Narrow Results

In recent years, deep learning has shown very competitive performance in seizure detection. However, most of the currently used methods either convert electroencephalogram (EEG) signals into spectral images and employ 2D-CNNs, or split the one-dimensional (1D) features of EEG signals into many segments and employ 1D-CNNs. Moreover, these investigations are further constrained by the absence of consideration for temporal links between time series segments or spectrogram images. Therefore, we propose a Dual-Modal Information Bottleneck (Dual-modal IB) network for EEG seizure detection. The network extracts EEG features from both time series and spectrogram dimensions, allowing information from different modalities to pass through the Dual-modal IB, requiring the model to gather and condense the most pertinent information in each modality and only share what is necessary. Specifically, we make full use of the information shared between the two modality representations to obtain key information for seizure detection and to remove irrelevant feature between the two modalities. In addition, to explore the intrinsic temporal dependencies, we further introduce a bidirectional long-short-term memory (BiLSTM) for Dual-modal IB model, which is used to model the temporal relationships between the information after each modality is extracted by convolutional neural network (CNN). For CHB-MIT dataset, the proposed framework can achieve an average segment-based sensitivity of 97.42%, specificity of 99.32%, accuracy of 98.29%, and an average event-based sensitivity of 96.02%, false detection rate (FDR) of 0.70/h. We release our code at https://github.com/LLLL1021/Dual-modal-IB.

Sign language recognition (SLR) is a useful tool for the deaf-mute to communicate with the outside world. Although many SLR methods have been proposed and have demonstrated good performance, continuous SLR (CSLR) is still challenging. Meanwhile, due to the heavy occlusions and closely interacting motions, there is a higher requirement for the real-time efficiency of CSLR. Therefore, the performance of CSLR needs further improvement. The highlights include: (1) to overcome these challenges, this paper proposes a novel video-based CSLR framework. This framework consists of three components: an OpenPose-based skeleton stream extraction module, a RGB stream extraction module, and a combination module of the BiLSTM network and the conditional hidden Markov model (CHMM) for CSLR. (2) A new residual network with Squeeze-and-Excitation blocks (SEResNet50) for video sequence feature extraction. (3) This paper combines the SEResNet50 module with the BiLSTM network to extract the feature information from video streams with different modalities. To evaluate the effectiveness of our proposed framework, experiments are conducted on two CSL datasets. The experimental results indicate that our method is superior to the methods in the literature.

Aiming at the low accuracy of the track signal intrusion detection (IDe) algorithm in the traditional cloud-side collaborative computing environment, this paper proposes a deep learning (D-L)-based track signal IDe method in the cloud edge collaborative computing environment. First, the main framework of the IDe method is constructed by comprehensively considering the backbone network, network transmission and ground equipment, and edge computing (EC) is introduced to cloud services. Then, the The CNN (Convolutional Neural Networks)-attention-based BiLSTM (Bi-directional Long Short-Term Memory) neural network is used in the cloud center layer of the system to train the historical data, a D-L method is proposed. Finally, a pooling layer and a dropout layer are introduced into the model to effectively prevent the overfitting of the model and achieve accurate detection of track signal intrusion. The purpose of introducing the pooling layer is to accelerate the model convergence, remove the redundancy and reduce the feature dimension, and the purpose of introducing the dropout layer is to prevent the overfitting of the model. Through simulation experiments, the proposed IDe method and the other three methods are compared and analyzed under the same conditions. The results show that the F1 value of the method proposed in this paper is optimal under four different types of sample data. The F1 value is the lowest of 0.948 and the highest of 0.963. The performance of the algorithm is better than the other three comparison algorithms. The method proposed in this paper is important for solving the IDe signal in the cloud-edge cooperative environment, and also provides a theoretical basis for tracking the signal IDe direction.

Aiming at the problem of low accuracy rate of current sentiment analysis methods for book review texts, a book review sentiment analysis method based on BERT-ABiLSTM hybrid model is proposed. First, the overall framework of sentiment analysis is constructed by integrating sentiment vocabulary and deep learning methods, and the fine-grained sentiment analysis is divided into three stages: topic identification, sentiment identification and thematic sentiment identification. Then, a dynamic character-level word vector containing contextual information is generated using a bidirectional encoder representation from transformers (BERT) pre-trained language model. Then, the contextual information in the text data is fully learned by introducing the bidirectional long short-term memory (BiLSTM) model. Finally, the accurate analysis of book review sentiment is achieved by using Attention mechanism to highlight important features and improve the efficiency of resource utilization. Through an experimental comparison with existing advanced algorithms, the proposed method in this study has improved at least 4.2%, 3.9% and 3.79% in precision, recall and F1 values, respectively. The experimental results show that the proposed BERT-ABiLSTM is higher than the existing models under different metrics, indicating that the proposed model has a good application prospect in the fields of book review analysis and book recommendation.

