PaperlayerView exact revision on arXiv

Source-linked AI summary

A Novel Feature Extraction for Robust EMG Pattern Recognition

Angkoon Phinyomark, Chusak Limsakul, Pornchai Phukpattaranont

arXiv:0912.3973v2cs.CV

TL;DR

Random WGN is difficult to remove from sEMG without potentially losing signal content, motivating robust feature extraction for EMG recognition. The paper introduces MMNF and MMDF by modifying frequency calculations, evaluates them against sixteen existing features under varied SNRs, and reports MMNF as the strongest robustness candidate, including in weak signals.

  • Problem

    Random WGN is difficult to remove from sEMG with existing filtering and denoising methods, although noise removal is important before EMG feature extraction and classification.

  • Method

    The paper evaluates sixteen time- and frequency-domain features under additive WGN at varied SNRs and introduces MMNF and MMDF using amplitude-spectrum statistics.

  • Results

    MMNF had the best robustness among evaluated features, with average error of 6% for strong and 10% for weak sEMG at 0 dB SNR; MMNF, WAMP, and HEMG also formed a stronger noisy-environment feature group than comparator groups.

  • Takeaways & Limitations

    MMNF is reported as a robust feature candidate for EMG pattern recognition, and MMNF combined with WAMP and HEMG is reported as a useful multi-feature vector in noisy environments.

  • Takeaways & Limitations

    The Willison amplitude threshold is selected within a setting-dependent range, with optimal robustness evaluated for the instrument gain used.

Abstract

from arXiv · show

Varieties of noises are major problem in recognition of Electromyography (EMG) signal. Hence, methods to remove noise become most significant in EMG signal analysis. White Gaussian noise (WGN) is used to represent interference in this paper. Generally, WGN is difficult to be removed using typical filtering and solutions to remove WGN are limited. In addition, noise removal is an important step before performing feature extraction, which is used in EMG-based recognition. This research is aimed to present a novel feature that tolerate with WGN. As a result, noise removal algorithm is not needed. Two novel mean and median frequencies (MMNF and MMDF) are presented for robust feature extraction. Sixteen existing features and two novelties are evaluated in a noisy environment. WGN with various signal-to-noise ratios (SNRs), i.e. 20-0 dB, was added to the original EMG signal. The results showed that MMNF performed very well especially in weak EMG signal compared with others. The error of MMNF in weak EMG signal with very high noise, 0 dB SNR, is about 5-10 percent and closed by MMDF and Histogram, whereas the error of other features is more than 20 percent. While in strong EMG signal, the error of MMNF is better than those from other features. Moreover, the combination of MMNF, Histrogram of EMG and Willison amplitude is used as feature vector in classification task. The experimental result shows the better recognition result in noisy environment than other success feature candidates. From the above results demonstrate that MMNF can be used for new robust feature extraction.

1 INTRODUCTION

The paper addresses the difficulty of recognizing sEMG patterns under random noise, especially WGN, by evaluating robust feature extraction as an alternative to noise removal. It introduces MMNF and MMDF and tests them against existing features across noisy conditions.

  • Motivation: Random noise overlapping EMG energy is difficult to remove with typical filtering, and denoising may also remove important sEMG components.WGN has broad-band, random frequency characteristics, and its amplitude can exceed the sEMG signal amplitude of about 50 µV-100 mV.
  • Motivation: Feature extraction is essential before classification because it highlights relevant sEMG structures while rejecting noise and unimportant signal components.The study therefore seeks features that tolerate WGN so preprocessing noise-removal algorithms are unnecessary.
  • Related work: Prior work evaluated EMG features mainly from class-separability perspectives, while robust EMG feature selection and modification remained insufficiently addressed.The paper also notes that earlier evaluated feature sets were not fully aligned with methods currently available.
  • Study design: Sixteen literature-based time- and frequency-domain features are tested with additive WGN at varied SNRs, alongside evaluation of signal-amplitude effects.The study considers these domains because of their computational simplicity and widespread research use.
  • Contribution: MMNF and MMDF modify mean and median frequency calculations by taking the mean and median of the amplitude spectrum instead of the power spectrum.These two calculations are presented as novel frequency-domain features for robust extraction.

2 EXPERIMENTS AND DATA ACQUISITION

The experiments record sEMG from forearm muscles under controlled acquisition settings and construct datasets for feature-error and noisy-classification evaluations. Motion, signal-strength, electrode-placement, and timing conditions are specified.

