- Introduction
- Project Overview
- System Architecture
- Hardware Specifications
- Sensor Specifications
- Circuit Diagram
- Working Principle
- Statistical Anomaly Detection
- Advanced Embedded Optimizations
- Display & Serial Output
- Software Libraries
- Project Structure
- Limitations
- Future Improvements
- Image References
- Author
Atoms is an embedded environmental monitoring system developed around the ESP32-S3 microcontroller. Instead of relying on manually configured thresholds, the system establishes its own environmental baseline through statistical calibration and continuously monitors for abnormal conditions using adaptive anomaly detection.
The project demonstrates how classical statistical methods can be implemented efficiently on resource-constrained embedded hardware while maintaining low memory usage and real-time operation.
Unlike many demonstration projects, Atoms intentionally avoids overstated claims regarding artificial intelligence or machine learning. Every decision is based on deterministic statistical calculations.
The system continuously monitors:
- Gas concentration
- Ambient temperature
- Ambient light intensity
After an initial calibration period, every sensor reading is compared against its learned baseline using z-score analysis.
Instead of asking,
"Is the value above a fixed threshold?"
the firmware asks,
"How unusual is this value compared to what has normally been observed?"
This allows the monitor to adapt naturally to different environments while remaining sensitive to genuine anomalies.
+----------------+
| ESP32-S3 |
+-------+--------+
|
+--------------+--------------+
| | |
MQ-2 DS18B20 LDR
| | |
+--------------+--------------+
|
Statistical Processing
|
Mean → Variance → Z-Score → Alert
|
SSD1306 OLED Display
|
Serial Monitoring
| Property | Specification |
|---|---|
| Board | ESP32-S3 DevKitC-1 |
| Processor | Xtensa LX7 Dual-Core |
| Clock Speed | Up to 240 MHz |
| SRAM | 512 KB |
| Flash | External SPI Flash |
| Wireless | Wi-Fi + Bluetooth LE |
| ADC | 12-bit |
| Communication | I2C, SPI, UART, GPIO, 1-Wire |
| Device | ESP32-S3 Pin |
|---|---|
| MQ-2 Analog Output | GPIO4 |
| DS18B20 Data | GPIO5 |
| LDR Analog Output | GPIO6 |
| OLED SDA | GPIO8 |
| OLED SCL | GPIO9 |
Purpose
Measures changes in combustible gas and smoke concentration.
Interface
Analog Output
Used for
- Smoke Detection
- LPG
- Hydrogen
- Methane
- Butane
Output
Analog value interpreted statistically rather than through fixed thresholds.
Purpose
Digital temperature measurement.
Communication
1-Wire protocol
Advantages
- Factory calibrated
- Digital output
- High accuracy
- Simple wiring
Only temperature is used within the anomaly detection system.
Humidity measurements were intentionally removed by replacing the earlier DHT22 with the DS18B20.
Purpose
Measures ambient light intensity.
Interface
Analog Output
Used to detect sudden environmental lighting changes.
| Property | Value |
|---|---|
| Resolution | 128 × 64 |
| Interface | I2C |
| Address | 0x3C |
Displays
- Gas Reading
- Light Reading
- Temperature
- Z-Scores
- Current Alert Status
The project wiring is defined in diagram.json.
Figure 1
Complete circuit wiring inside Wokwi.
Figure 2
Live OLED display during monitoring.
Figure 3
Serial Monitor showing sensor values and anomaly detection.
Immediately after power-up the system assumes the surrounding environment is in a normal state.
For approximately 16 seconds, sensor values are collected every 2 seconds.
These samples are used to compute
- Mean
- Variance
- Standard Deviation
These values become the initial baseline.
Once calibration completes:
Every sensor reading is converted into a z-score
z = (Current Reading - Mean) / Standard Deviation
This represents how unusual the current measurement is compared to historical observations.
| Z-Score | Status |
|---|---|
| < 2.0 | SAFE |
| 2.0 – 3.5 | CAUTION |
| > 3.5 | DANGER |
The highest z-score among all sensors determines the current system status.
Instead of remaining fixed forever, the baseline slowly adapts using an Exponential Moving Average.
Mean = Mean + α(Current − Mean)
where
α = 0.01
This allows slow environmental drift while preventing sudden spikes from permanently affecting the baseline.
The baseline updates only while the system remains in the SAFE state.
The firmware performs four major statistical operations.
Average of calibration samples.
Measures the spread of collected data.
Square root of variance.
Determines how far a reading is from normal.
Unlike threshold-based monitoring, this approach adapts automatically to different environments.
No artificial intelligence or machine learning techniques are used.
The firmware is optimized for resource-constrained environments and incorporates several key design patterns:
To minimize the RAM footprint on the ESP32-S3, the Baseline struct overlays variables that are mutually exclusive in time. It uses a union to share memory between the calibration-phase variable M2 (used for Welford's algorithm) and the monitoring-phase variable variance:
struct Baseline {
float mean;
union {
float M2; // Used ONLY during initial calibration
float variance; // Used ONLY during live monitoring
};
int sampleCount;
float alpha;
float minStdDev;
bool needsCalib;
};Sensor baselines are saved periodically (every 1 hour) to Non-Volatile Storage (NVS) using the Preferences library. To extend the flash memory lifespan and prevent corrupting the baseline with anomalous readings:
- Writes are completely bypassed if the current system alert status is not
SAFE. - Baseline updates are rate-limited to happen only when the state has stably drifted under normal conditions.
On boot, the system attempts to load baseline statistics from NVS to perform a "warm boot." To prevent using an outdated baseline (e.g., if the physical environment or sensors have changed since the last power cycle), it runs a sanity check on the first readings. If the initial sensor value has a Z-score greater than 5.0 relative to the saved NVS baseline, the baseline is flagged as stale, and a cold boot calibration is automatically triggered.
OLED displays:
- Gas Reading
- Temperature
- Light Level
- Individual Z-Scores
- Current Alert Status
Typical Serial Output:
Gas=512 z=0.31
Light=340 z=-0.12
Temp=24.1 z=0.05
Status=SAFE
| Library | Purpose |
|---|---|
| Wire | I2C Communication |
| Adafruit GFX | Graphics Rendering |
| Adafruit SSD1306 | OLED Driver |
| OneWire | 1-Wire Communication |
| DallasTemperature | DS18B20 Driver |
| Preferences | ESP32 Non-Volatile Storage (NVS) |
Atmos/
│
├── atoms.ino
├── diagram.json
├── libraries.txt
├── README.md
│
└── assets/
├── pic1.png
├── pic2.png
└── pic3.png
- Calibration assumes the environment is initially normal.
- MQ-2 cannot identify individual gas types.
- Rolling variance uses an EMA approximation rather than the complete Welford algorithm during live monitoring.
Possible future enhancements include:
- Per-sensor alert reporting
- Wi-Fi dashboard integration
- SD card logging
- Cloud telemetry
- Mobile monitoring application
| Image | Description |
|---|---|
|
Complete hardware circuit |
|
OLED monitoring interface |
|
Serial monitor output |
| ------- | ------------- |
Hariom Sharnam