JalSense — Hardware Deep Dive | JalOS Technologies

COMPONENTS

Complete Component Catalog

19 components — each with role, interface, specifications, and INR unit cost at prototype scale.

ESP32-S3

Espressif · Main MCU

₹250–350

Brain. Core 0 = sensor polling + edge AI. Core 1 = comms + storage. Deep sleep between cycles keeps avg current under 3 mA. ESP-NN hardware gives 7.2× TFLite speedup.

UART×3I2C×2SPI×445 GPIO20ch ADC

Cores	Dual Xtensa LX7 @ 240 MHz
RAM	512 KB SRAM + 8 MB PSRAM
AI accel	ESP-NN, 7.2× speedup, int8 ops
Active	~78 mW (23.9 mA @ 3.3V)
Deep sleep	~27 µW (7–150 µA)
TFLite arena	512 KB PSRAM tensor arena
Inference	5–60 ms per model

RYLR998 LoRa (SX1276)

Reyax · Primary uplink

₹450–650

Primary wireless uplink. 1 gateway covers 50+ nodes over 5 km. Sleep current <1 µA. SF7–SF12 adaptive spreading for range vs. speed tradeoff.

UART 9600AT Commands868 MHz ISM

Sensitivity	−148 dBm @ SF12
TX power	+20 dBm max
Airtime SF7	~328 ms for 100B payload
Airtime SF12	~2,636 ms for 100B payload
TX current	~120 mA peak
Sleep current	<1 µA
Max payload	251 bytes (raw LoRa)

SIM7080G NB-IoT

SIMCom · Cellular fallback

₹1,200–1,600

Auto-activates after 3 failed LoRa attempts. Band 28 (700 MHz) gives 50+ km rural coverage. PSM sleep: 3.8 µA. Native MQTT AT commands — no translation layer.

UARTNative MQTTBand 28/3/8

India bands	B28 (Jio/Airtel), B3, B8
PSM sleep	3.8 µA (T3412=1hr, T3324=5min)
TX current	~230 mA active
MQTT cmd	AT+SMQTCONN, AT+SMQTPUB
Rural range	50+ km at Band 28

Atlas EZO-pH

Atlas Scientific · pH

₹3,500–4,500

CPHEEO limit: 6.5–8.5. Below 6.5 = contamination. EZO circuit converts analog probe to I2C digital — eliminates ADC noise. Quarterly calibration.

I2C 0x63UART alt

Range	0.001–14.000 pH
Accuracy	±0.02 pH
Active	22 mA @ 3.3V
Sleep	2.6 mA
Read seq	Write 'R' → 900ms → read 20B ASCII
Calibration	Cal,7 → Cal,4 → Cal,10 (3-point)

Atlas EZO-EC

Atlas Scientific · TDS/Conductivity

₹3,800–4,800

Outputs EC, TDS, salinity, specific gravity simultaneously. Critical for Rajasthan/Gujarat bore wells where TDS often exceeds 2000 ppm. Auto temp compensation.

I2C 0x64Auto temp-comp

EC range	0.07–500,000 µS/cm
Accuracy	±2%
Outputs	EC, TDS (PPM), Salinity, SG
Budget alt	DFRobot SEN0244 (~₹600)

Atlas EZO-DO

Atlas Scientific · Dissolved O₂

₹4,200–5,200

DO <4 mg/L = hypoxia. Detects septic contamination and algae blooms. WHO limit: >6 mg/L. Auto compensation for temperature, salinity, and pressure.

I2C 0x61Press+temp comp

Range	0–saturation (~14 mg/L)
Response	~8 seconds
WHO limit	>6 mg/L drinking water

DFRobot SEN0189 Turbidity

DFRobot · Suspended solids

₹550–800

WHO limit: 1 NTU for drinking water. Monsoon runoff spikes above 100 NTU. IR LED transmittance → 0–5V analog → voltage divider → ADS1115 → NTU lookup table.

Analog 0–5Vvia ADS1115

Range	0.5–500 NTU
Pure water	4.1±0.3V output
Current	~10 mA

Ultrasonic Clamp-on Flow Meter

Industrial Grade · DN15–DN100

₹8,000–12,000

No pipe cutting. No moving parts = zero maintenance. Transit-time ultrasonic. Modbus RTU via MAX485 → ESP32 UART2. ±1–2% accuracy for NRW calculation.

RS485 Modbus RTUMAX485 adapter

Pipe size	DN15–DN100
Accuracy	±1–2%
Reg 0x0000	Flow rate (m³/h × 0.001)
Reg 0x0002	Cumulative total (m³)
Supply	24 VDC, IP68

4–20mA Pressure Transducer

Industrial Grade · 0–10 bar

₹2,000–3,000

Pressure drop = burst pipe or illegal tapping. Loop current immune to 100m+ cable noise. 100Ω shunt → 0.4–2.0V → ADS1115 A0 at 16-bit resolution.

4–20mA loop100Ω shuntADS1115 A0

Range	0–10 bar gauge
Signal	4mA=0bar, 20mA=10bar
Accuracy	±0.5% FS
Wetted	304 stainless steel, 1/4" BSP
Resolution	0.15 mbar via 16-bit ADS1115

A02YYUW Level Sensor

Waterproof Ultrasonic · Tank level

₹800–1,200

Continuous 4-byte UART output — no trigger pin needed. Mounts on tank lid looking down. Calculates fill %, volume in litres. Feeds demand forecasting model.

UART 96004-byte autoIP68

Range	30–450 cm
Accuracy	±3 cm
Packet	0xFF, distH, distL, checksum
Power	5V, ~10 mA

ADS1115 16-bit ADC

Texas Instruments · Signal conversion

₹300–500

16-bit resolution = 65,535 steps. Detects 0.125 mbar pressure changes — enough to find slow leaks. Converts 4–20mA and turbidity voltage to digital for ESP32.

I2C 0x48–0x4B4 channels

Resolution	16-bit (65,535 steps)
PGA range	±256mV to ±6144mV
A0	Pressure 4–20mA shunt
A1	Turbidity 0–3.3V
A2	Valve feedback 4–20mA

DS3231 RTC

Maxim · Accurate timekeeping

₹200–300

±2ppm TCXO = ±1.7 sec/month drift. Timestamps all readings even during power outages. Without RTC, readings during outages become legally inadmissible for CPHEEO reports.

I2C 0x68CR2032 backupTCXO

Accuracy	±2 ppm (±1.7 sec/month)
Backup life	5+ years on CR2032
Alarms	2× programmable + interrupt

INA226 Power Monitor

Texas Instruments · Battery telemetry

₹300–500

V/I/W transmitted to cloud with every report. Operators see battery health remotely. INA226 with 10mΩ shunt gives 10 mA current resolution — detects parasitic drain.

