🤖 Edge AI Flow — Nano 33 BLE Sense Rev2
－
⊙
＋
🤖 Edge AI Data Flow — TinyML Pipeline on Arduino Nano 33 BLE Sense Rev2
① Sensors
② Filtering
③ Feature Extraction
④a Training Path (PC / Cloud / Edge Impulse)
④b Inference Path (on nRF52840 + TFLM)
🔄
9-axis IMU
BMI270+BMM150 · SPI
🎤
PDM Microphone
MP34DT06JTR · 16 kHz
🌡️
Env. Sensors
HS3003 · BMP390 · I²C
👁️
Light / Color / Gesture
APDS-9960 · I²C
🔧
Low-pass / HPF
LPF / HPF · Kalman
⚖️
Normalization
min-max · z-score
🪟
Windowing
1s @ 100 Hz = 100 smp
⏱️
Time-domain Feats
mean · RMS · ZCR · peak
🎵
Frequency Features
FFT bins · MFCC (40×49)
📊
Statistical / Spatial
magnitude · variance · grad
🔀
Split
🗄️
Logged Dataset
CSV · Edge Impulse
🧠
Model Training
TF · Keras · Edge Impulse
✅
Validation
confusion · F1 · ROC
📦
Quantize → INT8
TFLite PTQ converter
💾
C-array Export
xxd -i → model.h
⚙️
TFLite-Micro Runtime
nRF52840 · CMSIS-NN
🏷️
Classification
softmax · argmax
🧩
Decision Logic
class + confidence
🔔
Output / Actuation
RGB LED · BLE · Serial
capture (cloud)
live (on-MCU)
💾 flash model.h (Arduino IDE / OTA-DFU)
📡 Serial/BLE → CSV capture
⚡ TinyML Scenarios
🤚 S1 — Gesture Recognition (IMU)
IMU @ 100 Hz → 1 s windows → mean / RMS / FFT per axis
→ small CNN ("punch" / "flex" / "idle") → INT8 → flash → BLE notify.
~20–40 KB Flash · <30 ms · 🛠 Arduino_BMI270_BMM150 + Arduino_TensorFlowLite
🎤 S2 — Keyword Spotting (PDM mic)
16 kHz audio → 1 s window → MFCC (40 × 49 frames)
→ tiny CNN ("yes" / "no" / "noise") → quantize → 🟢 RGB LED.
~100 KB Flash · ~200 ms · 🛠 PDM + Arduino_TensorFlowLite
⚡ S3 — Vibration Anomaly Detection (IMU on a motor)
IMU → LPF → 256-sample windows → FFT → autoencoder (healthy data only)
→ reconstruction error on edge → if > threshold → 🔴 LED + alert.
Unsupervised · only normal data needed · great for predictive maintenance
🌿 S4 — Environmental / HVAC Classification
HS3003 + BMP390 + APDS-9960 → 10-sample window of T/H/P + light
→ stats features → ("normal" / "open window" / "occupied") → 📱 BLE.
Sub-10 KB Flash · BLE · 🛠 Arduino_HS300x + Arduino_APDS9960 + ArduinoBLE
🔌 Sensor
🔧 DSP / Feature
🔀 Split
🧠 Training (cloud)
⚙️ Inference (on-MCU)
data flow
flash / capture
nRF52840: 1MB Flash · 256KB RAM · Cortex-M4F @ 64 MHz · model budget: 20–200 KB (INT8)