EMGFingerClassification

Step 1: Recording

1. Data Recording:
   • Data is recorded for all 5 fingers, each recorded 5 times.
   • Two recording sessions:
     • First session: 4-second gap between recordings.
     • Second session: 2-second gap between recordings.
2. Filtering:
   • Bandstop filter: 45 to 55Hz.
   • Bandpass filter: 1 to 400Hz.
3. Data Storage:
   • Recorded raw data is saved in CSV files.

Step 2: Processing and Labelling of Data

1. Data Splitting:
   • Split data into test and train datasets.
   • Test dataset: 4-second interval.
   • Train dataset: 2-second interval.
2. Data Smoothing:
   • Applied envelope on the dataset for smoothing.
   • 
3. Peak Detection:
   • Manual adjustment of prominence, distance, and height parameters to detect all peaks.
   • 

Step 3: Labelling Dataset

• Labels: Thumb, Index, Middle, Ring, Pinky.
• Window size: 500 for each finger.
• Baseline: 100ms window between two peak centers with a width of 100ms.

Step 4: Visualising Various Channels According to Our Labels

• 

Step 5: Created Datasets with Time Features Rather than Channel Data

• Time features considered:
  • MAV (Mean Absolute Value)
  • VAR (Variance)
  • WL (Waveform Length)
  • RMS (Root Mean Square)

Step 6: Exported CSV Files

• CSV files for 4-second, 2-second, and combined data intervals are exported.

Step 7: Training Machine Learning Models

• Open TrainMLModels.ipynb.
• Run the initial cells based on your requirement:
  • Combined data.
  • Data of an individual subject.

Step 8: Metrics and Model Evaluation

• Defined metrics for the confusion matrix.
• Implemented three ML models:
  • SVM (Support Vector Machine)
  • Random Forest
  • CNN (Convolutional Neural Network)
• 

Tips for Improving Accuracy

1. Change Window Size: Experiment with different window sizes for feature extraction.
2. Change Number of Epochs for CNN: Adjust the number of training epochs for the Convolutional Neural Network.
3. Add More Time Features: Incorporate additional time-domain features for better representation of the data.
4. Frequency Domain Analysis: Perform frequency domain analysis to extract frequency-based features.
5. Change Sampling Rate: Experiment with different sampling rates during data recording to capture more detailed signal information.

Additional Information

• Recording Software: Provided in this repository.
• Processing and Labelling: Performed in LabellingDataset.ipynb.

Ensure all paths to images and GIFs are correctly set in the markdown. Adjust parameters and model training steps as necessary for your specific dataset and use case.

Credits

• This project is supported by Upside Down Labs, who provided valuable guidance and the necessary hardware.