Identifying Suitable Brain Regions and Trial Size Segmentation for Positive/Negative Emotion Recognition.

Considering the use of a multi-modal framework to enhance emotion recognition, we propose to combine electroencephalography (EEG) and electrooculogram (EOG) through decision level fusion(DLF) and feature level fusion(FLF) for emotion recognition. By using different temporal window sizes to segment the signal, we explore the duration of the emotion of the EOG signal and the EEG signal. Then, some temporal window sizes that are friendly to both EOG signal and EEG signal are selected for segmentation and emotion recognition. According to the different degree of dependence of subjects, the accuracy of the proposed algorithm on subject-dependent and subject-independent is verified on the DEAP dataset. For subject-dependent, using feature level fusion strategy with a window size of 6 seconds, the accuracy is 0.9562 in terms of arousal, and 0.9558 in terms of valence. For subject-independent, using feature level fusion strategy with a window size of 5 seconds, the accuracy is 0.8638 in terms of arousal, and 0.8542 in terms of valence. The experimental results show that the proposed algorithm can better enhance emotion recognition.

The Wigner–Ville transform was used twice for analysis of periodic broadband noisy signals. The first time use is for an extraction of informative part (part in future) from the signal when time window is sliding across a signal from a beginning to the end. Time window overlap N% is applied. This corresponds to N% time of existing window for a part of analyzed signal. The time window size is equal to the part duration time T1. That allows increasing the signal to noise ratio for this part within the window. Then the narrow Fourier filtration (which corresponds to the most informative part of the signal) is applied for the output of time window. Broadband signal becomes narrow band after that. Second time the Wigner–Ville approach is used to extract a periodicity in filtered signal. In this case second window was applied. This window size is at least two times more parts’ repetition time T2. Overlap is M%. This signal processing was successfully used for acoustic and electromagnetic signals of human motion due to footsteps and for underwater clicks of beaked whales in passive acoustic identification. [This work was supported by Army Armament Research, Development, and Engineering Center Contract No. W15QKN-09-C-0163.]

Temporal networks can be used to model systems that evolve over longer time scales such as networks of disease spread, for instance, HIV/AIDS disease that is propagated within the population over a relatively long period. Analyzing temporal networks can be done by considering the network either as a series of snapshots (aggregation over a time window) or as a dynamic object whose structure changes over time. The first approach is used in this paper and requires specifying a size of time window that delimits snapshot size. To our best knowledge, there is not yet studies on setting the size of the window in a methodical basis. In real, existing works rely on a static or a regular value of time window size to capture snapshots over time.

Threshold methods in financial networks are important tools for obtaining important information about the financial state of a market. Previously, absolute thresholds of correlation coefficients have been used; however, they have no relation to the length of time. We assign a threshold value depending on the size of the time window by using the P-value concept of statistics. We construct a threshold network (TN) at the same threshold value for two different time window sizes in the Korean Composite Stock Price Index (KOSPI). We measure network properties, such as the edge density, clustering coefficient, assortativity coefficient, and modularity. We determine that a significant difference exists between the network properties of the two time windows at the same threshold, especially during crises. This implies that the market information depends on the length of the time window when constructing the TN. We apply the same technique to Standard and Poor’s 500 (S&P500) and observe similar results.

Electroencephalogram (EEG) based emotion recognition plays a vital role in affective computing and braincomputer interface applications. In this paper, we propose AMGCN, a novel adaptive graph neural network with multi-scale attention for robust emotion recognition from EEG signals. The model first extracts differential entropy (DE) features and maps them to a spatial 2D topological grid. A deep CNN is employed to capture local spatial-spectral representations. We introduce a multi-scale attention module that adaptively reweights EEG features across spatial, spectral, and temporal dimensions, enhancing emotion-relevant representations by jointly modeling brain-region topology, frequency, and samples. Furthermore, GCN is used to model individualized brain connectivity via a learnable adjacency matrix. Experimental results on the DEAP dataset demonstrate that AM-GCN achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 5. 1 8 \%}$</tex> accuracy for valence and 94.86 % for arousal, outperforming several state-of-the-art methods. These results highlight the model's ability to integrate spatial, spectral, and relational features, offering both high accuracy and interpretability.

