README: EMG-Based Prosthetic Hand Using Neural Network

This repository contains the implementation of an intelligent, multi-channel Electromyography (EMG) controlled prosthetic hand. The system utilizes a Convolutional Neural Network (CNN) to classify muscle activity into nine distinct hand gestures, which are then executed in real-time by a 3D-printed prosthetic mechanism.

---

Project Overview

The goal of this project is to develop a cost-effective and scalable assistive device for individuals with upper-limb loss. By bridging the gap between biomedical signal processing and deep learning, the system provides a more natural and intuitive control interface compared to traditional threshold-based myoelectric limbs.

Key Features:

• Multi-Channel Sensing: Uses five Muscle BioAmp Patchy sensors for detailed forearm muscle activity acquisition.
• Deep Learning Classification: A CNN model identifies 9 gestures with approximately 80-85% accuracy.
• Real-Time Actuation: Low-latency control of 5 high-torque MG996R servo motors.
• Tendon-Driven Design: 3D-printed (PLA) hand using the InMoov modular design for human-like motion.

---

Hardware Requirements

Component	Specification/Function
Microcontroller	Arduino Nano (ATmega328P)
EMG Sensors	5x Muscle BioAmp Patchy
Actuators	5x MG996R High-Torque Servos
Prosthetic Hand	3D-Printed PLA (Tendon-driven)
Power Supply	7.4V Li-ion battery (for servos) + 5V USB (for logic)

---

Software Pipeline

The software architecture is divided into three primary stages:

1. Signal Acquisition & Preprocessing:

   • Analog signals are amplified and filtered (74.5 Hz to 149.5 Hz) via hardware.
   • Digital smoothing is performed using a Moving Average Filter.
   • Data is segmented using a sliding window (50 samples, 50% overlap).

2. Feature Extraction: Six time-domain features are computed for each channel to represent muscle patterns:

   • Root Mean Square (RMS)
   • Mean Absolute Value (MAV)
   • Waveform Length (WL)
   • Zero Crossing (ZC)
   • Slope Sign Change (SSC)
   • Variance (VAR)

3. CNN Architecture:

   • Input: Structured $5 \times 6 \times 1$ feature matrix.
   • Layers: Two Conv2D layers (32/64 filters), Batch Normalization, Max Pooling, and two Dense layers (128/64 units).
   • Output: Softmax layer for 9 gesture classes.

---

Supported Gestures

The system is trained to recognize and actuate the following:

• Open Hand / Close Hand
• Individual Finger Movements (Thumb, Index, Middle, Ring, Little)
• Combined Gestures: Victory Sign, Thumbs Up

---

Performance

• Generic Model: ~80% accuracy (generalized across 10 users).
• Custom Model: ~85% accuracy (user-specific calibration).
• System Stability: Employs confidence-based filtering (threshold = 0.7) to prevent unintended activations.

---

Repository Structure

• /arduino: Contains firmware for EMG data streaming and servo control.
• /python: Scripts for data collection, feature extraction, and CNN training.
• /models: Pre-trained .h5 model files and scaler.pkl.

---

How to Use

1. Hardware Setup: Connect sensors to A0-A4 and servos to D3, D5, D6, D9, D10.
2. Data Collection: Run the Python data collection script while performing gestures to build a custom dataset.
3. Training: Train the model using the provided CNN architecture script.
4. Inference: Run the live classification script to control the prosthetic hand in real-time.

---