I2C 0x40–0x4F16 addr

Bus voltage	0–36V, 1.25 mV resolution
With 10mΩ	10 mA current resolution
Calibration	CAL = 0.00512 / (I_max × R_shunt)

CN3791 MPPT Solar Charger

CONSONANCE · Solar power

₹400–600

Standalone IC — no firmware needed. Trickle → CC → CV phases protect battery longevity. Set charge current via single resistor: R = 0.5 / I_max_amps.

Analog feedbackPWM buck

Input range	4.5–28V DC
Max charge	2A
Efficiency	~85–90%
Battery reg	4.2V ±1%

Li-Ion Battery 10,000 mAh

18650 Cells · Energy storage

₹1,000–1,500

72-hour offline at full duty cycle. 500+ charge cycles = 5+ year life. Protected PCB prevents over/under-voltage. Recharges from 10W solar in ~5–6 hours.

3.7V nominalProtected PCB

Capacity	10,000 mAh @ 3.7V
Cycles	500+ to 80% retention
Offline	72 hours @ 46 mA avg

u-blox NEO-M8N GPS

u-blox · Location + tamper

₹1,000–1,500

Records install GPS. If device moves (theft/tamper), coordinate changes → alert. Polled 30s once/week. Also syncs RTC time on boot. Multi-constellation reduces urban multipath.

UART 9600NMEA/UBX

Constellations	GPS+GLONASS+BeiDou+Galileo
Accuracy	±2.5m cold, ±0.8m warm
Cold TTFF	~25 seconds
Active	~60 mA acquisition
Usage	30s ON once/week then sleep

DN25 Motorized Ball Valve

Industrial Grade · Flow control

₹5,500–8,000

Proportional control (not just on/off). Position feedback via 4–20mA confirms valve moved. Emergency shutoff on contamination. Manual override lever for maintenance.

0–10V signal4–20mA feedback24VDC

Control	0V=closed, 10V=fully open
Rotation	≤7 seconds
Accuracy	±1% via potentiometer
Pressure	10–15 bar rated

4-ch Opto-isolated Relay

Industrial Grade · Pump + solenoid

₹300–500

Optocoupler isolates ESP32 GPIO from 24V pump inductive spikes. Active-LOW: GPIO pulled LOW = relay ON. CH1: pump, CH2: solenoid, CH3: alarm siren, CH4: spare.

GPIO active-LOWOptocoupler

Channels	4 independent
Rating	10A @ 250VAC / 30A @ 28VDC
Control	<5 mA per channel from 3.3V GPIO

MicroSD Card Module

SPI Interface · Offline buffer

₹200–400

Stores all readings during connectivity gaps. On reconnect, uploads backlog in order — cloud backfills TimescaleDB. 32GB = 15+ years of 5-min interval readings.

SPI 20MbpsFAT32512B blocks

Log format	CSV: ts,pH,EC,DO,turb,flow,press,lvl
Buffer	7-day window before overwrite
Write power	~100 mA, ~1 µA sleep

SIGNAL CHAIN

Signal Chain — Physical Measurement to Digital Value

Every sensor uses a different electrical interface. This is the complete path from physical phenomenon in the water to a digital number inside the ESP32.

I2C Digital BuspH · EC · DO · RTC · INA226 · ADS1115

2 shared wires (SDA + SCL) with 4.7 kΩ pull-ups to 3.3V

ESP32 is master. All sensors share the same bus — unique addresses prevent conflicts. Atlas sensors need 900ms after "R" command before reading back ASCII. I2C mutex semaphore in FreeRTOS prevents concurrent bus access from multiple tasks.

Wire.beginTransmission(0x63); Wire.write('R'); Wire.endTransmission();
delay(900); Wire.requestFrom(0x63, 20); // returns "7.245\r"

4–20mA Current LoopPressure · Valve feedback

Current → 100Ω shunt resistor → 0.4–2.0V → ADS1115 A0

4–20mA loops are immune to noise on cable runs 100m+. A 100Ω shunt converts current to voltage. ADS1115 at 16-bit resolution detects 0.15 mbar pressure changes — enough to find slow leaks. Protection: 1N4148 diodes clamp the shunt from reverse current damage.

V_shunt = I × 100Ω  |  4mA→0.4V (0 bar)   20mA→2.0V (10 bar)
Pressure_bar = (V_shunt − 0.4) × 10 / 1.6

Analog VoltageTurbidity sensor 0–5V

0–5V sensor output → 10kΩ/20kΩ voltage divider → ADS1115 A1

Turbidity sensor max output is 5V but ADS1115 accepts 3.3V. A voltage divider scales it: V_out = 5V × 20k / (10k + 20k) = 3.33V max. Non-linear NTU lookup table stored in ESP32 flash maps voltage to NTU values calibrated during deployment.

V_out = V_in × 20k/(10k+20k) = 3.33V max at ADS1115 A1
NTU = firmware_lookup_table(V_out) // calibrated non-linear curve

RS485 Modbus RTUUltrasonic flow meter

MAX485 chip → differential RS485 → ESP32 UART2

RS485 differential signalling (A/B pair) is immune to noise over 1200m. MAX485 transceiver chip converts to single-ended UART. ESP32 sends Modbus RTU request frame, waits ~50ms, receives 9-byte response with flow rate and cumulative total. DE/RE pins on MAX485 toggled for half-duplex TX/RX switching.

TX: 01 04 00 00 00 02 71 CB  // slave=1, func=04, reg=0x0000, count=2, CRC
RX: 01 04 04 00 0A 00 00 FA 33  // flow=10×0.001=0.01m³/h, total=0m³

UART Auto-OutputA02YYUW level sensor

4-byte continuous UART packets at 9600 baud — no trigger pin required

Sensor continuously sends distance measurements. ESP32 reads on UART1. Header byte 0xFF identifies packet start. Distance in mm from two data bytes. Checksum validation in firmware. Far simpler than JSN-SR04T trigger/echo which needs precise timing from two GPIO pins.

Packet: [0xFF][dist_high][dist_low][checksum]
dist_mm = (dist_high × 256) + dist_low  |  level = tank_height − dist_mm

UART AT CommandsLoRa · NB-IoT · GPS

3 hardware UARTs — one per radio module, managed by dedicated FreeRTOS tasks on Core 1

LoRa (UART0), NB-IoT (UART1), GPS (UART2) each have their own hardware UART peripheral. A binary semaphore per UART prevents task collisions. Responses are ASCII (AT) or binary (GPS NMEA). GPIO interrupt on DIO0 pin signals LoRa TX complete to release semaphore.