Traditional electroencephalograph (EEG)-based emotion recognition requires a large number of calibration samples to build a model for a specific subject, which restricts the application of the affective brain computer interface (BCI) in practice. We attempt to use the multi-modal data from the past session to realize emotion recognition in the case of a small amount of calibration samples. To solve this problem, we propose a multi-modal domain adaptive variational autoencoder (MMDA-VAE) method, which learns shared cross-domain latent representations of the multi-modal data. Our method builds a multi-modal variational autoencoder (MVAE) to project the data of multiple modalities into a common space. Through adversarial learning and cycle-consistency regularization, our method can reduce the distribution difference of each domain on the shared latent representation layer and realize the transfer of knowledge. Extensive experiments are conducted on two public datasets, SEED and SEED-IV, and the results show the superiority of our proposed method. Our work can effectively improve the performance of emotion recognition with a small amount of labelled multi-modal data.

A novel semi-supervised meta learning method for subject-transfer brain–computer interface

Link quality estimation plays a vital role in wireless sensor networks, which can improve transmission efficiency and reduce node energy consumption by selecting the links with good quality. Packet Reception Ratio (PRR), Received Signal Strength Indicator (RSSI) and Link Quality Indicator (LQI) are typical metrics for link quality estimation. However, there is a lack of quantitative analysis on the convergence of these three metrics in existing studies. More specifically, what the size of time window should be used to process these metrics to reflect link quality more accurately? To solve this problem, this paper conducted comprehensive and quantitative analysis on the convergence of PRR, RSSI and LQI through collecting and analyzing large amounts of experimental data. Experimental results show that only about 10 packets are enough for RSSI and LQI to evaluate the link quality accurately, while 50 packets are needed for PRR under the same conditions.

Motor intention in the posterior parietal cortex

This paper proposed an automatic steady-state visual evoked potential (SSVEP) component selection (SCS) measure for a high-performance SSVEP-based brain-computer interface (SBCI) system. First, multi-electrode raw electroencephalogram signals are spatially pre-processed using a blind source separation technique resulting in multi-source components. The SCS measure of each component is then calculated by continuous wavelet transform (CWT), and the ensemble features that contain the weighted CWT energy of individual SSVEP harmonic are extracted. Second, the SSVEP component with maximal SCS measure is considered to have the highest signal-to-noise ratio. In our SBCI system, six stimulus frequencies served as the input patterns. Offline analyses were performed, through which the common electrode locations, the time window size, and the number of harmonics were defined. Thereafter the results of our method were compared with those of others. We next carried out an online test of the SBCI for 11 subjects using eight common electrode locations, a 1.5-s time window, and the first and second harmonics. The test results showed that our method achieved an average accuracy of 95.2 % and a practical bit rate of 68.2 bits/min.

Remaining useful life prediction methods are extensively researched based on failure or suspension histories. However, for some applications, failure or suspension histories are hard to obtain due to high reliability requirement or expensive experiment cost. In addition, some systems' work condition cannot be simulated. According to current research, remaining useful life prediction without failure or suspension histories is challenging. To solve this problem, an individual-based inference method is developed using recorded condition monitoring data to date. Features extracted from condition data are divided by adaptive time windows. The time window size is adjusted according to increasing rate. Features in two adjacent selected windows are regarded as the inputs and outputs to train an artificial neural network. Multi-step ahead rolling prediction is employed, predicted features are post-processed and regarded as inputs in the next prediction iteration. Rolling prediction is stopped until a prediction value exceeds failure threshold. The proposed method is validated by simulation bearing data and PHM-2012 Competition data. Results demonstrate that the proposed method is a promising intelligent prognostics approach.