The improvement of edge perception layer anomaly detection performance has an immeasurable driving effect on the development of smart cities. However, many existing anomaly detection methods often suffer from problems such as ignoring the correlation between multiple source temporal sequences and losing key features of a single temporal sequence. Therefore, a new anomaly detection method using BiLSTM and attention mechanism is proposed. First, a fusion algorithm TCDCD was formed by combining Data Correlation Detection (DCD) and Temporal Continuity Detection (TCD) to preprocess Edge Perception Data (EPD). Then, BiLSTM is employed to gather deep-level features of EPD, and the attention mechanism is utilized to enhance important features that contribute to anomaly detection. Ultimately, the SoftMax classifier is employed to categorize abnormal data. The experimental findings from the SWaT and WADI datasets demonstrate that the suggested method achieves better performance than other newer anomaly detection methods. Among them, the accuracy, precision, recall and F1 of the proposed method on the SWaT dataset were 96.62%, 94.32%, 96.02% and 94.30%, respectively. In terms of performance, it is superior to traditional EPD anomaly detection models, and has good representational and generalization capabilities.

Lying state detection is a typical psychological calculation problem with significant spatiotemporal dynamic changes. However, most existing detection methods have not fully considered the dynamic changes in psychological states, and the distribution information of time series in multimodal spaces has not effectively utilized. The current detection systems lack adaptive fusion methods for multimodal features and it is difficult to extract their spatiotemporal dependencies. Therefore, a novel speech lie detection model was proposed that combines a Convolutional Neural Network (CNN) with a Bidirectional Long Short-Term Memory (BiLSTM) neural network and multimodal feature fusion of Spatiotemporal Attention Mechanism (SAM). CNN has the ability to extract local spatial features, while BiLSTM can handle long sequences and long-term dependencies in bidirectional information flows, and the contextual information in sequences can be captured. The proposed model combined the short-term stationary characteristics in time and the diversity of semantic environments in space, and introduced SAM to fuse multimodal features of temporal and spatial dependencies as feature vectors for the detection model of lying psychological states. The simulation experiment results on the Open-source Real Life Trial lie database show that the average lie detection rate reaches 88.09%. In general, the proposed speech lie detection model has a significant detection accuracy improvement compared to the existing lie detection models.

Network classification plays a crucial role in various domains like social network analysis and bioinformatics. While Graph Neural Networks (GNNs) have achieved significant success, they struggle with the problem of over-smoothing and capturing global information. Additionally, GNNs require a large amount of data, hindering performance on small datasets. To address these limitations, we propose a novel approach utilizing Deng’s entropy, capturing network topology and node/edge information. This entropy is calculated at multiple scales, resulting in an entropy sequence that incorporates both local and global features. We embed the networks by combining the entropy sequences for edges and nodes into a matrix, which then are fed into a bidirectional long short-term memory network to perform network classification. Our method outperforms GNNs in the bioinformatics, social, and molecule domains, achieving superior classification power on nine out of eleven benchmark datasets. Further experiments with both real-world and synthetic datasets highlight its exceptional performance, achieving an accuracy of 97.24% on real-world complex networks and 100% on synthetic complex networks. Additionally, our approach proves effective on datasets with a small number of networks and unbalanced classes and excels at distinguishing between synthetic and real-world networks.

In recent times, video event detection gained high attention in the researcher’s community, because of its widespread applications. In this paper, a new model is proposed for detecting different human actions in the video sequences. First, the videos are acquired from the University of Central Florida (UCF) 101, Human Motion Database (HMDB) 51 and Columbia Consumer Video (CCV) datasets. In addition, the DenseNet201 model is implemented for extracting deep feature values from the acquired datasets. Further, the Improved Gray Wolf Optimization (IGWO) algorithm is developed for selecting active/relevant feature values that effectively improve the computational time and system complexity. In the IGWO, leader enhancement and competitive strategies are employed to improve the convergence rate and to prevent the algorithm from falling into the local optima. Finally, the Bi-directional Long Short Term Memory (BiLSTM) network is used for event classification (101 action types in UCF101, 51 action types in HMDB51, and 20 action types in CCV). In the resulting phase, the IGWO-based BiLSTM network achieved 94.73%, 96.53%, and 93.91% accuracy on the UCF101, HMDB51, and CCV datasets.

Recent advancements in Natural Language Processing (NLP) have made sentiment analysis an essential component of a variety of NLP jobs, including recommendation systems, question answering, and business intelligence products. While sentiment analysis research has, to put it mildly, been widely pursued in English, Telugu has barely ever attempted the task. The majority of research works concentrate on analysing the sentiments of Tweets, news, or reviews containing Hindi and English words. There is a growing interest among academics in studying how people express their thoughts and views in Indian languages like Bengali, Telugu, Malayalam, Tamil and so on. Due to a paucity of labelled datasets, microscopic investigation on Indian languages has been published to our knowledge. This work suggested a sentence-level sentiment analysis on multi-domain datasets that has been collected in Telugu. Deep learning models have been used in this work because it demonstrates the significant expertise in sentiment analysis and is widely regarded as the cutting-edge model in Telugu Sentiment Analysis. Our proposed work investigates a productive Bidirectional Long Short-Term Memory (BiLSTM) Network and Bidirectional GRU Network (BiGRU) for improving Telugu Sentiment Analysis by encapsulating contextual information from Telugu feature sequences using Forward-Backward encapsulation. Further, the model has been deployed by merging the domains so as to predict the accuracy and other performance metrics. The experimental test findings show that the deep learning models outperform when compared with the baseline traditional ML methods in four benchmark sentiment analysis datasets. There is evidence that the proposed sentiment analysis method has improved precision, recall, F1-score and accuracy in certain cases. The proposed model has achieved the F1-score of 86% for song datasets when compared with the other existing models.