Sword Strike Classifier

Real-time gesture classification of sword strike types using IMU data, deployed on an Arduino Nano 33 BLE Sense. Built with Edge Impulse for data pipeline management, CNN-based model training, and TFLite deployment.

---

Overview

This project implements an end-to-end TinyML pipeline for classifying distinct sword strike motions (e.g., overhead, lateral, diagonal, idle) from inertial measurement unit (IMU) data. The trained model runs entirely on-device via TFLite Micro, with live inference results streamed over BLE to a browser-based web interface — no cloud dependency at inference time.

Edge Impulse project: https://studio.edgeimpulse.com/public/1014843/live

---

Repository Structure

sword_strike_classifier/
├── data_collect/
│   └── Arduino firmware for collecting IMU data over BLE
│
├── live_inference/
│   └── Arduino firmware for real-time sword strike classification
│
├── logger.html
│   └── Web interface for data collection and labeling
│
├── inference.html
│   └── Live dashboard for visualizing sensor streams and model predictions
│
└── README.md

---

Pipeline

1. Data Collection

data_collect/ contains an Arduino sketch that streams raw 6-axis IMU data (accelerometer + gyroscope) from the Nano 33 BLE Sense over BLE Serial. logger.html connects to the board in-browser, lets you label gesture windows in real time, and exports CSV files for upload to Edge Impulse.

2. Model Training (Edge Impulse)

The full training pipeline lives in the public Edge Impulse project linked above. Key configuration:

• Input block: fixed-length time-series windows of raw acc/gyro samples
• Processing block: spectral features (FFT-based frequency domain representation)
• Learning block: DNN trained on labeled strike classes
• Data augmentation: noise injection and axis jitter to improve generalization across users and grip styles
• Export: INT8 quantized TFLite model compiled into a C header for deployment

3. On-Device Inference

live_inference/ runs the exported TFLite Micro model directly on the nRF52840 MCU. Inference results are transmitted over BLE and displayed in real time through inference.html in the browser — no USB connection required during use.

---

Hardware

Component	Detail
Board	Arduino Nano 33 BLE Sense
MCU	Nordic nRF52840 (ARM Cortex-M4F, 64 MHz)
IMU	LSM9DS1 — 3-axis accelerometer + 3-axis gyroscope
SRAM	256 KB
Flash	1 MB
Connectivity	Bluetooth Low Energy (BLE)

---

Dependencies

Arduino libraries:

• Arduino_LSM9DS1
• ArduinoBLE
• Arduino_TensorFlowLite (TFLite Micro)

Toolchain:

• Edge Impulse Studio (model training, DSP pipeline)
• Edge Impulse CLI (optional, for local dataset management)
• Arduino IDE 2.x with the Mbed OS Nano board package

---

Usage

Data collection:

1. Flash data_collect/data_collect.ino to the Nano 33 BLE Sense
2. Open logger.html in a Web Bluetooth-compatible browser (Chrome recommended)
3. Connect to the board, select a gesture label, and record strike windows
4. Export and upload CSV data to Edge Impulse

Live inference:

1. Train and export the model from Edge Impulse as an Arduino library
2. Place the exported library in live_inference/ and flash sword_inference.ino
3. Open inference.html in browser, connect over BLE
4. Perform strikes — predicted class and confidence appear in real time

---

Context

Developed for EE 446: TinyML at the University of Washington. Demonstrates the full embedded ML workflow from IMU data collection through DNN training, INT8 quantization, TFLite Micro deployment, and BLE-based inference streaming on a resource-constrained microcontroller.