Meteorological factors, which are periodic and regular in a long run, have an unignorable impact on human health. Accurate health risk prediction based on meteorological factors is essential for optimal allocation of resource in healthcare units. However, due to the non-stationary and non-linear nature of the original hospitalization sequence, traditional methods are less robust in predicting it. This study aims to investigate hospital admission prediction models using time series pre-processing algorithms and deep learning approach based on meteorological factors. Using the electronic medical record data from Panyu Central Hospital and meteorological data of Panyu district from 2003 to 2019, 46,089 eligible patients with lower respiratory tract infections (LRTIs) and four meteorological factors were identified to build and evaluate the prediction models. A novel hybrid model, Cascade GAM-CEEMDAN-LSTM Model (CGCLM), was established in combination with generalized additive model (GAM), complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and long-short term memory (LSTM) networks for predicting daily admissions of patients with LRTIs. The experimental results show that CGCLM multistep method proposed in this paper outperforms single LSTM model in the prediction of health risk time series at different time window sizes. Moreover, our results also indicate that CGCLM has the best prediction performance when the time window is set to 61 days (RMSE = 1.12, MAE = 0.87, R2 = 0.93). Adequate extraction of exposure-response relationships between meteorological factors and diseases and suitable handling of sequence pre-processing have an important role in time series prediction. This hybrid climate-based model for predicting LRTIs disease can also be extended to time series prediction of other epidemic disease.

Relationship between braking intensity and driver heart rate as a function of the size of the measurement window and its position

Spiking Neural Networks (SNNs) are brain- inspired computing models incorporating unique temporal dynamics and event-driven processing. Rich dynamics in both space and time offer great challenges and opportunities for efficient processing of sparse spatiotemporal data compared with conventional artificial neural networks (ANNs). Specifically, the additional overheads for handling the added temporal dimension limit the computational capabilities of neuromorphic accelerators. Iterative processing at every time-point with sparse inputs in a temporally sequential manner not only degrades the utilization of the systolic array but also intensifies data movement.In this work, we propose a novel technique and architecture that significantly improve utilization and data movement while efficiently handling temporal sparsity of SNNs on systolic arrays. Unlike time-sequential processing in conventional SNN accelerators, we pack multiple time points into a single time window (TW) and process the computations induced by active synaptic inputs falling under several TWs in parallel, leading to the proposed parallel time batching. It allows weight reuse across multiple time points and enhances the utilization of the systolic array with reduced idling of processing elements, overcoming the irregularity of sparse firing activities. We optimize the granularity of time-domain processing, i.e., the TW size, which significantly impacts the data reuse and utilization. We further boost the utilization efficiency by simultaneously scheduling non-overlapping sparse spiking activities onto the array. The proposed architectures offer a unifying solution for general spiking neural networks with commonly exhibited temporal sparsity, a key challenge in hardware acceleration, delivering 248X energy-delay product (EDP) improvement on average compared to an SNN baseline for accelerating various networks. Compared to ANN based accelerators, our approach improves EDP by 47X on the CIFAR10 dataset.

Fault seismic attribute volumes (such as volumetric coherence and curvature) represent an efficient and objective way to visualize and identify faults in seismic cubes. Fault geometric attributes such as length, height, and fault segmentation can be extracted from such fault seismic attribute volumes. We evaluate the strengths and pitfalls of using coherence volumes for characterization of fault geometry. The results are obtained using a database from the Barents Sea, which contains 35 3D seismic cubes, together with conceptual synthetic seismic models. A high signal-to-noise ratio is a requirement for the extraction of accurate fault geometric data. Noise attenuation methods improve fault visualization, but our results indicate that the effect of noise attenuation on the extracted fault geometric attributes is only clear in areas of low signal-to-noise ratios. The choice of coherence algorithm is important when extracting fault length data. Semblance-based coherence performs better than gradient structure tensor-based coherence in low-displacement areas near the fault tips, and it produces more accurate fault length data. Faults can appear segmented in coherence volumes if relatively similar reflectors are juxtaposed across a fault. In such areas, it is important that the interpreter does not overlook the fault. The size of the analysis window used in coherence calculations controls the resolution and continuity of the imaged faults. Our results support an optimal temporal window size of one to two times the dominant period of the seismic data (typically 7–17 samples in conventional 4 ms sampled 3D seismic data). Larger temporal window sizes can result in an overestimation of fault height, especially for small faults. A large spatial window can smear out segmentation along the fault and make the fault traces wider. Even though a large spatial window can have some positive effects, we recommend using a relatively small spatial window (five traces) when extracting subtle fault geometric attributes.