EMG Signal Processing & Analysis | Biosignal Filtering & FFT

This research presents an EMG-based prosthesis control system that uses machine learning to classify electromyography signals from the forearm for controlling prosthetic hand movements. The system consists of three main components: a Myo armband with 8 channels for acquiring EMG signals, an Arduino microcontroller for data processing, and a decision tree classifier trained on 12 datasets with a maximum depth of 32. The prosthesis is controlled by five servo motors, with each finger movement executed by one motor. The system achieved 98% accuracy in classifying 12,000 EMG signals, demonstrating the effectiveness of machine learning approaches for prosthetic control. Future improvements could include adding more datasets, using more complex machine learning algorithms, and incorporating additional sensors for enhanced human-body interaction.

Myoelectric prosthetic hands detect muscle activity (EMG) from residual limbs and convert it to motion. Traditional machine learning approaches require extensive training data and large model parameters, causing user burden and phantom pain. The proposed solution transforms EMG signals into cylindrical space where circumferential position and signal intensity are represented. Kernel density estimation estimates the distribution of incoming signals, enabling comparison with reference actions without complex classifiers.

This section covers advanced EMG signal processing techniques for prosthetic control. The presenter explains how frequency spectrum analysis extracts characteristic peaks from EMG signals, enabling identification of specific muscle movements. The video demonstrates sampling at 100 samples per millisecond to capture signal data for processing. Finally, the presenter introduces neural networks as classifiers that correlate extracted frequency features with known movement patterns, enabling automated recognition of complex muscle activation patterns for independent prosthetic control.

Prosthetic control systems classify EMG signals into distinct categories representing different intended movements. In this project, three states were identified: flexion attempt (muscle contraction toward palm closure), neutral/rest position, and extension attempt (muscle relaxation toward palm opening). Initial single-model AI approaches struggled with accurate classification across all states, particularly during transitions between positions. Parallel dual-model architectures (stacked classifiers) proved more effective by simultaneously detecting presence/absence of movement while distinguishing between flexion and extension.

Classification-based EMG control uses machine learning algorithms to identify specific movement patterns from multiple EMG channels. The system extracts features from the EMG signals and trains classifiers to recognize different intended actions (e.g., hand open, hand close, wrist flexion, wrist extension). The classifier outputs discrete decisions at specific time points, which are then used to control the prosthesis. This approach can handle more complex control tasks than simple threshold-based methods.

The Short-Time Fourier Transform (STFT) divides a time signal into overlapping segments, applies a window function to each segment, and computes the Fourier transform separately to obtain a time-frequency representation called a spectrogram, which captures how frequency content changes over time; however, STFT has fixed window size limiting its ability to adapt to varying frequency resolutions, whereas the Wavelet Transform overcomes this limitation by using a scalable 'mother wavelet' that can be dilated (shrunk or expanded) to provide better time localization for high-frequency components and better frequency localization for low-frequency components, enabling more flexible analysis of non-stationary signals.

Joint time-frequency analysis methods provide simultaneous information about amplitude, time, and frequency components, which is impossible with conventional time-domain or frequency-domain analysis alone. These methods include Short-Time Fourier Transform (STFT), Continuous Wavelet Transform, Discrete Wavelet Transform, Wavelet Packet Decomposition, Wigner-Ville Decomposition, and Empirical Mode Decomposition. Applications include noise removal, arrhythmia detection, sleep apnea detection, biometric identification, emotion detection, epileptic seizure detection, driver distraction detection, and drug/alcohol detection. Fourier transform fails for non-stationary signals (where frequency components change over time) because it loses time information. All biomedical signals are inherently non-stationary, necessitating advanced analysis methods. STFT uses sliding windows where shorter windows provide better time resolution but poorer frequency resolution, requiring optimal window selection to balance both.

Biosignals are inherently non-stationary with transient events occurring at specific times. Time-frequency representations localize both time and frequency content, unlike stationary signal analysis. Short-time Fourier transform works for locally stationary signals. Wavelet transforms and synchro-squeezing algorithms offer superior localization for highly non-stationary biosignals. The choice depends on signal characteristics rather than following trends, as each method reveals different aspects of the signal's temporal-frequency structure.

