AI-on-the-Edge: ESP32 Temperature Anomaly Detection

A complete workflow for building, training, and deploying a lightweight LSTM Autoencoder anomaly detector for temperature data on the ESP32 microcontroller—without TensorFlow or TFLite. This project uses PyTorch, ONNX, and C++ for efficient, real-time anomaly detection on edge devices.

---

Table of Contents

1. Project Overview
2. Setup & Quickstart
3. File & Data Explanations
4. Interpreting Training Output
5. Automated Deployment Steps
6. How Model Weights & LSTM Logic Are Used
7. Contributing

---

Project Overview

• Goal: Detect abnormal temperature spikes on an ESP32 using a trained LSTM Autoencoder.
• Workflow:
  1. Generate or collect temperature data
  2. Train an LSTM Autoencoder in PyTorch
  3. Export the model to ONNX and convert weights to C header
  4. Extract normalization and threshold parameters to C header
  5. Deploy and run anomaly detection on ESP32 (Arduino/PlatformIO)

---

Setup & Quickstart

1. Clone & Install Dependencies

pip install -r requirements.txt

2. Generate Synthetic Data (or collect real data)

python generate_synthetic_data.py --minutes 1440 --anomalies 10 --outfile temperature_log.csv

• This creates a realistic temperature time-series with injected anomalies.

3. Train the Model

python train_lstm_autoencoder.py

• Trains the LSTM autoencoder and exports ONNX, scaler, and threshold.

4. Convert Model Weights for ESP32

python convert_to_header.py --onnx lstm_autoencoder.onnx --header lstm_autoencoder_weights.h

• Converts ONNX weights to a C header for ESP32 deployment.

5. Extract Normalization and Threshold Parameters

python extract_params_to_header.py

• Generates model_params.h with min/max normalization values and anomaly threshold for direct use in your ESP32 code.

6. Deploy to ESP32

• Use esp32_anomaly_detector.ino and include both lstm_autoencoder_weights.h and model_params.h in your Arduino/PlatformIO project.
• The sketch now automatically uses the correct normalization and threshold values.
• Upload to ESP32 and connect your DHT22 sensor.

---

One-Step Automated Deployment

For a fully automated setup, simply run:

python prepare_esp32_deployment.py

This script will execute all the above steps in sequence, generating all required files for ESP32 deployment in one go.

---

File & Data Explanations

File	Type/Format	Purpose/Role
generate_synthetic_data.py	Python script	Generate synthetic training data
temperature_log.csv	CSV	Training data for model
train_lstm_autoencoder.py	Python script	Train and export LSTM autoencoder
lstm_autoencoder.pth	PyTorch model	Trained model weights
scaler.save	Joblib binary	Fitted scaler for normalization
anomaly_threshold.npy	NumPy binary	Anomaly detection threshold
lstm_autoencoder.onnx	ONNX model	Model for conversion to C/C++
convert_to_header.py	Python script	Convert ONNX weights to C header
lstm_autoencoder_weights.h	C header	Model weights for ESP32 (autogenerated, safe to delete)
extract_params_to_header.py	Python script	Extracts scaler min/max and threshold to C header
model_params.h	C header	Normalization and threshold constants for ESP32 (autogenerated, safe to delete)
generate_lstm_cpp.py	Python script	Auto-generates C++ LSTM forward pass from ONNX
lstm_forward_pass.cpp	C++ source	(Optional) Standalone LSTM forward pass (autogenerated, safe to delete)
prepare_esp32_deployment.py	Python script	One-step deployment: runs all steps and generates all required files
data_collector.ino	Arduino sketch	Collect and log real sensor data
esp32_anomaly_detector.ino	Arduino sketch	Run anomaly detection on ESP32
requirements.txt	Text	Python dependencies

• Note: All autogenerated files (lstm_autoencoder_weights.h, model_params.h, lstm_forward_pass.cpp) can be safely deleted and will be recreated by the deployment script. Always back up headers if you want to preserve a specific version.
• The LSTM logic and weights are always up-to-date and automated—no manual editing required.
• The one-step deployment script (prepare_esp32_deployment.py) is available in both the Setup & Quickstart and the file table above.

---

Interpreting Training Output

• During training, you'll see output like:

  Epoch 1/30, Loss: 0.495228
  Epoch 2/30, Loss: 0.326717
  ...
  

• Loss: Measures how well the model reconstructs normal temperature patterns. Lower loss = better reconstruction.
• Decreasing loss: Indicates the model is learning.
• Threshold: The script prints and saves a suggested anomaly threshold. Use this value on the ESP32 to flag anomalies.

---

Automated Deployment Steps

1. Run all Python scripts in order:
   • generate_synthetic_data.py
   • train_lstm_autoencoder.py
   • convert_to_header.py
   • extract_params_to_header.py
2. Copy the generated headers (lstm_autoencoder_weights.h, model_params.h) to your ESP32 project.
3. Upload esp32_anomaly_detector.ino to your ESP32.
4. No manual copying of normalization or threshold values is needed!
5. The LSTM forward pass logic is now implemented directly in the sketch and uses the latest weights automatically.

Bonus: One-Step Deployment

For a fully automated setup, you can simply run:

python prepare_esp32_deployment.py

This script will execute all the above steps in sequence, generating all required files for ESP32 deployment in one go.

---

How Model Weights & LSTM Logic Are Used

• The deployment pipeline automatically extracts the latest model weights from ONNX and writes them to lstm_autoencoder_weights.h.
• The ESP32 sketch (esp32_anomaly_detector.ino) includes this header and uses the weights directly in the LSTM cell math.
• The LSTM forward pass logic is implemented in C++ in the sketch, so every time you retrain and regenerate weights, your ESP32 will use the new model without manual changes.
• Tip: If you want to keep a specific set of weights, back up the header before rerunning the deployment script.

---

Contributing

• Contributions, bug reports, and improvements are welcome!
• Please open an issue or submit a pull request.
• When contributing, ensure your code is robust (error handling, clear comments) and follows the modular structure of this project.

---

Project maintained by fellow developers and contributors.