  • Signal acquisition: sEMG was recorded from flexor carpi radialis and extensor carpi radialis longus using two electrode pairs on the right forearm.Electrodes were separated by 2 cm; acquisition used a 10-500 Hz band-pass filter, 60 dB amplification, and 1 kHz sampling with a 16-bit converter.
  • First dataset: Four upper-limb motions were recorded, and signals were divided into strong and weak sEMG categories to test signal-strength effects.Strong signals came from specified muscle-motion combinations, while other motions and electrode channels were treated as weak signals.
  • First dataset: Ten datasets were collected for each motion, using 256 ms signal segments to satisfy a response-time constraint below 300 ms.This dataset supported the MSE criterion for measuring noise effects on feature values.
  • Classification dataset: A second dataset evaluated classification using seven upper-limb motions, resting, and eight right-forearm electrode positions.The dataset was acquired by Carleton University in Canada and was used to assess feature-space quality with WGN.

3 METHODOLOGY

The methodology organizes EMG features into time- and frequency-domain groups and evaluates their suitability for pattern recognition. Time-domain features provide computational simplicity, while frequency-domain features represent fatigue, neural abnormalities, and recognition information.

  • Feature domains: Eighteen features are described across time and frequency domains, including thirteen time-domain variables measured from signal segments.The listed time-domain group includes IEMG, MAV, modified MAV variants, RMS, and other features.
  • Time domain: Time-domain features are widely used because of computational simplicity and support onset, muscle-contraction, and muscle-activity detection.The paper characterizes these features as popular linear techniques in EMG pattern recognition.
  • Frequency domain: Frequency-domain features represent muscle fatigue and neural abnormalities and can also be used in EMG pattern recognition.The frequency-spectrum methods include autoregressive coefficients, mean and median frequencies, and their modified forms.

3.1 Time Domain Feature Extraction

The time-domain feature section defines amplitude-, slope-, and threshold-based descriptors for sEMG segments, including weighted modifications of mean absolute value. These features support activity detection and robustness evaluation.

  • Integrated EMG: Integrated EMG sums the absolute sEMG amplitudes over a signal segment.The feature is used as an onset index for detecting muscle activity preceding assistive-device control commands.
  • Mean Absolute Value: Mean Absolute Value averages the absolute values of the sEMG signal and is used to detect muscle contraction levels.It is described as a popular feature for myoelectric control applications.
  • Modified MAV features: Modified Mean Absolute Value 1 extends MAV with a weighting window function.The weighting function changes the contribution of samples within the signal segment.
  • Modified MAV features: Modified Mean Absolute Value 2 improves the MMAV1 window by using a continuous weighting function.Its calculation applies the smooth window to the weighted signal amplitudes.

3.1.5 Mean Absolute Value Slope

Mean Absolute Value Slope (MAVSLP) is introduced as a modified MAV feature based on differences between adjacent-segment MAVs. The supplied passages also contextualize related amplitude-based features and their comparison with RMS.

  • MAVSLP modifies MAV by determining differences between MAVs of adjacent signal segments.
  • Simple Square Integral uses sEMG signal energy as a feature.
  • Variance of EMG uses sEMG signal power and can exploit the near-zero mean of the EMG signal.
  • Root Mean Square relates to standard deviation and is modeled for amplitude-modulated Gaussian processes.
  • Prior comparisons report MAV processing as equal to or better than RMS processing, while RMS retains more power-property information.

3.1.9 Waveform Length

Waveform Length (WL) measures cumulative waveform length over a time segment and relates to amplitude, frequency, and time. The supplied material also introduces threshold-based frequency features and groups several amplitude-derived features by similar experimental patterns.

  • 3.1.9 Waveform Length: Waveform Length is the cumulative length of the waveform over a time segment, related to amplitude, frequency, and time.
  • 3.1.9 Waveform Length: Amplitude-based features in Sections 3.1.1–3.1.9 show similar experimental patterns, motivating selection of one robust representative.
  • Zero Crossing counts threshold-qualified crossings of the sEMG amplitude through the zero axis to approximate frequency-domain properties.
  • Slope Sign Change counts positive-to-negative slope changes across three consecutive segments using a threshold to avoid interference.

3.1.12 Willison Amplitude

Willison Amplitude (WAMP) counts threshold-exceeding amplitude differences between adjacent sEMG segments to reduce noise effects. The section also relates WAMP to motor-unit firing and muscle contraction, with threshold selection dependent on instrument gain.

  • 3.1.12 Willison Amplitude: Willison Amplitude counts adjacent-segment sEMG amplitude differences that exceed a predefined threshold to reduce noise effects.
  • 3.1.12 Willison Amplitude: WAMP is related to motor unit action-potential firing and muscle contraction level.
  • Thresholds for ZC, SSC, and WAMP are normally selected between 10 and 100 mV, depending on instrument gain.
  • Histogram of EMG divides sEMG elements into b equally spaced segments and returns the count in each segment.