LoRa TX: AT+SEND=0,20,hex_payload\r\n → +OK\r\n
NB-IoT: AT+SMQTPUB="topic",45,"payload"\r\n → +SMQTPUB:0\r\n

GPIO Digital + PWM DACRelay · Valve control

GPIO LOW = relay ON · PWM → RC filter → op-amp ×3 → 0–10V valve

Relay module is active-LOW: GPIO output LOW activates relay coil via optocoupler. Valve proportional control: ESP32 PWM (8-bit, 10 kHz) → RC low-pass filter (R=10kΩ, C=10µF, fc=1.6Hz) averages to 0–3.3V DC → LM324 op-amp unity+gain buffer ×3 amplifies to 0–10V. Valve feedback 4–20mA read back on ADS1115 A2 confirms position.

Valve 50%: 50% PWM duty → 1.65V after RC → ×3 → 4.95V → valve 49.5% open
Feedback: 4–20mA → 100Ω → 0.4–2.0V → ADS1115 A2 → position%

EDGE AI

Edge AI Pipeline — On-Device Intelligence

TFLite Micro with ESP-NN hardware acceleration. INT8 quantised models, 512KB PSRAM tensor arena, <50ms total inference. Works completely offline — no cloud needed for real-time decisions.

Why edge AI matters: Contamination events don't wait for connectivity. If the node detects contaminated water while offline, it still must close the valve and trigger the alarm within seconds. Cloud AI handles historical analysis; edge AI handles the immediate autonomous response.

Step 1 — Core 0

FreeRTOS sensor_task

Priority 5, 4KB stack. Polls all sensors every 30 seconds. Acquires I2C mutex semaphore → reads Atlas pH/EC/DO → reads ADS1115 (pressure, turbidity) → reads flow meter via Modbus → reads DS3231 timestamp. Writes to ring buffer in PSRAM.

xTaskCreate priority 5 · Core 0

Step 2 — Core 0

Kalman Filter

Per-channel 1D Kalman filter removes sensor noise spikes. State estimate updated each reading. pH spike from dirty probe → covariance increase → spike discarded. Runs in <1ms per channel.

1D Kalman, O(1) per channel

Step 3 — Core 0

5-Point Moving Average

Ring buffer of last 5 filtered readings per channel. Smooths transient artefacts after Kalman. Output: one representative value per sensor per 30s polling cycle.

Ring buffer, O(1)

Step 4 — Core 1

Feature Vector Assembly

12 normalised values assembled into input tensor: [pH/14, EC/500000, DO/20, turbidity/1000, temp/100, flow/100, pressure/10, Δpressure/10, level/100, Δlevel/100, power/50, tod/24]. Shape: float32[1,12].

Input tensor float32[1,12]

Step 5 — Core 1

Anomaly Detector

Dense neural network: Input[12] → Dense(64, ReLU) → Dense(32, ReLU) → Dense(1, Sigmoid). INT8 quantised, 45KB, 15ms inference. Output: anomaly_score [0.0–1.0]. Threshold: >0.75 = anomaly flag.

MLP INT8 · 45KB · 15ms

Step 6 — Core 1

Quality Classifier

3-layer MLP: Input[12] → Dense(32, ReLU) → Dense(16, ReLU) → Dense(3, Softmax). INT8, 30KB, 20ms. Output: confidence scores for [POTABLE, BORDERLINE, CONTAMINATED]. Threshold: CONTAMINATED >0.8.

MLP INT8 · 30KB · 20ms

Step 7 — Core 1

Leak Detector

Decision tree on 60-second pressure waveform (120 samples). Rapid drop (>3 bar in 10s) = burst pipe. Low-amplitude oscillation (±0.2 bar at 0.5Hz) = slow leak. INT8, 15KB, 8ms.

Decision Tree INT8 · 15KB · 8ms

Step 8 — Core 1

Decision Engine

Priority queue: Emergency > Safety > AI > Schedule > Manual. CONTAMINATED + anomaly >0.9 → close valve (0V) + relay CH3 alarm + QoS2 alert packet. Valve position confirmed via feedback within 10 seconds.

Rule + Model Fusion

Step 9 — Core 0

CRON Scheduler

EEPROM-stored CRON schedule survives power cycles. Sources: manual (dashboard), cloud-pushed (MQTT downlink), sensor-triggered (e.g., open valve when level <20%). Operates fully offline. Audit log per actuation.

CRON in EEPROM · Offline-first

Step 10 — Core 1

MessagePack Assembly

15 fields packed into ~45 bytes: [uint32 timestamp, float32×8 sensors, uint8×3 classes, int16×2 RSSI/SNR, uint8 flags]. Total inference overhead: ~50ms per 30s cycle = 0.28% CPU time on Core 1.

MessagePack ~45B · 0.28% CPU

Step 11 — Cloud

LSTM Demand Forecast

72-hour demand prediction per deployment. Retrained monthly with latest TimescaleDB data. Output: hourly valve scheduling pushed back to edge as CRON update via MQTT downlink.

LSTM · TimescaleDB features

Step 12 — Cloud

NRW + Predictive Maintenance

NRW balance weekly per zone. Sensor drift scoring via rolling regression — predicts calibration failure 7–14 days ahead. Battery discharge curve analysis. Outputs maintenance work orders to field app.

Gradient Boost + Water Balance

DATA FLOW

Micro-Level End-to-End Data Flow

10 layers from physical measurement to push notification. Every protocol, every data format, every timing value, every component — micro-level detail. Scroll each layer independently.

PHYSICAL SENSING

in pipe / tank / pump

pH Probe

I2C 0x63 · UART alt

Output: "7.245\r" ASCII
Read: Write 'R' → 900ms
Accuracy: ±0.02 pH

22 mA active · 2.6 mA sleep

EC/TDS Probe

I2C 0x64 · auto temp-comp

Output: "1250.5,500.3,0.6,1.0002\r"
Fields: EC,TDS,Salinity,SG
Range: 0.07–500k µS/cm

22 mA active · 2.6 mA sleep

DO Probe

I2C 0x61 · press+temp comp

Output: "8.4,95.2\r"
Fields: mg/L, % saturation
Response: ~8 seconds

22 mA active · 2.6 mA sleep

Turbidity Sensor

0–5V analog → 10k/20k divider

Output: 0–3.33V (scaled)
Pure water: 4.1V → 3.33V max
NTU: lookup table in FW

10 mA always-on

Flow Meter

RS485 Modbus RTU 9600

Reg 0x0000: flow rate
Reg 0x0002: cumulative m³
MAX485 → ESP32 UART2

24V supply · poll every 30s

Pressure Xducer

4–20mA → 100Ω shunt

0.4V=0bar, 2.0V=10bar
ADS1115 A0 · 16-bit
Resolution: 0.15 mbar

20 mA always-on (loop)

