Deep learning and adaptive time-frequency analysis for motor imagery EEG analysis

Abstract

Brain–computer interface (BCI) technology aims to establish a direct communication pathway between the human brain and external devices by decoding neural activity. Among various BCI paradigms, motor imagery (MI) has been widely investigated due to its non-invasive nature and practical applicability. Electroencephalography (EEG) remains one of the most commonly used modalities for MI-based BCI systems. However, the effective preprocessing and decoding of MI EEG signals are still challenged by several intrinsic characteristics, including low signal-to-noise ratio, nonstationarity, nonlinear dynamics, limited spatial resolution, and complex inter-channel dependencies.

Despite considerable progress achieved by conventional signal processing and deep learning approaches, several key challenges remain unresolved. First, computational inefficiency in multichannel adaptive time-frequency decomposition, which hinders real-time application. Second, Decoding robustness is compromised by low signal-to-noise ratios and volume conduction effects. Third, low-channel wearable systems suffer performance degradation due to severe loss of spatial information. Finally, accurately characterizing nonstationary and nonlinear dynamics remains difficult, as traditional methods are constrained by resolution limits, cross-term interference, and limited adaptability to diverse neural patterns.

To address these challenges, this PhD research proposes a unified framework that integrates adaptive time–frequency analysis and deep learning-based decoding for MI EEG signals. The main contributions of this thesis are organized into four technical studies.

First, to reduce the computational burden of multivariate adaptive signal decomposition while preserving inter-channel correlations, Fast Multivariate Empirical Mode Decomposition (FMEMD) and Noise-Assisted FMEMD (NA-FMEMD) are proposed. These methods improve computational efficiency and decomposition stability, enabling effective extraction of MI-related EEG rhythms and facilitating multi-channel feature learning.

Second, to enhance EEG decoding performance under low signal-to-noise and limited spatial resolution conditions, a Parallel Multi-Scale Attention Network (PMSA-Net) is developed. By employing parallel multi-scale feature extraction and channel attention mechanisms, the proposed network effectively captures spatial dependencies and multi-scale representations, leading to improved classification accuracy and robustness.

Third, to address the challenges within performance Degradation in low-channel EEG systems, a Temporal–Spectral Cross-Fusion Network (TSCF-Net) is introduced. This architecture integrates complementary temporal–spatial and spectral–spatial features through a multi-modal fusion strategy, thereby compensating for the loss of spatial information and improving decoding performance in low-channel scenarios.

Finally, to overcome the limitations of extracting characterization of nonstationary and nonlinear signal dynamics, a theory-guided Quadratic Time–Frequency Network (QTFN) is proposed. The network learns adaptive basis functions and generates high-resolution, cross-term-free time–frequency representations. Despite being trained solely on synthetic signals, QTFN demonstrates strong generalization capability when applied to real MI EEG data.

Overall, this thesis systematically addresses critical challenges in MI EEG preprocessing and decoding, providing efficient signal analysis methods, robust deep learning architectures, and interpretable time–frequency representations, thereby contributing to the advancement of practical and scalable BCI systems.

Date of Award	18 Jul 2026
Original language	English
Awarding Institution	University of Nottingham
Supervisor	Ruibin Bai (Supervisor), Alain Chong (Supervisor) & Yi Pan (Supervisor)