User identification system based on 2D CQT spectrogram of EMG with adaptive frequency resolution adjustment

Area of Science:

• Biometric Systems
• Signal Processing
• Machine Learning

Background:

• Electromyogram (EMG) signals offer unique individual characteristics for user identification.
• Conventional 1D feature extraction from EMG signals is limited by nonlinear and abnormal signal patterns, hindering accuracy.
• Multidimensional features, particularly time-frequency information, are crucial for enhancing EMG-based identification.

Purpose of the Study:

• To develop an improved user identification system leveraging the unique properties of EMG signals.
• To enhance identification accuracy by utilizing customized time-frequency resolution through Constant Q Transform (CQT).
• To explore the efficacy of 2D features derived from CQT for robust user identification.

Main Methods:

• Data preprocessing of electromyogram (EMG) signals.
• Conversion of EMG signals into 2D images using Constant Q Transform (CQT) with customized time-frequency resolution.
• Feature extraction using convolutional neural networks (CNNs) and classification.

Main Results:

• The proposed system achieved a high identification accuracy of 97.5%.
• This represents a significant improvement of 15.4% over traditional 1D feature methods.
• Accuracy was also improved by 2.1% compared to systems using Short-Time Fourier Transform (STFT) based features.

Conclusions:

• Constant Q Transform (CQT)-based 2D features provide a superior approach for electromyogram (EMG) user identification.
• The proposed CNN-based system effectively utilizes these multidimensional features for enhanced accuracy.
• This method overcomes limitations of conventional 1D feature extraction, paving the way for more reliable biometric systems.