Level Sensor

UART 9600 auto-output

4B packet: FF,dH,dL,CK
dist_mm = dH×256+dL
Level = height − dist

10 mA · continuous output

INA226 Power

I2C 0x40 · 10mΩ shunt

Reg 0x01: shunt voltage
Reg 0x02: bus voltage
Reg 0x04: current

330 µA · 1.1ms conversion

↓

I2C · RS485 · UART · 4-20mA

raw sensor structs · 30-second polling interval

FreeRTOS sensor_task priority 5 · Core 0

SIGNAL CONDITIONING

ADC conversion + filtering

ADS1115 ADC

I2C 0x48 · PGA ±4.096V

A0: pressure 0.4–2.0V
A1: turbidity 0–3.33V
A2: valve feedback 0.4–2.0V
16-bit · 860 SPS max

<150 µA · 1.1ms/conv

→

Kalman Filter

1D per sensor channel

x̂ₙ = x̂ₙ₋₁ + K(zₙ − x̂ₙ₋₁)
Noise cov Q=0.001, R=0.1
Removes spike artefacts
Runs <1ms per channel

CPU only · no extra HW

→

Moving Average

5-point ring buffer

avg = Σ(last 5) / 5
Per sensor channel
Removes residual noise
O(1) memory per channel

~0.1ms per channel

→

Unit Conversion

firmware lookup tables

pH: ASCII→float parse
EC: ASCII→float parse
Pressure: V→bar formula
Turbidity: V→NTU table
Level: dist_mm→fill%

<1ms total

→

DS3231 Timestamp

I2C 0x68 · ±2ppm TCXO

Unix epoch uint32
±1.7 sec/month drift
CR2032 backup when off
Syncs to GPS on boot

<500 µA · 3µA on backup

→

sensor_reading_t struct

PSRAM ring buffer

{ uint32 ts;
float ph, ec, do, turb;
float flow, press, lvl;
float V, I, W;
uint8 flags; }
~48 bytes per record

xQueue push → Core 1

↓

xQueue IPC

sensor_reading_t struct · 48 bytes · QueueLength=16

FreeRTOS xQueueSend from Core 0 → Core 1

EDGE AI INFERENCE

ESP32-S3 Core 1 · TFLite Micro

Feature Normalise

Core 1 · ml_task priority 5

12 features: [pH/14,
EC/500000, DO/20,
turbidity/1000, temp/100,
flow/100, press/10,
Δpress/10, level/100,
Δlevel/100, power/50,
time_of_day/24]
Input tensor: float32[1,12]

~0.2ms

→

Anomaly Detector

TFLite Micro INT8

MLP: 12→64→32→16→1
Activation: ReLU+Sigmoid
Model size: 45KB
PSRAM arena: 512KB
Threshold: score>0.75
Output: float32 [0,1]

~15ms ESP-NN accel

→

Quality Classifier

TFLite Micro INT8

MLP: 12→32→16→3
Activation: ReLU+Softmax
Model size: 30KB
Classes: POTABLE(0)
BORDERLINE(1)
CONTAMINATED(2)
Threshold: class2>0.80

~20ms ESP-NN accel

→

Leak Detector

Decision Tree INT8

Input: 60s press waveform
(120 float32 samples)
Burst: drop>3bar in 10s
Slow leak: ±0.2bar@0.5Hz
Model size: 15KB
Output: bool+enum

~8ms

→

Decision Engine

Rule + model fusion

Priority queue:
1. Emergency shutoff
2. Safety interlock
3. AI-commanded
4. Scheduled CRON
5. Manual override
CONTAMINATED+score>0.9
→ close valve (0V cmd)
→ relay CH3 alarm
→ QoS2 MQTT alert

~2ms rule evaluation

↓

AI output + sensor data

annotated_reading_t · anomaly_score, quality_class, leak_flag, valve_cmd

xQueueSend to storage_task · Core 1

LOCAL STORAGE + TX QUEUE

Core 1 · DS3231 + MicroSD + xQueue

MessagePack Encoder

Core 1 · mpack library

15 fields → ~45 bytes wire:
0xce [4B] timestamp
0xca [4B] × 8 floats
0xcc [1B] × 3 uint8
0xd1 [2B] × 2 int16
Total: ~45 bytes vs
JSON ~215 bytes (79% save)

~0.5ms encoding

→

MicroSD Logger

SPI 20MHz · FAT32

Path: /YYYYMMDD/log.csv
Append 512B aligned write
7-day circular buffer
~100mA write burst
Daily rotation at midnight
Checksum per record

~0.5s per write burst

→

TX Ring Buffer

PSRAM · 32-item queue

pending_ack[32] array
QoS 1: 5-min telemetry
QoS 2: emergency alert
QoS 0: 60s heartbeat
Priority: alert > telemetry
Retry: 3× on no ACK

in-RAM, no latency

→

Actuator Control

GPIO + PWM + I2C

Valve: PWM→RC→op-amp
Relay CH1: pump GPIO
Relay CH2: solenoid GPIO
Relay CH3: alarm GPIO
Valve feedback: ADS1115 A2
Safety: watchdog 5s

7s valve rotation

↓

TX Queue dispatch

MessagePack payload ~45B → RF transmission every 5 minutes

lora_task priority 4 · Core 0 · or NB-IoT fallback after 3 failures

RADIO TRANSMISSION

SX1276 LoRa · SIM7080G NB-IoT fallback

LoRa Packet Build

SX1276 SPI · SF7 default

MHDR: 0x40 (unconf UL)
FHDR: DevAddr(4B)+FCtrl(1B)
FCnt(2B)+FOpts(0B)
FPort: 10 (telemetry)
11 (alert) 12 (cmd)
FRMPayload: 45B MsgPack
encrypted w/ AppSKey
MIC: AES-CMAC NwkSKey(4B)
Total frame: ~59 bytes

~0.5ms frame assembly

→

LoRa RF Transmit

868 MHz · BW125kHz

SF7 (default): ~328ms TX
SF12 (max range): ~2636ms
TX power: +20 dBm max
TX current: ~120mA peak
DIO0 interrupt: TX done
ADR: RSSI<−120→↑SF
Airtime duty cycle: <1%

UART: AT+SEND=0,59,hex

NB-IoT Fallback

SIM7080G · Band 28/3/8

Activates: 3 LoRa failures
PSM: T3412=1hr,T3324=5min
Sleep: 3.8µA (PSM off)
TX: ~230mA active
AT+SMQTPUB direct
MQTT over TCP 8883 TLS
Session: persistent clean=0

~2-5s to acquire + TX

→

Backlog Upload

On reconnect · SD card

Read CSV from SD in order
Re-encode to MsgPack
Batch upload to MQTT
QoS 1 per record
Cloud fills TimescaleDB gaps
Up to 7 days buffered
~50KB/day at 5min intervals

