Abstract
EMG is a signal which could reflect the information of human body's movement in residual limb muscle group. As a source of information, it could control the bionic prosthesis to complete the multi-degree of freedom of action. EMG prosthesis controlled the bionic prosthesis through the body's own physiological signal. It is remarkable significant for the disabled users that if the bionic prosthesis could understand the movement of the disabled users will. The purpose of the EMG signal acquisition is driving a bionic hand which could help the hand disabilities to achieve grasping, pushing, taking, holding etc. EMG prosthetic is a complex system, which involving multidisciplinary disciplines. And this paper mainly introduces the method of EMG extraction.
Introduction
The primary challenge in bionic limb development is the "translation" problem: how to convert the noisy, stochastic electrical signals generated by residual muscles (Electromyography, or EMG) into precise mechanical movements. While early prosthetics relied on simple external drivers or basic triggers, modern users require "natural" control for complex actions like grasping, pushing, and holding. However, EMG signals are notoriously difficult to process—they are non-stationary, susceptible to electromagnetic interference, and vary significantly based on muscle fatigue and electrode placement. This research evaluates the primary extraction methods used to overcome these hurdles.
Methodology and Signal Processing Pipeline
The transition from raw muscle activity to prosthetic movement requires a sophisticated signal processing pipeline. As illustrated in the flowchart below, this process typically involves preprocessing, activation-level estimation, and activity detection before reaching the final classification stage through sequential matching techniques.
The extraction of features from these signals can be categorized into four primary methodological approaches:
1. Time Domain Analysis
Time-domain methods are the most traditional and computationally efficient. They treat the EMG as a random signal and calculate features such as Integral EMG (IEMG), Zero Crossing (ZC), and Variance. While easy to implement for real-time applications, these methods are often unstable and highly sensitive to external noise and electrode positioning.
2. Frequency Domain Analysis
To address the instability of time-domain features, researchers utilize power spectrum analysis. Key metrics include Average Power Frequency and Median Frequency. These are particularly effective for identifying muscle fatigue, which manifests as a predictable shift in the signal's frequency spectrum, allowing for more robust pattern recognition.
3. Time-Frequency Analysis
Because EMG signals are non-stationary, traditional Fourier Transforms often prove inadequate. The Wavelet Transform is a superior alternative, employing variable window sizes—short windows for high frequencies and wide windows for low frequencies. This allows for high-fidelity real-time processing that captures the temporal evolution of the signal's local characteristics.
4. Nonlinear Dynamic Analysis
Recognizing the human body as a complex, nonlinear system, this approach applies parameters like the Lyapunov index, entropy, and correlation dimension. These metrics capture underlying muscle firing patterns that linear models might overlook, offering a path toward higher-fidelity control by acknowledging the signal's origin in the central nervous system.
Discussion and Key Insights
A critical finding of this research is the "Stability Gap"—the inherent trade-off between computational simplicity and reliability. While time-domain features are simple, they lack the robustness required for multi-degree-of-freedom gesture recognition. Frequency-domain features offer better stability, but the Wavelet Transform shows the greatest potential due to its dual-domain localization.
Furthermore, the paper highlights that as early as 1992, systems achieved 80% recognition accuracy for basic commands. However, achieving true biomimetic movement requires moving beyond linear assumptions. Because EMG signals are products of a complex nonlinear system, nonlinear dynamical analysis is likely the key to achieving intuitive, natural control.
Conclusion
The complexity of EMG signals is such that no single extraction method can fully capture the information necessary for a truly biomimetic hand. Future advancements will likely depend on the integration of multiple technologies. As processing power increases, the real-time implementation of complex Wavelet transforms and nonlinear algorithms will become standard. The ultimate goal is a "perception function" that allows bionic hands to interact intelligently with the environment based on the user's direct motor intent.
Summary updated