This repository contains firmware and Python utilities for collecting "digit" writing gestures from a 6-DoF IMU (LSM6DSOX) attached to an ESP32, extracting features, training simple classifiers (kNN/SVM) and running real-time inference over serial.

Contents

• ml_digits.ino — ESP32 logger that streams CSV rows for every IMU sample.
• write.py — Python helper to collect a single labeled writing segment and save it to imu_digits/.
• train_svm_digits.py / train_knn_digits.py — training scripts (feature extraction + model training).
• infer.py — serial realtime inference using a trained knn_digit_cls.joblib model.
• collect folders (imu_digits/, 4_seconds/, etc.) hold recorded CSVs.
• morse.ino, firmware.ino — other firmware examples in the repo.

Quickstart

1. Install Python deps:

   pip install numpy pandas scikit-learn joblib scipy pyserial

2. Collect a sample (on Windows):

   • Set PORT in write.py (e.g. COM3) and run python write.py.
   • Type a digit label (0–9), then write the digit on the device; the script will save a CSV to imu_digits/.

3. Train a classifier:

   • Edit the training script to point at your data directory (look for DATA_DIR).
   • Run the script (e.g. python train_svm_digits.py or the KNN script) to produce a .joblib model file.

4. Real-time inference:

   • Ensure the model (knn_digit_cls.joblib) is present in the repo root.
   • Edit PORT in infer.py and run python infer.py to stream from the board and see predictions.

Security & secrets

• morse.ino contains placeholders for Wi‑Fi and Telegram tokens. Do NOT commit real secrets into the repo. Instead use a local, untracked header or build-time configuration to inject them.

Notes and suggestions

• CSV format: t_ms,ax,ay,az,gx,gy,gz,state,label — state marks writing vs rest.
• If your CSVs contain formatting issues (e.g. - characters as placeholders), use pandas.to_numeric(..., errors='coerce') and .fillna() before numeric conversion.
• This repository is intentionally lightweight and designed for experimentation. Consider adding unit tests for feature extraction and a small example dataset to make training reproducible.