~1KB/sec upload rate

↓

LoRa RF 868 MHz

Encrypted LoRaWAN frame · NwkSKey+AppSKey AES-128 · MIC 4B

2–15km range · SF7–SF12 adaptive · <1% duty cycle ISM band

LORA GATEWAY

RAK7268 · 1 per 5km · 50 nodes

RAK7268 Gateway

SX1301 8-ch concentrator

8 simultaneous channels
BW: 125/250/500 kHz
All SFs simultaneously
Listens all devices in range
RX window: continuous
LTE/Ethernet backhaul
Solar-capable mounting

50 nodes · 5km radius

→

UDP Packet Forwarder

UDP port 1700 → ChirpStack

Semtech pkt fwd protocol
PUSH_DATA type 0x02:
{ rxpk: [{ time, tmms,
chan, rfch, freq, stat,
modu, datr:"SF7BW125",
codr:"4/5", rssi:-95,
lsnr:5.2, size:59,
data:"QEd..." (base64)}]}
JSON over UDP

<10ms forwarding latency

→

Gateway Bridge

ChirpStack GW bridge

Translates UDP→MQTT
Topic: gateway/{eui}/event/up
Re-encodes to Protobuf
Forwards to NS MQTT
Handles PULL_DATA
for downlink scheduling

TLS MQTT to NS

↓

UDP Semtech Protocol → MQTT Protobuf

{ rxpk: [{datr:"SF7BW125", rssi:-95, lsnr:5.2, data: base64_payload}] }

LTE/Ethernet backhaul · <50ms end-to-end gateway → NS

CHIRPSTACK NS

OTAA · MIC verify · Dedup · AppSKey decrypt

Frame Deduplication

Redis · 200ms window

Multiple GW send same frame
Dedup key: DevAddr+FCnt
Window: 200ms
Select: best RSSI/SNR
Forward: 1 copy only
Redis bloom filter store

200ms dedup window

→

OTAA Session

NwkSKey + AppSKey

1× at activation

→

MIC Verification

AES-CMAC NwkSKey

~0.5ms crypto

→

FRMPayload Decrypt

AES-128-CTR AppSKey

FRMPayload(45B) → plaintext
Key: AppSKey (128-bit)
Nonce: DevAddr+FCnt+dir
Hardware AES on ESP-S3
Decrypted payload passed
to Application Server
via gRPC integration

~0.5ms crypto

→

Application Server

ChirpStack AS · gRPC

Receives plaintext payload
Adds context: device metadata
zone_id, deploy_location
Publishes to EMQX MQTT
Topic: jalsense/{id}/telemetry
QoS 1 · retain=false

<5ms total NS processing

↓

MQTT v5 · TLS 1.3 · X.509 mutual auth

topic: jalsense/{device_id}/telemetry · QoS 1 · payload: MsgPack 45B

Persistent session · LWT: jalsense/{id}/status {"status":"offline"}

EMQX + KAFKA

Message broker → event streaming

EMQX Broker

MQTT v5 · 10k+ connections

Topics:
jalsense/+/telemetry (QoS1)
jalsense/+/alert (QoS2)
jalsense/+/command (QoS1)
jalsense/+/heartbeat (QoS0)
LWT: jalsense/+/status
Session: persistent clean=0
TLS 1.3 · X.509 per-device
Storage: RocksDB ~1KB/session

10k+ concurrent conns

→

EMQX Rule Engine

SQL-like rules → Kafka

Rule: SELECT * FROM
"jalsense/+/telemetry"
→ Kafka raw-telemetry
Rule: SELECT * FROM
"jalsense/+/alert"
→ Kafka alerts
No transformation at EMQX
Pass-through to Kafka

<1ms rule eval

→

Kafka Topics

Partition key: device_id hash

raw-telemetry: 5 partitions
retention: 30 days
Snappy compress: ~4:1
alerts: 3 partitions
retention: 24 hours
valve-commands: 5 parts
Replication factor: 3
Idempotent producer: on

~100k msg/sec capacity

→

Consumer Groups

3 independent groups

telemetry-writer:
reads raw-telemetry
writes TimescaleDB
commit every 5k msgs
alert-engine:
reads raw-telemetry+alerts
evaluates thresholds
ml-pipeline:
reads raw-telemetry
batch 100 msgs or 30s

Independent lag

↓

Kafka consumer → TimescaleDB + ML + Alert Engine

Kafka offset commit every 5,000 records · exactly-once semantics on alerts

Partition key: device_id hash → ordered processing per device

CLOUD STORAGE + ML

TimescaleDB · ML Pipeline · Alert Engine

TimescaleDB

PostgreSQL hypertable

Table: telemetry
Chunk interval: 1 day
Schema: (time,device_id,
ph,ec,do,turbidity,
flow_rate,pressure,
level,power_w,
anomaly_score,quality_class,
valve_pct,rssi,snr)
Index: (device_id,time DESC)
Compress after: 1 day (90%)

Raw 90d→hourly 2yr→daily ∞

→

Continuous Aggregates

TimescaleDB auto-refresh

telemetry_1h: refresh hourly
AVG(ph,ec,do,turb,flow)
MAX(anomaly_score)
COUNT(readings)
telemetry_1d: refresh daily
SUM(flow×15/60) daily_liters
MAX(anomaly) daily_max
Query <50ms for 7-day range
Query <20ms for 2-year hourly

Auto background refresh

→

ML Pipeline

Python workers · scheduled

02:00 daily: LSTM demand
72-hr forecast per device
Input: 7-day hourly history
Output: demand_forecasts tbl
03:00 Sunday: NRW analysis
flow_in−consumed−loss=NRW%
04:00 daily: pred maintenance
sensor drift regression
battery discharge curve

Cron: 02:00/03:00/04:00

→

Alert Engine

Kafka consumer · real-time

Evaluate per-reading:
a) Operator thresholds
b) Edge AI score in packet
c) Rate-of-change rules
pH drop >0.5 in 1min
Dedup: Redis bloom filter
30-min TTL per alert type
Route: GPS zone → engineer
Escalation: 15m→supervisor
30m→ops manager
60m CRITICAL→nodal SMS

<100ms per reading eval

↓

FCM HTTP v1 · WebSocket · Webhook

POST https://fcm.googleapis.com/v1/projects/{id}/messages:send · topic: zone_{zone_id}

OAuth2 service account · Bearer token · 4KB max payload

L10

DELIVERY TO OPERATORS

FCM · WebSocket · WhatsApp · Auto-reports

FCM Push (Android)

Firebase Cloud Messaging v1

Topic: "zone_{zone_id}"
All engineers subscribed
Notification payload:
title: "CRITICAL: Contam."
body: "Zone 3 · pH 5.2"
Data payload:
device_id, severity,
anomaly_score, action
Deep link → device detail
Retry: 24hr exponential

<2 seconds delivery

WebSocket Dashboard

React + Socket.IO

Event: "telemetry_update"
Event: "alert_created"
Event: "node_status_change"
GIS map: node turns RED
Alert triage queue update
Valve control panel sync
Live chart data stream
Connected: govt operators

