surface level.Using Riddick's empirical model in Eq. 1 that calculates the impact of wind speed on concentration accuracy, 47 for a range of wind speeds between 1 m s −1 and 3.5 m s −1 , the over-or under-estimation of the concentration is expected to vary between 0.35 and −0.28x respectively.A plot of the over-or under-estimation using Riddick's empirical model appears in Supplementary Information, Fig. S6. Figure 4c shows the calculated CH 4 concentration using the artificial neural network regression model and shows a decrease from 10 2 to 10 1 ppm as distance is increased from 0 to 0.3 m.Ulrich et al. 16 performed a set of atmospheric measurements on the same test pad using a Picarro Cavity Ringdown Spectroscopy (CRDS) CH 4 analyzer at a horizontal distance of 1.5 m and observed a decrease from 10 2 to 10 1 ppm from 0 to 0.5 m in height in wind speeds of 2.5 m s −1 .Their results predict a rapid drop in CH 4 concentration within the first half meter from 10 2 to 10 1 ppm.This report also observed fluctuations of up to an order of magnitude in CH 4 concentration at timescales of minutes, which are consistent with our observations.Ultimately, the consistency of our results with the more sophisticated CRDS methane analyzer technology validates the capability of our MPES platform to obtain accurate concentration measurements in the field for methane emissions monitoring.To account for the effects of weather, multiple sensors that predict concentration spatially distributed around the leak point would need to be coupled with atmospheric modeling to provide accurate quantification of emission rates.Finally, because of the limited duration of the field test, repeating the measurements before and after returning for calibration was not performed.In a future effort, long term stability of the sensor responses after prolonged in-field testing should be measured.

Conclusions
Low-cost continuous monitoring solutions for natural gas emissions are vital to address the contribution of methane emissions to global climate change.We have demonstrated in this work sensors with 10-40 ppm CH 4 limit of detection that are suitable for near surface monitoring of underground pipeline.Artificial neural networks were trained with high test accuracy to predict the methane concentration.For the first time, a mixed potential sensor coupled with an internet of things platform was evaluated in the field under real world conditions.The concentrations predicted by our sensor match with previously reported measurements on significantly more sophisticated analyzer technology, indicating our sensor's ability to function as a high accuracy, but low cost and durable sensor system.ECS Sensors Plus, 2024 3 011402

Figure 1 .
Figure 1.Schematic showing a fully connected ANN with 1 hidden layer and 5 neurons in the hidden layer to predict CH 4 concentrations from sensor signals.

Figure 2 .
Figure 2. (a) Signal vs concentration for 40-5000 ppm of CH 4 measured on a multi-electrode mixed potential sensor.The CH 4 sensitive ITO vs Pt pair exhibits the closest match to a linear relationship in this range.(b) An artificial neural network was trained on the sensor data to predict methane concentrations.The predicted vs true values are plotted.The solid line represents a perfect model where the predicted value is equal to the true value.

Figure 3 .
Figure 3. (a) Sensor signal as a function of horizontal position from the emission source.(b) Estimated concentration of CH 4 measured for different emission rates from 22-26 SLPM.The horizontal black bands show the range of training data while data points outside the range were extrapolated using the linear regression mode.Error bars represent standard deviation of 50 data points at 5 s intervals.

Figure 4 .
Figure 4. Sensor data from (a) 10/26/22 and (b) 10/27/22 as a function of vertical distance for an emission rate of 20 SLPM.(c) Concentration as a function of vertical distance shows a decrease in concentration from 100 to 10 ppm over a distance of 0.3 m.The horizontal black line shows the lower range of laboratory collected calibration data while data points below this range the range were extrapolated.Error bars represent standard deviation of 150 data points at 5 s intervals.