3.2 Frequency Domain Feature Extraction

The frequency-domain features include AR coefficients, modified median and mean frequencies based on the amplitude spectrum, and traditional median and mean frequencies based on power spectrum. Because amplitude-spectrum variation is lower, the modified frequency features also vary less than their traditional counterparts.

  • Autoregressive features: An autoregressive model represents each sEMG sample as a linear combination of previous samples plus white-noise error, using AR coefficients as features.
  • Autoregressive features: The study evaluates AR orders from first through tenth order, extending beyond the previously suggested fourth-order model.
  • Modified frequency features: Modified Median Frequency divides the amplitude spectrum into two regions with equal amplitude.
  • Modified frequency features: Modified Mean Frequency averages frequency using amplitude-spectrum weights, dividing the weighted frequency sum by total spectrum intensity.
  • Traditional frequency features: Traditional MDF and MNF use the power spectrum, whereas MMDF and MMNF use the amplitude spectrum.
  • Comparison: Amplitude-spectrum variation is lower than power-spectrum variation, so MMNF and MMDF vary less than traditional MNF and MDF.

3.3 Evaluation methods

The evaluation measures feature robustness using percentage error between clean and noisy sEMG features, with added WGN tested across decreasing SNRs. Classification rate is also used to assess recognition performance in noisy conditions.

  • Robustness metrics: Percentage error evaluates how robust each EMG feature is to WGN.The clean and noisy feature vectors are compared; smaller PE indicates better performance.
  • Experimental design: Figures 6 and 7 summarize PE across feature families, parameter settings, signal strengths, and SNR conditions.Figure 6 examines representative feature groups and parameters; Figure 7 compares nine selected features for strong and weak sEMG signals.
  • Noise evaluation: WGN is added to original sEMG signals at SNRs ranging from 20 to 0 dB.Average PE is calculated for each motion using repeated datasets, with WGN added ten times at each noise level.
  • Experimental design: The evaluation covers four motions, two channels, and 80 datasets for each feature and noise level.Ten repeated datasets are used for each motion, while noise levels vary across the tested SNR range.
  • Recognition evaluation: Classification rate evaluates recognition-system performance with noisy sEMG signals.The study also uses recognition performance to assess robust features in a pattern-recognition setting.

4 RESULTS AND DISCUSSION

Robustness was evaluated across representative EMG features under additive WGN, signal strengths, and noisy classification settings. MMNF performed best for weak and strong signals, while combining MMNF with WAMP and HEMG improved noisy-environment discrimination.

  • Robustness evaluation: SNRs from 20 dB to 0 dB were used to test sixteen traditional and two novel features under additive WGN.Nine representative features were then selected for detailed robustness evaluation.
  • Feature selection: RMS showed the strongest noise-tolerance performance among amplitude-based time-domain features, while ZC and WAMP used a 10 mV threshold.SSC used a 30 mV threshold; HEMG used its second bin and AR used first order.
  • Strong-signal robustness: 6% average PE was reported for MMNF in strong sEMG at 0 dB SNR, with 3.5% for wrist extension and hand open from extensor carpi radialis longus.At SNRs above 10 dB, MMNF had the smallest average PE among strong-signal features.
  • Weak-signal robustness: About 10% average PE was reported for MMNF in weak sEMG at 0 dB SNR, while other features exceeded 20%.MMDF and HEMG followed MMNF in weak-signal robustness; wrist extension from flexor carpi radialis had 5% PE.
  • Noisy classification: The MMNF, WAMP, and HEMG combination provided stronger discriminatory power in noisy classification than the Hudgins and Oskoei multi-feature groups.Single-feature recognition results were generally weaker, supporting the use of combined feature vectors.

5 CONCLUSION

The study introduced MMNF as a robust EMG feature under WGN and evaluated it alongside traditional and other novel features. MMNF showed low error at both strong and weak signal levels and contributed to an effective multi-feature vector for noisy recognition.

  • Conclusion: MMNF was identified as the best feature for EMG robustness under WGN among sixteen traditional and two novel features.The evaluation covered strong and weak sEMG signals across SNR values from 20 dB to 0 dB.
  • Conclusion: 6% average error occurred for MMNF in strong sEMG and 10% in weak sEMG at 0 dB SNR.At 20 dB SNR, average error was 0.4% for both strong and weak sEMG signals.
  • Conclusion: MMNF, WAMP, and HEMG were used as candidates for a multi-source feature vector in EMG pattern recognition.The paper recommends testing new MMNF-based multi-feature sets with other classifier types in future work.
Loading 0912.3973v2…