How UrbanSentinel works

Two halves that each won a national prize, joined into one loop: a testing module that reads water in seconds and a plasma reactor that treats it. The testing module decides how hard the reactor works and when it stops.

The closed loop

Large treatment plants already adjust UV and ozone doses from indirect signals. UrbanSentinel closes the loop on the compliance parameters themselves, at the scale of a single outfall or village water point, without grid power.

  1. 1

    Sense the water

    Four optical and electrical channels read the sample in place. No reagents and no sample transport.

  2. 2

    Identify the contaminant

    An int8 neural network of about 40–80 kB, running on an ESP32-C6, estimates BOD, COD and coliform.

  3. 3

    Set the plasma

    The same chip sends the driver a power and frequency matched to the contaminant class.

  4. 4

    Treat with plasma

    A zero-voltage-switching driver turns a 12–18 V solar battery into 10–12 kV at the electrode.

  5. 5

    Verify and stop

    The sensor re-tests. Treatment ends when the reading meets the standard, not after a fixed time.

  6. 6

    Record and trace

    Every batch is logged with its spectral fingerprint for compliance records and source tracing.

Built, in testing The connected version, in which the module drives the plasma's switching stage directly, is built and being tested. Published results come from the earlier side-by-side configuration.

The testing module

A pocket-sized reader that estimates the laboratory's slowest results in seconds, and flags anything it has never seen before.

ChannelWhat it readsStatus
UV–visible absorbance, 200–400 nmOrganic load, dyes, protein and nucleic-acid bandsBuilt
Mid-infrared (ATR)Bond vibrations: C–H, O–H, N–H and C=OBuilt
Capacitance and resistanceDissolved solids, conductivity and ionic strengthBuilt
Dispersive Raman, 785 nmMolecular fingerprints for classifying dischargesUpgrade in testing

Testing module

Validated, 28 samples
Processor
ESP32-C6, an AI and IoT microcontroller; runs fully offline
Model
int8 neural network, about 40–80 kB
Estimates
BOD, COD, total and faecal coliform
Anomaly check
Autoencoder trained on clean water; unfamiliar samples are flagged for a laboratory test
Display
1.47-inch screen with a safe, risky or danger score
Accuracy
96.4% mean agreement with the laboratory
Optics
LEDs, a 650 nm laser, a photodiode and a CCD sensor, salvaged from old electronics
The ESP32-C6 testing board. Back and front of the testing board: the ESP32-C6 microcontroller and its 1.47-inch display showing the safe, risky and danger bands.
Salvaged detector optics. The detector's optics came from old electronics: RGB laser, photodiode, UV and IR LEDs, and a CCD/CMOS sensor.
AI testing module in the field. The sensing module reading surface water. An ESP32-C6 microcontroller runs a compact neural network on the device itself, so the result appears on the 1.47-inch screen within seconds and without an internet connection.

Plasma reactor

Built
Input
12–18 V from a solar-charged battery
Output
10–12 kV at the electrode
Driver
Zero-voltage switching, resonant near 45 kHz with a 40 µH tuning inductance
Power
55 W peak
Heating
+2.9 °C over 30 minutes at 55 W: non-thermal
Electrodes
Positive electrode 2 cm above the water; negative electrode immersed
Chemistry
•OH, O₃, H₂O₂, reactive nitrogen species and UV, formed in the water
Added chemicals
None

The plasma reactor

A cold, atmospheric-pressure plasma at the water's surface, driven by a circuit that is almost entirely salvaged.

Series L–C tuning puts the driver at resonance. At about 45 kHz the MOSFETs switch softly, which keeps heat and switching losses low, and with a 40 µH tuning inductance the driver delivers its 55 W peak. Because the treatment works through plasma chemistry rather than heat, the water barely warms.

Frequency test. The oscilloscope shows the switching waveform while the discharge forms at the electrode.
Plasma driver with frequency control. The driver board of the purification stage. A series L–C circuit puts it at resonance near 45 kHz, where the transistors switch softly and the reactor delivers its 55 W peak.
High-voltage switching stage. Salvaged MOSFETs mounted on an aluminium heat sink with ferrite inductors. This zero-voltage-switching stage steps a 12–18 V solar battery up towards the 10–12 kV needed for plasma.
How plasma cleans water. Reactive oxygen and nitrogen species formed at the gas–liquid interface cleave azo dyes and inactivate microbes.Illustration made with Gemini Notebook

How the AI chooses the dose

A three-second disinfection should never pay for a 25-minute dye cycle. Each sample is treated only until the re-test shows it meets the standard.

Contaminant class the AI identifiesPlasma powerTimeTreatment ends when
Microbial: coliform, E. coli25–30 W2–3 sColiform estimate reaches 0 CFU/100 mL
Organic load: BOD and COD60–70 W15 minCOD ≤ 25 mg/L and BOD ≤ 5 mg/L
Textile dye90–110 W25 minColour ≤ 150 Pt-Co
Unfamiliar discharge, PFAS suspected100–120 W15–20 minSample sent for laboratory LC-MS/MS

Discharge targets from Bangladesh's Environment Conservation Rules 2023, as used in the project's tests.

Tracing a discharge to its source

Every treated batch leaves a record of what was in the water. Units placed at sewer junctions and river outfalls combine three independent clues:

  • Fingerprint: the logged spectrum matches a discharge class, such as azo dye or tannery chrome-organic.
  • Position: the plume starts between the last clean unit and the first contaminated one.
  • Timing: arrival times along the river fit the flow from one drain.

A Bayesian model fuses the clues into a ranked list of likely sources with confidence scores, never a single unqualified accusation. High confidence comes only when all three agree.

Designed

Designed and modelled. The first field evidence is planned for the Phase 1 pilot. Every unit still classifies and flags anomalies offline, while forecasting and attribution run on a gateway, aiming to warn water intakes 2–6 hours ahead in Phase 2.

Safety and by-products

  • Nitrate and nitrite checked on every pilot batch against WHO guideline values (50 and 3 mg/L).
  • pH restored by aeration and limestone; firmware caps the maximum exposure.
  • The 10–12 kV stage sits in a sealed enclosure with a lid interlock, a current-limited secondary, residual-current protection and automatic shut-off.
  • Ozone off-gas is vented through activated carbon.

How it compares

Laboratory testingOnline spectral probesUrbanSentinel
Time to result18 hours to 5 daysContinuousSeconds
Typical settingCentral laboratoryLarge treatment plantsOutfalls, intakes, community water points, portable
Treats the waterNoNoYes, with high-voltage plasma
Ends treatment at complianceNoNoYes
Points to the sourceSlow forensic workNoDesigned
Needs grid powerYesYesNo, solar battery