Multi-Layer Graph Attention Network for Sleep Stage Classification Based on EEG.

Graph neural networks have been successfully applied to sleep stage classification, but there are still challenges: (1) How to effectively utilize epoch information of EEG-adjacent channels owing to their different interaction effects. (2) How to extract the most representative features according to confused transitional information in confused stages. (3) How to improve classification accuracy of sleep stages compared with existing models. To address these shortcomings, we propose a multi-layer graph attention network (MGANet). Node-level attention prompts the graph attention convolution and GRU to focus on and differentiate the interaction between channels in the time-frequency domain and the spatial domain, respectively. The multi-head spatial-temporal mechanism balances the channel weights and dynamically adjusts channel features, and a multi-layer graph attention network accurately expresses the spatial sleep information. Moreover, stage-level attention is applied to easily confused sleep stages, which effectively improves the limitations of a graph convolutional network in large-scale graph sleep stages. The experimental results demonstrated classification accuracy; MF1 and Kappa reached 0.825, 0.814, and 0.775 and 0.873, 0.801, and 0.827 for the ISRUC and SHHS datasets, respectively, which showed that MGANet outperformed the state-of-the-art baselines.

Steady-State Conduction Current Performance for Multilayer Polyimide/SiO2 Films.

The steady-state electrical conduction current for single and multilayer polyimide (PI) nanocomposite films was observed at the low and high electric field for different temperatures. Experimental data were fitted to conduction models to investigate the dominant conduction mechanism in these films. In most films, space charge limited current (SCLC) and Poole-Frenkel current displayed dominant conduction. At a high electric field, the ohmic conduction was replaced by current-voltage dependency. Higher conduction current was observed for nanocomposite films at a lower temperature, but it declined at a higher temperature. PI nanocomposite multilayer films showed a huge reduction in the conduction current at higher electric fields and temperatures. The conclusions derived in this study would provide the empirical basis and early breakdown phenomenon explanation when performing dielectric strength and partial discharge measurements of PI-based nanocomposite insulation systems of electric motors.

Uncertainties of reverberation time estimation via adaptively identified room impulse responses.

This paper investigates the reverberation time estimation methods which employ backward integration of adaptively identified room impulse responses (RIRs). Two kinds of conditions are considered; the first is the "ideal condition" where the anechoic and reverberant signals are both known a priori so that the RIRs can be identified using system identification methods. The second is that only the reverberant speech signal is available, and blind identification of the RIRs via dereverberation is employed for reverberation time estimation. Results show that under the "ideal condition," the average relative errors in 7 octave bands are less than 2% for white noise and 15% for speech, respectively, when both the anechoic and reverberant signals are available. In contrast, under the second condition, the average relative errors of the blindly identified RIR-based reverberation time estimation are around 20%-30% except the 63 Hz octave band. The fluctuation of reverberation times estimated under the second condition is more severe than that under the ideal condition and the relative error for low frequency octave bands is larger than that for high octave bands under both conditions.

Reverberation time estimation from speech signals based on blind room impulse response identification.

A reverberation time (RT) estimation method is presented which consists of three steps, the anechoic speech is first recovered by maximizing the skewness of the linear prediction residual of the reverberant speech, then room impulse response (RIR) is identified using the recovered anechoic and reverberant speech, finally RIR is truncated to compensate the estimation errors and RT is estimated using the Schroeder's method. Simulations show that the proposed method successfully estimates RT less than 1.4 s and is insensitive to the speech content such as the number of long pauses and sharp offsets.