Search

Detecting respiration and heartbeat information from human chest movement is crucial for non‐contact vital sign detection. Due to the small frequency difference between respiration and heartbeat signals and the relatively low amplitude of the heartbeat signal, it is challenging to separate and extract them completely. This letter proposes a successive variational mode extraction method based on spectrum trend for the extraction of respiration and heartbeat signals. The proposed method adaptively determines the initial centre frequency and extraction order of the desired signal modes by analyzing the spectrum trend of the signal to be decomposed. Additionally, a recursive framework is introduced to sequentially extract the desired signal modes. Both simulation and practical experiment results show that the method is effective in extracting respiration and heartbeat signals, significantly improving the accuracy of estimating respiration and heartbeat frequencies. [ABSTRACT FROM AUTHOR]

An aggressive driving state impacts drivers' decisions, which could potentially lead to accidents. Real-time recognition of driving state is particularly important for improving road safety. However, the majority of modeling in existing studies relies on a single algorithm, which may lead to unreliable predictions. This paper proposes a stacking ensemble aggressive driving state recognition model using electroencephalography (EEG), which is able to combine different heterogeneous classification algorithms. Five types of classification algorithms and their variants are tested and compared to identify suitable base classifiers. All of these classifiers are optimized by Bayesian optimizer before the comparison. Three stacking ensemble recognition models using different meta-classifiers (i.e., logistic regression, random forest, and AdaBoost) and an equal-weight voting ensemble recognition model are established. The aforementioned recognition models are evaluated by using a dataset collected from a car-following simulated driving experiment. Fast Fourier transformation (FFT) and wavelet packet transformation (WPT) are adopted to extract features from raw EEG data. The results suggest that the stacking ensemble recognition models outperform the best single (i.e., support vector machine) model; the random Forest stacking recognition model achieves the best performance and the accuracy is increased from 81.21% to 84.23% using FFT features and from 86.45% to 87.38% using WPT features. [ABSTRACT FROM AUTHOR]