Real-time <100ms

WhatsApp Webhook

WhatsApp Business API

Supervisor channel message:
"[CRITICAL] Zone 3 · T+0s
pH 5.2 · Turbidity 180NTU
Valve auto-closed.
Engineer Rahul notified.
Tap to view dashboard."
Template message (pre-approved)
Webhook POST from alert engine

<5 seconds delivery

Auto Reports

Cron job · PDF generation

Daily 00:00: quality report
PDF+CSV per deployment
CPHEEO format
Weekly: NRW summary
production−billed−loss
Monthly: SLA performance
uptime%, alert response time
On-demand: tender data pack
historical export for govt

Cron: 00:00 daily

T+0 to notification: Physical contamination detected → edge AI closes valve (T+2s) → LoRa packet transmitted (T+5s) → EMQX received (T+8s) → alert created (T+12s) → FCM push delivered (T+25–45s). Entire pipeline end-to-end: under 45 seconds.

Offline resilience: Up to 7 days of data stored on MicroSD with DS3231 timestamps. On reconnect, backlog uploads in chronological order. TimescaleDB auto-backfills. Redis deduplication prevents duplicate alerts from replayed records.

INSTALLATION

Installation Flow — Site Visit to Live Monitoring

8 steps from first site visit to node going GREEN on the government GIS map.

Site Survey

Measure pipe diameter — select correct flow meter transducer spacing
Check pipe material (cast iron needs ultrasonic coupling gel)
Survey sun exposure — south-facing, no shadow 8AM–4PM
Test NB-IoT signal: AT+CSQ, value >10 = acceptable
Measure tank dimensions for level sensor offset calculation
Photograph + record GPS in field app before any work

Mechanical Installation

Mount enclosure on DIN rail or IP67 wall bracket
Clamp flow transducers (upstream + downstream per DN spacing table)
Thread pressure transducer into 1/4" BSP tapping (3 turns Teflon)
Mount A02YYUW on tank lid, face down, clear of inlet splash
Insert Atlas probes through bulkhead fittings into sample port
Mount turbidity in bypass loop — protects from direct flow debris
Route cables through IP67 glands into enclosure
Mount valve inline between isolation valves for maintenance access

Electrical Wiring

Solar panel → CN3791 input (red=+6V, black=GND)
Battery → CN3791 output + INA226 shunt in series on BAT+
Atlas sensors → shared I2C (SDA, SCL, 3.3V, GND)
ADS1115 → same I2C bus (ADDR pin sets address)
Pressure 4–20mA → 100Ω shunt → ADS1115 A0
Turbidity 0–5V → 10k/20k divider → ADS1115 A1
MAX485 transceiver ↔ flow meter RS485 ↔ ESP32 UART2
PWM → RC filter → op-amp ×3 → valve 0–10V input

Firmware Provisioning

Power on → boots in provisioning mode (WiFi AP: JalSense-XXXX)
Field engineer connects phone to JalSense AP
JalOS App: scan QR code on enclosure → auto-fills DevEUI
Enter: site name, pipe DN, LoRa gateway ID, NB-IoT APN
App pushes config to device over WiFi → stored in NVS encrypted
DS3231 RTC synced to GPS time on first fix

Sensor Calibration

pH: pH7 buffer→Cal,7→rinse→pH4→Cal,4→rinse→pH10→Cal,10
EC: Cal,dry → immerse in 84 µS/cm standard → Cal,84
DO: 30s in air → Cal (100% air saturation)
Turbidity: pour reference solution → enter NTU value in app
Pressure: record zero-flow offset for baseline calibration
Level: enter actual tank depth (cm) → app auto-calculates offset
All calibration values stored NVS + synced to cloud cal DB

OTAA Join + Go-Live

LoRa OTAA JoinRequest → ChirpStack assigns DevAddr + session keys
Field app confirms gateway receipt of test packet <10 seconds
GPS fix → tamper baseline position stored in cloud
Dashboard shows node GREEN on GIS map
Operator configures supply schedule (e.g., valve OPEN 6AM–8AM)
Alert thresholds set: pH <6.5, turbidity >4 NTU

Cloud Configuration

Assign node to zone (GPS-based field engineer routing)
Configure MQTT LWT: jalsense/{id}/status offline message
Enable continuous aggregate refresh for this device in TimescaleDB
First automated daily report generated at midnight
LSTM demand model initialised with default parameters (no history yet)
Node added to NRW zone calculation for this municipal cluster

Maintenance Schedule

Monthly: pH probe clean + single-point recal check
Quarterly: full 3-point pH, EC recal, DO membrane check
6-monthly: turbidity reference verification
Annual: flow transducer coupling gel, pressure zero-point
As-needed: predictive maintenance alert → field app work order
Solar panel clean after monsoon (silt reduces output 20–30%)

CLOUD

Cloud Pipeline — MQTT to Analytics

From EMQX message receipt to TimescaleDB storage to ML inference to alert routing. Every stage designed for 10→10,000 node scale.

EMQX MQTT
Broker

Persistent sessions · LWT · QoS 1/2 store-and-forward

Receives MessagePack from JalSense via LoRa gateway or NB-IoT. Topics: jalsense/{device_id}/telemetry (QoS 1, 5-min interval), jalsense/{device_id}/alert (QoS 2, immediate). LWT: if TCP drops without DISCONNECT, broker publishes jalsense/{id}/status {"status":"offline"} after 30s. X.509 per-device certificates for mutual TLS 1.3 auth. Persistent sessions (CleanStart=false) queue messages during connectivity gaps.

Apache Kafka
Event Stream

5 partitions by device_id hash · 30-day retention · 3× replication

EMQX rule engine forwards all messages to Kafka topics: raw-telemetry (5 partitions, partition key = device_id hash → ordered processing per device), alerts (3 partitions), valve-commands. Snappy compression: ~4:1 ratio. Idempotent producer prevents duplicate writes. Replay window: 30 days allows reprocessing historical data with updated ML models.

TimescaleDB
Hypertable

Chunk interval: 1 day · Compression: 90% after 1 day · Index: (device_id, time DESC)

Hypertable telemetry partitioned by time (1-day chunks) and device_id (hash, 2 spaces). Raw data retained 90 days → continuous aggregates: hourly averages for 2 years, daily averages forever. Partial index on anomaly_score > 0.5 keeps anomaly queries fast. Typical query latency: <50ms for 7-day raw range, <20ms for 2-year hourly aggregate.

ML Pipeline
Scheduled

LSTM daily · NRW weekly · Predictive maintenance daily

Python ML workers consume from Kafka. LSTM demand forecasting (02:00 AM daily per deployment): reads 7-day hourly history from TimescaleDB, outputs 72-hour demand forecast stored back to demand_forecasts table. NRW analysis (03:00 Sunday): flow_produced − flow_billed − losses = NRW% per zone. Predictive maintenance (04:00 daily): rolling regression on calibration residuals detects sensor drift 7–14 days ahead.

