ESP32 Neural Network Offloading with MQTT

This project demonstrates offloading neural network computations from an ESP32 device to an edge server using MQTT. The ESP32 device connects to a Wi-Fi network, registers itself on the edge server, and communicates through MQTT topics to perform neural network inference layer by layer.

Features

• Wi-Fi Connectivity: The ESP32 connects to a specified Wi-Fi network.
• MQTT Communication: The ESP32 uses MQTT to send and receive messages from an edge server.
• Neural Network Offloading: Inference layers are run on the ESP32, and results are sent to the edge server.
• TensorFlow Lite Micro: The project utilizes TensorFlow Lite Micro for running machine learning models on the ESP32.
• NTP Time Synchronization: The device syncs with an NTP server for accurate timestamps.
• UUID Generation: Each message and device is uniquely identified using UUIDs.

Dependencies

• ESP32 Board Support Package (via Arduino IDE)
• Libraries:
  • WiFi.h: For Wi-Fi connectivity.
  • PubSubClient.h: For MQTT communication.
  • ArduinoJson.h: For handling JSON data.
  • tensorflow/lite/micro: For running TensorFlow Lite models on microcontrollers.
  • UUID.h: For generating UUIDs.

Configuration

Edit the conf.h file to set up the following configurations:

• Wi-Fi Credentials: SSID and PWD
• MQTT Broker: MQTT_SRV, MQTT_PORT
• NTP Server: NTP_SRV, NTP_GMT_OFFSET, NTP_DAYLIGHT_OFFSET

Model Setup

The project supports up to 5 layers of a TensorFlow Lite model, which are loaded and run sequentially. Each layer is defined in its own header file (e.g., layer_0.h, layer_1.h).

Ensure that the TensorFlow Lite model layers are properly converted and placed in the model_layers directory.

Topics

The following MQTT topics are used for communication:

• devices/: Device registration.
• DeviceUUID/model_data: Topic to send model data for inference.
• DeviceUUID/model_inference: Topic to trigger inference for a specific layer.
• DeviceUUID/model_inference_result: Topic to receive inference results.

Running the Project

1. Flash the code to the ESP32.
2. Ensure that the MQTT broker is running and accessible.
3. The device will:
   • Connect to Wi-Fi.
   • Register itself on the edge server.
   • Subscribe to necessary MQTT topics.
   • Perform neural network inference layer by layer upon receiving input data and an offloading layer index.

Example Workflow

1. Device Registration: The ESP32 registers itself to the edge server by publishing a registration message.
2. Receive Input Data: The device receives input data for the neural network through the model_data topic.
3. Run Inference: The device runs inference for each layer of the neural network and publishes the results to the model_inference_result topic.
4. Publish Results: Inference results, including the output and inference time for each layer, are published to the edge server.

Notes

• The size of the MQTT packet is set to 3 KB (MQTT_MAX_PACKET_SIZE).
• The device restarts automatically after completing the test, ensuring fresh initialization for the next session.
• Ensure that the edge server is properly configured to handle incoming MQTT messages and process the neural network inference results.

License

This project is licensed under