ptrajdos / noise_tollerant_semg_ensemble

A noise-tolerant dual multiclassifier system with fuzzy model applied to the sEMG-based control a bionic upper limb prosthesis

System Requirements

Requires: Python>=3.9.7 All required packages may be installed via pip. Tested on: Ubuntu Linux 22.04, macOS Sonoma 14.5

Setup

To download test data, create the virtual envirionment, and install required packages type:

make create_env

To clean virtual environment type:

make clean

The experimental results will be located in: ./experiments_results

Experiments

To run all experiments type:

make run

One class classifiers

To run the experiments on one-class classifiers type:

make run_one_class

Results will be placed in: ./experiments_results/outlier_detection_experiment_snr2. Directory structure:

• Single set
  • A[1-9]_Force_Exp_low_windowed.pickle -- raw results (for a single set) as numpy arrays
  • A[1-9]_Force_Exp_low_windowed.pdf -- raw results (for a single set) as boxplots. It include results for all quality criteria, all classifiers, all SNR levels, and signal spoilers.
  • A[1-9]_Force_Exp_low_windowed.md -- raw results (for a single set) in tabular form (average and standard deviation). It include results for all quality criteria, all classifiers, all SNR levels.
  • A[1-9]_Force_Exp_low_windowed_trends.pdf -- trends in quality criteria (median, Q1, Q3) over all SNR levels. Results for all quality criteria, signal spoilers, and base one-class classifiers.
• Ranking over all sets
  • ALL_trends_ranks.pdf -- average ranks plots for all SNR levels.
  • ALL_trends_ranks.md -- average ranks and statistical tests in tabular form.

Ensemble classifiers

To run experiments on commitees with outlier detection type:

make run_commitees

Results will be placed in: ./experiments_results/results_channel_combination_ensemble_fast_full. Directory structure:

• Single set:
  • A[1-9]_Force_Exp_low_windowed.pickle -- raw results (for a single set) as numpy arrays.
  • A[1-9]_Force_Exp_low_windowed_snr_m1.pdf -- boxplots (for a single set) for different SNR values, criteria, ensemble sizes, number of contaminated channels.
  • A[1-9]_Force_Exp_low_windowed_snr_m2.pdf -- boxplots (for a single set) for different SNR values, criteria, ensemble sizes. Averaged over the number of contaminated channels.
  • A[1-9]_Force_Exp_low_windowed_noise_gs.pdf -- boxplots (for a single set) for different SNR values, and quality criteria. The impact of changing number of channels included in the committee.
  • A[1-9]_Force_Exp_low_windowed_noise_gs.pdf -- boxplots (for a single set) for SNR=6, and quality criteria. The impact of changing number of channels included in the committee.
• Rankings over all sets:
  • *ALL_snr_m1_ranks.pdf" -- Average ranks plots for different SNR values, criteria, ensemble sizes.
  • *ALL_snr_m1_ranks.md" -- Average ranks tables and statistical tests for different SNR values, criteria, ensemble sizes.
  • *ALL_snr_m2_ranks.pdf" -- Average ranks plots for different SNR values, criteria, ensemble sizes. Averaged over the number of contaminated channels.
  • *ALL_snr_m2_ranks.md" -- Average ranks tables and statistical tests for different SNR values, criteria, ensemble sizes. Averaged over the number of contaminated channels.
  • *ALL_noise_gs_ranks.pdf" -- Investigation of the impact of the K parameter -- classical ranks.
  • *ALL_noise_gs_ranks.md" -- Investigation of the impact of the K parameter -- classical ranks. Tabular version with statistical tests
  • *ALL_noise_gs_ranks_alt.pdf" -- Investigation of the impact of the K parameter -- Different values of K are ranked.
  • *ALL_noise_gs_ranks_alt.md" -- Investigation of the impact of the K parameter -- Different values of K are ranked. Tabular version with statistical tests

Ensemble classifiers

To run experiments for comparison with reference methods type:

make run_reference

Results will be placed in: ./experiments_results/results_channel_combination_ensemble_full2. Directory structure:

• Single set:
  • A[1-9]_Force_Exp_low_windowed.pickle -- raw results (for a single set) as numpy arrays.
  • A[1-9]_Force_Exp_low_windowed_snr_m1.pdf -- boxplots (for a single set) for different SNR values, criteria, ensemble sizes, number of contaminated channels.
  • A[1-9]_Force_Exp_low_windowed_snr_m2.pdf -- boxplots (for a single set) for different SNR values, criteria, ensemble sizes. Averaged over the number of contaminated channels.
  • A[1-9]_Force_Exp_low_windowed_noise_gs.pdf -- boxplots (for a single set) for different SNR values, and quality criteria. The impact of changing number of channels included in the committee.
  • A[1-9]_Force_Exp_low_windowed_noise_gs.pdf -- boxplots (for a single set) for SNR=6, and quality criteria. The impact of changing number of channels included in the committee.
• Rankings over all sets:
  • *ALL_snr_m1_ranks.pdf" -- Average ranks plots for different SNR values, criteria, ensemble sizes.
  • *ALL_snr_m1_ranks.md" -- Average ranks tables and statistical tests for different SNR values, criteria, ensemble sizes.
  • *ALL_snr_m2_ranks.pdf" -- Average ranks plots for different SNR values, criteria, ensemble sizes. Averaged over the number of contaminated channels.
  • *ALL_snr_m2_ranks.md" -- Average ranks tables and statistical tests for different SNR values, criteria, ensemble sizes. Averaged over the number of contaminated channels.
  • *ALL_noise_gs_ranks.pdf" -- Investigation of the impact of the K parameter -- classical ranks.
  • *ALL_noise_gs_ranks.md" -- Investigation of the impact of the K parameter -- classical ranks. Tabular version with statistical tests
  • *ALL_noise_gs_ranks_alt.pdf" -- Investigation of the impact of the K parameter -- Different values of K are ranked.
  • *ALL_noise_gs_ranks_alt.md" -- Investigation of the impact of the K parameter -- Different values of K are ranked. Tabular version with statistical tests