A Fusion of Time-Domain Descriptors for Improved Myoelectric Hand Control.
Abstract: This paper presents a new feature extraction algorithm for the challenging problem of the classification of myoelectric signals for prostheses control. The algorithm employs the orientation between a set of descriptors of muscular activities and a nonlinearly mapped version of them. It incorporates information about the Electromyogram (EMG) signal power spectrum characteristics derived from each analysis window while correlating that with the descriptors of previous windows for robust activity recognition. The proposed idea can be summarized in the following three steps: 1) extract power spectrum moments from the current analysis window and its nonlinearly scaled version in time-domain through Fourier transform relations, 2) compute the orientation between the two sets of moments, and 3) apply data fusion on the resulting orientation features for the current and previous time windows and use the result as the final feature set. EMG data collected from nine transradial amputees performing six classes of movements with different force levels is used to validate the proposed features. When compared to other well-known EMG feature extraction methods, the proposed features produced an improvement of at least 4%.
The main idea here is that the paper considers the process of extracting the time domain features (TDD in the figure below) as a basic step only. We also consider extracting the same set of features from a nonlinearly mapped version of the original EMG signal.
Abstract: This paper presents a new feature extraction algorithm for the challenging problem of the classification of myoelectric signals for prostheses control. The algorithm employs the orientation between a set of descriptors of muscular activities and a nonlinearly mapped version of them. It incorporates information about the Electromyogram (EMG) signal power spectrum characteristics derived from each analysis window while correlating that with the descriptors of previous windows for robust activity recognition. The proposed idea can be summarized in the following three steps: 1) extract power spectrum moments from the current analysis window and its nonlinearly scaled version in time-domain through Fourier transform relations, 2) compute the orientation between the two sets of moments, and 3) apply data fusion on the resulting orientation features for the current and previous time windows and use the result as the final feature set. EMG data collected from nine transradial amputees performing six classes of movements with different force levels is used to validate the proposed features. When compared to other well-known EMG feature extraction methods, the proposed features produced an improvement of at least 4%.
The main idea here is that the paper considers the process of extracting the time domain features (TDD in the figure below) as a basic step only. We also consider extracting the same set of features from a nonlinearly mapped version of the original EMG signal.
Then the final extracted features are fused (multiplied) by the features extracted from the n'th previous window (this is called steps in the code). For example to multiply by the features extracted from the 3rd previous windows then you set steps =3.