Alert Engine
Real-time

Redis deduplication · GPS zone routing · 30min/supervisor escalation ladder

Kafka consumer evaluates every incoming reading in real-time: (a) operator-configured thresholds per device, (b) edge AI anomaly_score forwarded in packet, (c) rate-of-change rules (pH drop >0.5 in 1 minute). Redis bloom filter deduplicates: same alert type suppressed 30 minutes. GPS zone lookup assigns field engineer. FCM HTTP v1 → push notification in <2 seconds. Escalation: 15min unack → supervisor; 30min → ops manager; 60min CRITICAL → government nodal officer SMS.

ALERTS

Alert Flow — Contamination to Phone in 45 Seconds

Step-by-step timeline from edge AI detection to push notification on a field engineer's Android phone.

T + 0s — TRIGGER

Edge AI detects contamination

TFLite quality classifier: CONTAMINATED confidence 94% >0.80 threshold. Decision engine: close valve (GPIO PWM→0V), relay CH3 alarm ON, QoS2 flag set in TX queue.

T + 2s — LOCAL ACTION

Valve closes in 7 seconds, alarm sounds

Motorised valve receives 0V → closes. Relay CH3 activates siren. DS3231 timestamps event. SD card logs: all 15 sensor values, anomaly score, quality confidence, valve command.

T + 5s — LORA TX

QoS 2 alert packet transmitted

LoRaWAN FPort=11 (alert). Confirmed uplink (MHDR=0x80). MIC appended with NwkSKey AES-CMAC. If no ACK in 5s: NB-IoT activates as fallback. AT+SMQTPUB with QoS 2.

T + 8s — CLOUD RECEIPT

ChirpStack MIC-verifies, EMQX receives

ChirpStack deduplicates (200ms window, best RSSI gateway selected). Decrypts FRMPayload with AppSKey. EMQX ACKs QoS 2 (4-step handshake). Kafka routes to alert-engine consumer.

T + 12s — ALERT RECORD

PostgreSQL alert record created

Schema: (id, device_id, timestamp, type=CONTAMINATION, severity=CRITICAL, sensor_values JSON, ai_scores JSON, zone_id, assigned_engineer_id, status=OPEN, created_at).

T + 18s — ROUTING

GPS zone lookup → FCM topic publish

Zone lookup: GPS coordinates → zone_id → assigned field engineer. FCM HTTP v1 POST to topic zone_{zone_id}. Webhook fires to WhatsApp Business API for supervisor channel.

T + 25s — DELIVERY

Push notification on engineer's phone

"CRITICAL: Contamination at Tank B7, Zone 3. pH 5.2, Turbidity 180 NTU. Valve auto-closed. Tap to view." Deep link → device detail screen with full sensor timeline.

T + 45s — DASHBOARD

Government GIS map node turns RED

WebSocket event alert_created pushes to all connected dashboard clients. Node icon RED. Alert in triage queue. Supervisor sees full sensor history, AI confidence, and valve status.

Component	State	mA	Duty %	Avg mA	Notes
ESP32-S3	Active (sensing + AI)	24	10%	2.40	Core 0: sensors; Core 1: comms
ESP32-S3	Deep sleep	0.01	90%	0.009	ULP coprocessor only
Atlas EZO-pH	Active read	22	3%	0.66	Sleep 2.6 mA between reads
Atlas EZO-EC	Active read	22	3%	0.66	Same sleep pattern
Atlas EZO-DO	Active read	22	3%	0.66	8s response, awake 10s/read
DFRobot Turbidity	Always on	10	100%	10.00	Power-gate via transistor to save
Ultrasonic Flow Meter	Active poll	80	5%	4.00	24V via boost converter
Pressure Transducer	Always on	20	100%	20.00	4–20mA loop — biggest draw
ADS1115	Convert bursts	0.15	5%	0.008	Negligible
DS3231 RTC	Always on	0.5	100%	0.50	CR2032 backup on outage
INA226	Always on	0.33	100%	0.33	Continuous battery monitoring
LoRa (TX burst)	Transmitting	120	0.5%	0.60	5-min interval, ~2s per TX
LoRa (idle)	Sleep	1	99.5%	1.00	Not in continuous RX
SIM7080G NB-IoT	PSM T3412=1hr	0.004	99%	0.004	Fallback only, 3.8µA PSM
NEO-M8N GPS	Weekly 30s fix	50	0.05%	0.025	30s ON once/week
MicroSD	Write burst	100	1%	1.00	~0.5s write per 5-min cycle
Relay (1 active)	Pump control	80	5%	4.00	Not always active
TOTAL AVERAGE				~46 mA	10,000 mAh / 46mA = 217hrs (9 days)

BOM & COST

BOM & Cost Analysis

34 line items, full Bill of Materials with INR unit costs at prototype (1–10 units) and production (100+ units) scale. Sources: Robu.in, Mouser India, IndiaMART, JLCPCB.

