Alaafia PCB local

Fall Sensor - ESP32-C3 Wearable Fall Detection System

Overview

3D render of the Fall Sensor carrier PCB (v1, single-layer)

An intelligent wearable fall detection system built on the ESP32-C3 Super Mini with BMI160 IMU, targeting elderly users with waist/lower-back placement. The system combines a 4-stage rule-based state machine with an on-device CNN (TFLite Micro, int8 QAT) for high-confidence detection. A persistent alerted state holds the alert active until the person's posture changes or they self-dismiss, preventing the common problem of fall alerts clearing while the person is still on the floor.

Real-time data streams over WiFi TCP to a PyQtGraph dashboard and InfluxDB/Grafana stack for threshold tuning and post-event analysis. A dual-purpose button handles both "I'm okay" dismissal and manual "I need help" alerts.

Hardware

| Board | 5x7cm perfboard | Power/signal zone separated |

Hand-soldered perfboard build, ESP32-C3 Super Mini + BMI160

Pin assignments (ESP32-C3 Super Mini):

Power path: LiPo -> TP4056 (charge management) -> MT3608 (boost to 5V) -> ESP32-C3 5V pin -> onboard LDO -> 3.3V rail -> BMI160 VCC

Project Structure

Fall_Sensor/
├── firmware_idf/                    # ESP-IDF project (active)
│   ├── CMakeLists.txt
│   ├── partitions.csv               # custom 3MB app partition for ESP32-C3 4MB flash
│   ├── sdkconfig.defaults
│   ├── main/
│   │   ├── CMakeLists.txt
│   │   ├── fall_sensor.cpp          # app_main() entry point
│   │   ├── main.cpp                 # state machine + BMI160 + tasks
│   │   ├── inference.cpp            # TFLite Micro wrapper + consecutive filter
│   │   ├── inference.h
│   │   ├── wifi_stream.cpp          # WiFi TCP server for dashboard streaming
│   │   ├── wifi_stream.h
│   │   └── fall_model.h             # auto-generated model C header
│   └── components/
│       ├── esp-tflite-micro/        # git submodule (espressif official)
│       ├── DFRobot_BMI160/          # git submodule (unmodified)
│       └── DFRobot_BMI160_wrapper/  # local CMakeLists.txt wrapper (tracked)
├── firmware/                        # legacy PlatformIO attempt - kept for reference
├── docs/
│   ├── ERRORS_AND_FIXES.md
│   └── ML_README.md
├── hardware/                        # KiCad schematics, BOM, perfboard layout
├── visualization/

Cloning

This project uses git submodules. Clone with:

git clone --recurse-submodules <repo-url>

If already cloned:

git submodule update --init --recursive

Building and Flashing

Open ESP-IDF terminal in VS Code (Ctrl+Shift+P -> ESP-IDF: Open ESP-IDF Terminal):

cd firmware_idf
idf.py set-target esp32c3
idf.py build
idf.py -p COM7 flash monitor

Exit monitor: Ctrl+]

On first build after a clean or target change, delete sdkconfig (not sdkconfig.defaults) before running set-target to avoid stale target conflicts.

How It Works

State Machine

STATE_NORMAL
  accelMag < 3.0 m/s²  ->  STATE_FREEFALL  (timeout 400ms)
  
STATE_FREEFALL
  accelMag > 20.0 m/s² ->  STATE_IMPACT    (window 500ms)
  
STATE_IMPACT
  gyroMag < 25 deg/s   ->  STATE_POST_IMPACT_STILLNESS
  
STATE_POST_IMPACT_STILLNESS
  stillness 2.5s + orientation change + NN confirms  ->  STATE_FALL_CONFIRMED
  
STATE_FALL_CONFIRMED
  captures fall orientation, triggers buzzer, fires FALL event
  ->  STATE_FALL_ALERTED  (persistent - does not auto-clear)
  
STATE_FALL_ALERTED
  re-buzzes every 10s
  posture change > 2.0 m/s² delta  ->  STATE_NORMAL
  button press ("I'm okay")        ->  STATE_NORMAL

Any state: button press outside STATE_FALL_ALERTED -> manual "I need help" alert

ML Inference

• Model: lightweight 1D CNN, int8 QAT, 16.9KB flash, 47.1KB RAM
• Input: 512-sample sliding window of accel Y (in g) at 200Hz = 2.56 second window
• Consecutive filter: 3 consecutive runs >= 80% probability required before confirming
• Falls back to rule-based alone if window not yet full

Sensor Configuration

BMI160 ODR set to 200Hz via direct register write after I2cInit():

Wire.beginTransmission(0x68);
Wire.write(0x40); Wire.write(0x09);  // ACC_CONF: 200Hz
Wire.endTransmission();

Wire.beginTransmission(0x68);
Wire.write(0x42); Wire.write(0x09);  // GYR_CONF: 200Hz
Wire.endTransmission();

BMI160 init order matters: softReset() must be called before I2cInit(). Reversed order causes silent all-zero readings.

Serial/WiFi Output Format

DATA,<timestamp_ms>,<accelMag>,<gyroMag>,<state>
FALL,<timestamp_ms>

State values: 0=NORMAL, 1=FREEFALL, 2=IMPACT, 3=POST-IMPACT STILLNESS, 4=FALL CONFIRMED, 5=FALL ALERTED

Output goes to both USB Serial (printf) and WiFi TCP simultaneously.

Visualization

Dashboard (PyQtGraph)

cd visualization
python dashboard.py

Switch between serial and WiFi at the top of the file:

CONNECTION_MODE = "serial"   # or "wifi"
WIFI_HOST       = "192.168.x.x"   # ESP32-C3 IP printed on boot
WIFI_PORT       = 3333

The ESP32-C3 IP is printed on boot: [WiFi] Connect dashboard with: TCP <IP>:3333

Live dashboard showing a detected fall event: freefall, impact, then post-impact stillness

InfluxDB + Grafana

cd visualization
docker compose up -d

• InfluxDB: http://localhost:8086 (bucket: fallsensor, org: smartlab)
• Grafana: http://localhost:3000

ML Pipeline

See docs/ML_README.md for full training and quantization details.

Quick retrain after any model or dataset change:

cd tests/three_classes
venv_compat\Scripts\activate    # Python 3.10 required - TF not yet on 3.13+
python train_qat.py
python convert_to_header.py     # outputs firmware_idf/main/fall_model.h
cd ../../firmware_idf
idf.py build
idf.py -p COM7 flash monitor

Current Model Performance

Python Dependencies

pip install -r requirements.txt

ML training scripts require Python 3.10 in a separate venv (venv_compat) - TensorFlow does not support Python 3.13+.

Future Work

• PCB design: 2-layer board, BMI160 and ESP32-C3 Super Mini as soldered modules on carrier, discrete power section
• Enclosure: TPU 3D printed, waist/belt-clip form factor
• BLE alert path for phone app connectivity
• SIM/GPS module on existing 5V rail for cellular fallback
• BMI160 INT1/INT2 for hardware step detection and low-power motion-wake
• Collect real elderly fall simulation data for model fine-tuning

Troubleshooting

See docs/ERRORS_AND_FIXES.md

Author

Smart Ayodele ayosmart129@gmail.com