The Wavelet Transform is a time-frequency analysis technique that overcomes the limitations of FFT by using multi-resolution windowing with wavelets—finite-duration oscillations that can simultaneously provide good time and frequency resolution. Unlike FFT which uses fixed windows creating a trade-off between time and frequency resolution, wavelet analysis adapts resolution: at low frequencies it provides high frequency resolution (since low-frequency components last longer), and at high frequencies it provides high time resolution (since high-frequency components last shorter). The process involves scaling (stretching/compressing) and shifting a mother wavelet across the signal to identify which frequencies are present at specific time instants through correlation matching.

This section covers advanced techniques for analyzing non-stationary physiological signals. The Short-Time Fourier Transform (STFT) addresses Fourier Transform limitations by analyzing signal segments over specific time windows, visualized through spectrograms showing frequency changes over time. This technique reveals muscle fatigue (decreasing median frequency) and seizure activity in EEG. However, STFT involves a fundamental tradeoff between time and frequency resolution—shorter windows provide better time resolution but worse frequency resolution. The Wavelet Transform overcomes this limitation by providing variable resolution: better time resolution at higher frequencies and better frequency resolution at lower frequencies, achieved through recursive filtering that decomposes signals into narrower frequency bands.

This section presents the complete software architecture for implementing digital filters on microcontrollers. It covers designing filter structures containing pre-computed coefficients and output buffers, writing initialization functions that calculate RC constants and filter coefficients from user-specified parameters, and implementing update functions that perform sample shifting and recursive filtering calculations. The video demonstrates integrating these filters with sensor data acquisition systems, showing how to pass accelerometer data through the filter and transmit both raw and filtered outputs. Practical considerations include ensuring adequate sampling rates to satisfy Nyquist requirements and maintaining stability in IIR filter implementations.

This extensive section covers deploying filters on microcontrollers. MPLAB IDE generates peripheral templates for I2S configuration, simplifying hardware abstraction. The main loop manages system state while DMA handles data movement. DMA callback functions trigger when buffers fill, invoking processing routines asynchronously. Pass-through implementations copy data from ADC to DAC buffers, establishing baseline behavior. Integrating filters involves converting I2S data to floating-point, applying the filter algorithm sample-by-sample, then converting back to I2S format. The implementation maintains state variables storing delayed samples across iterations. Despite apparent mathematical simplicity, real-time deployment requires careful attention to coefficient accuracy, data type conversions, buffer management, and timing constraints. The WYSIWYG myth of digital design proves misleading—debugging can be equally challenging as analog design despite apparent simplicity. The complete workflow demonstrates deploying a notch filter on a $50 Curiosity Nano board with free tools, showing how mathematical concepts translate to working hardware implementations.

Digital filters can be implemented in microcontrollers by converting their transfer functions into discrete-time equations that require past input and output values, which are stored in vectors and updated at each sampling period to calculate the current output.

Implementing digital filters requires converting transfer functions to difference equations and programming them on microcontrollers. The implementation chain includes: (1) Anti-aliasing filter at input to prevent aliasing; (2) Analog-to-digital converter (ADC) to digitize the signal; (3) Microcontroller executing the difference equation algorithm; (4) Digital-to-analog converter (DAC) at output with smoothing filter. For a filter with H(z) = (0.21z²) / (z² - 1.6z + 0.81), the difference equation is y[n] = 0.21x[n] + 0.21x[n-1] - 1.6y[n-1] + 0.81y[n-2]. The microcontroller maintains state variables (previous inputs and outputs) and computes each new output sample iteratively in a loop.

This seminar teaches digital filter implementation through practical hands-on experience using Arm microcontrollers (NUCLEO-F446RE), allowing participants to design and implement filters like moving average, FIR, and IIR filters while observing real-time waveform changes and audio effects, providing a comprehensive understanding of digital signal processing theory through direct experimentation.