Time-frequency feature extraction method for EEG signals utilizing fractional-order transient-extracting transform

In light of the challenges in capturing transient features of electroencephalographic (EEG) signals under the motor imagery (MI) paradigm, this paper proposes a Fractional-order transient-extracting transform (FOTET). Transient features refer to short-duration, non-stationary waveform segments that reflect critical neural activity during MI, and their accurate extraction is essential for effective brain-computer interface performance. FOTET enhances transient feature extraction and time-frequency energy aggregation by introducing a fractional-order transient extracting operator and an iterative optimization process, which can efficiently capture weak transient signal features while overcoming the limitations of the traditional methods. Moreover, the method can balance the time and frequency resolution by adjusting the fractional order parameterα. The experimental results, based on data from 10 healthy subjects performing four-class MI tasks, indicate that FOTET can accurately extract transient features in noisy environments, effectively distinguishing EEG signals across different classes. When combined with the dense convolutional network-long short-term memory, it achieves a classification accuracy of 96.71% whenα=0.16, significantly surpassing results obtained using traditional time-frequency analysis methods, effectively validating the superiority of FOTET in EEG signal feature extraction.