#	Component	Supplier / Source	Qty	Proto (₹)	Prod 100+ (₹)	Line Total
COMPUTE SUBSYSTEM
1	ESP32-S3 Development Module	Espressif / Robu.in	1	300	200	300
2	DS3231 RTC Module (with CR2032)	Maxim / Robu.in	1	250	150	250
3	MicroSD Card Module (SPI)	Generic / Robu.in	1	300	180	300
4	MicroSD Card 32 GB Samsung	Amazon India	1	450	350	450
Compute Subtotal						1,300
WATER QUALITY SENSORS
5	Atlas Scientific EZO-pH Circuit + Probe	Atlas Scientific (direct)	1	4,000	3,400	4,000
6	Atlas Scientific EZO-EC Circuit + Probe	Atlas Scientific (direct)	1	4,300	3,600	4,300
7	Atlas Scientific EZO-DO Circuit + Probe	Atlas Scientific (direct)	1	4,700	4,000	4,700
8	DFRobot SEN0189 Turbidity Sensor	DFRobot / Robu.in	1	650	400	650
9	DS18B20 Waterproof Temperature Probe	Maxim / Robu.in	1	100	60	100
Water Quality Subtotal						13,750
FLOW & PRESSURE SUBSYSTEM
10	Ultrasonic Clamp-on Flow Meter (DN25)	Industrial / IndiaMART	1	10,000	7,500	10,000
11	4–20mA Pressure Transducer 0–10 bar	Industrial / Moglix	1	2,500	1,800	2,500
12	A02YYUW Ultrasonic Level Sensor	DFRobot / Robu.in	1	1,000	650	1,000
13	ADS1115 16-bit ADC Module	Adafruit / Robu.in	1	400	250	400
14	MAX485 RS485 Transceiver Module	Generic / Robu.in	1	80	50	80
Flow & Pressure Subtotal						13,980
TRANSMISSION SUBSYSTEM
15	RYLR998 LoRa Module (SX1276)	Reyax / Robu.in	1	550	380	550
16	SIM7080G NB-IoT/GNSS Module	SIMCom / Waveshare	1	1,400	1,100	1,400
17	u-blox NEO-M8N GPS Module	u-blox / Robu.in	1	1,200	900	1,200
18	NB-IoT SIM Card (Airtel IoT)	Airtel Business	1	200	150	200
19	LoRa Antenna 868 MHz 3 dBi	Generic / Robu.in	1	150	100	150
Transmission Subtotal						3,500
POWER SUBSYSTEM
20	10W Monocrystalline Solar Panel	Loom Solar / Amazon	1	700	500	700
21	CN3791 MPPT Charger Module	Generic / Robu.in	1	500	300	500
22	Li-Ion Battery Pack 10,000 mAh 3.7V	Generic 18650 pack	1	1,200	900	1,200
23	INA226 Power Monitor Module	TI / Robu.in	1	400	250	400
24	USB-C Port + Cable Gland Set	Generic	1	300	180	300
Power Subtotal						3,100
ACTUATOR SUBSYSTEM
25	DN25 Motorized Ball Valve (0–10V, 24V)	Industrial / IndiaMART	1	6,500	4,500	6,500
26	4-Channel Opto-isolated Relay Module	Generic / Robu.in	1	400	250	400
27	LM324 Op-Amp + RC filter passives	LCSC / Robu.in	1	100	60	100
Actuator Subtotal						7,000
ENCLOSURE & MECHANICAL
28	IP67 ABS Enclosure 200×150×80mm	Fibox / IndiaMART	1	2,000	1,400	2,000
29	DIN Rail + Wall Mount Bracket	Generic	1	300	200	300
30	IP67 Cable Glands (set of 6)	Hawke / IndiaMART	1	400	250	400
31	Terminal Blocks + DIN Rail (set)	Phoenix Contact	1	500	350	500
Enclosure Subtotal						3,200
PCB & MISCELLANEOUS
32	Custom PCB 2-layer 100×80mm	JLCPCB / PCBWay	1	800	350	800
33	Passive components (R, C, connectors)	LCSC India	1	400	200	400
34	Wire, connectors, heat shrink set	Generic	1	300	180	300
PCB & Misc Subtotal						1,500
TOTAL BOM COST PER UNIT						₹ 47,330

Biggest cost driver: Atlas sensors (pH+EC+DO = ₹13,000) = 27% of total BOM. Budget alternative: DFRobot analog sensors (~₹1,500 total). Saves ₹11,500/unit but reduces accuracy from ±0.02 to ±0.1 pH and needs more frequent calibration.

Prototype Scale (1–10 units)

BOM₹ 47,330

Assembly (manual)₹ 3,000

Testing + calibration₹ 2,000

PCB (5-pc run)₹ 1,200

Total Unit Cost₹ 53,530

Production Scale (100+ units)

BOM₹ 33,100

PCBA + test jig₹ 1,500

Testing + calibration₹ 800

PCB panel₹ 350

Total Unit Cost₹ 35,750

Government Pricing Model

Hardware (to utility)₹ 80k–1.2L

SaaS/month/node₹ 3,000–5,000

Annual maintenance₹ 8,000–12,000

HW gross margin~55–65%

SaaS gross margin>80%

Budget Version (₹28K target)

Swap Atlas → DFRobotSave ₹ 11,500

Swap 4–20mA → digitalSave ₹ 2,000

Remove GPSSave ₹ 1,200

Remove DO sensorSave ₹ 4,700

Budget BOM~₹ 27,930

JalSense Hardware Deep Dive

Every Component · Every Signal · Every Byte

System Overview — 6 Subsystems

Water Quality Sensors

Flow & Pressure

Compute Subsystem

Transmission

Power Subsystem

Actuator Subsystem

Complete Component Catalog

ESP32-S3

RYLR998 LoRa (SX1276)

SIM7080G NB-IoT

Atlas EZO-pH

Atlas EZO-EC

Atlas EZO-DO

DFRobot SEN0189 Turbidity

Ultrasonic Clamp-on Flow Meter

4–20mA Pressure Transducer

A02YYUW Level Sensor

ADS1115 16-bit ADC

DS3231 RTC

INA226 Power Monitor

CN3791 MPPT Solar Charger

Li-Ion Battery 10,000 mAh

u-blox NEO-M8N GPS

DN25 Motorized Ball Valve

4-ch Opto-isolated Relay

MicroSD Card Module

Signal Chain — Physical Measurement to Digital Value

2 shared wires (SDA + SCL) with 4.7 kΩ pull-ups to 3.3V

Current → 100Ω shunt resistor → 0.4–2.0V → ADS1115 A0

0–5V sensor output → 10kΩ/20kΩ voltage divider → ADS1115 A1

MAX485 chip → differential RS485 → ESP32 UART2

4-byte continuous UART packets at 9600 baud — no trigger pin required

3 hardware UARTs — one per radio module, managed by dedicated FreeRTOS tasks on Core 1

GPIO LOW = relay ON · PWM → RC filter → op-amp ×3 → 0–10V valve

Edge AI Pipeline — On-Device Intelligence

FreeRTOS sensor_task

Kalman Filter

5-Point Moving Average

Feature Vector Assembly

Anomaly Detector

Quality Classifier

Leak Detector

Decision Engine

CRON Scheduler

MessagePack Assembly

LSTM Demand Forecast

NRW + Predictive Maintenance

Micro-Level End-to-End Data Flow

Installation Flow — Site Visit to Live Monitoring

Site Survey

Mechanical Installation

Electrical Wiring

Firmware Provisioning

Sensor Calibration

OTAA Join + Go-Live

Cloud Configuration

Maintenance Schedule

Cloud Pipeline — MQTT to Analytics

Persistent sessions · LWT · QoS 1/2 store-and-forward

5 partitions by device_id hash · 30-day retention · 3× replication

Chunk interval: 1 day · Compression: 90% after 1 day · Index: (device_id, time DESC)

LSTM daily · NRW weekly · Predictive maintenance daily

Redis deduplication · GPS zone routing · 30min/supervisor escalation ladder

Alert Flow — Contamination to Phone in 45 Seconds

Edge AI detects contamination

Valve closes in 7 seconds, alarm sounds

QoS 2 alert packet transmitted

ChirpStack MIC-verifies, EMQX receives

PostgreSQL alert record created

GPS zone lookup → FCM topic publish

Push notification on engineer's phone

Government GIS map node turns RED

Power Budget

BOM & Cost Analysis