Prototype 4

This is intended as a human input device (like a keyboard) that predicts your intention from muscle signals. Using a neural network, trained through a process where you press keys at the exact same time that you make movements with the limb wearing the device, it learns correlations between the muscle signals and intended keys, and can press them for you.

Electrode map

Datasheet

Features:
- Battery-powered
- Wireless, using Bluetooth Low Energy (BLE)
- Supports 17 electrodes (8 pairs + 1 ground)
- Transmits 8 signals at 8-bit resolution, 500Hz sampling rate
- Linux graphical user interface for:
  - Mapping keyboard key presses to muscle signals
  - Training a neural network to predict key presses from signals
  - Simulating key presses based on neural network predictions
- GNURadio integration for plotting the raw signals + FFT
Hardware overview:
- A fabric glove/sleeve with a central component (arduino + power supply) and 8 distributed signal clean-up units close to individual electrodes.
- Circuit Schematic
- PCB of one distributed signal clean-up unit
Software dependencies:
- Python3 (see requirements.txt)
  - BLE-GATT
  - matplotlib
  - pynput
  - tensorflow
- ArduinoIDE
  - ArduinoBLE
  - Arduino LSM9DS1 (for accelerometer/gyroscope)
Components:
- 1x Arduino Nano 33 BLE Sense
- 1x LM324 Operational Amplifier
- 8x INA128 Instrumentational Amplifier, SOIC package
- Any portable 6V power supply (like 2x CR2032)
- 17x Electrodes, e.g. Prym 390120 (don't laugh, they work)
- A forearm sleeve from non-conductive fabric
- Additional standard components (see schematic)
Power ratings
- Supply Voltage: 4.5-6V
- Power dissipation:
  - Idle: ~86.9mW (16.9mA x 5.14V)
  - Transmitting at 6-7kB/s: 92.5mW (18.0mA x 5.14V)
Weight: ~85g (at 9.81m/s² gravitational acceleration)

If you require assembly instructions or further information, please contact the author.

See blog post Wireless Prototype
Components:

Video Demo: here
Date: 2021-05-12