TinyML on ESP32

What is TinyML?

TinyML brings machine learning to microcontrollers with <1MB RAM. Run inference in milliseconds with milliwatt power consumption.

Setup & Installation

# Install TensorFlow Lite for Microcontrollers
# In Arduino IDE: Install "TensorFlowLite" library

# Or PlatformIO
[env:esp32dev]
platform = espressif32
board = esp32dev
lib_deps = 
  tensorflow/tensorflow@^2.10.0
  arduino-libraries/Arduino TensorFlowLite

Gesture Recognition

#include 
#include < MPU6050.h>
#include "gesture_model.h"

MPU6050 mpu;
tflite::MicroInterpreter* interpreter;

void setup() {
  mpu.initialize();
  
  // Load model
  static tflite::MicroModel model = 
    tflite::GetModel(gesture_model_tflite);
  
  // Create interpreter
  static tflite::MicroInterpreter interp(model, resolver, tensor_arena, kTensorArenaSize);
  interpreter = &interp;
  interpreter->AllocateTensors();
}

void loop() {
  // Read accelerometer
  int16_t ax, ay, az;
  mpu.getAcceleration(&ax, &ay, &az);
  
  // Run inference
  TfLiteTensor* input = interpreter->input(0);
  input->data.f[0] = ax;
  input->data.f[1] = ay;
  input->data.f[2] = az;
  
  interpreter->Invoke();
  
  // Get prediction
  TfLiteTensor* output = interpreter->output(0);
  int gesture = argmax(output->data.f);
}

Anomaly Detection

Detect equipment failures from vibration patterns:

// Autoencoder for anomaly detection
float reconstruction_error = calculate_error(input_data);

if (reconstruction_error > THRESHOLD) {
  // Anomaly detected!
  sendAlert("MACHINE_ANOMALY");
}

Model Training

# Python training script
import tensorflow as tf

# Load sensor data
X_train, y_train = load_data()

# Create model
model = tf.keras.Sequential([
  tf.keras.layers.Dense(32, activation='relu'),
  tf.keras.layers.Dense(16, activation='relu'),
  tf.keras.layers.Dense(3, activation='softmax')
])

# Convert to TFLite
converter = tf.lite.TFLiteConverter.from_keras_model(model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
tflite_model = converter.convert()

# Save for ESP32
with open('gesture_model.tflite', 'wb') as f:
  f.write(tflite_model)

Model Optimization

Technique	Size Reduction	Accuracy Impact
Post-training quantization	4x	Minimal
Pruning	2-10x	Low
Knowledge distillation	10x	Minimal