Search

Prediction of Remaining Useful Life (RUL) for Rolling Element Bearings (REB) has attracted widespread attention from academia and industry. However, there are still several bottlenecks, including the effective utilization of multi-sensor data, the interpretability of prediction models, and the prediction across the entire life cycle, which limit prediction accuracy. In view of that, we propose a knowledge-based explainable life-cycle RUL prediction framework. First, considering the feature fusion of fast-changing signals, the Pearson correlation coefficient matrix and feature transformation objective function are incorporated to an Improved Graph Convolutional Autoencoder. Furthermore, to integrate the multi-source signals, a Cascaded Multi-head Self-attention Autoencoder with Characteristic Guidance is proposed to construct health indicators. Then, the whole life cycle of REB is divided into different stages based on the Continuous Gradient Recognition with Outlier Detection. With the development of Measurement-based Correction Life Formula and Bidirectional Recursive Gated Dual Attention Unit, accurate life-cycle RUL prediction is achieved. Data from self-designed test rig and PHM 2012 Prognostic challenge datasets are analyzed with the proposed framework and five existing prediction models. Compared with the strongest prediction model among the five, the proposed framework demonstrates significant improvements. For the data from self-designed test rig, there is a 1.66 % enhancement in Corrected Cumulative Relative Accuracy (CCRA) and a 49.00 % improvement in Coefficient of Determination (R2). For the PHM 2012 datasets, there is a 4.04 % increase in CCRA and a 120.72 % boost in R2. [Display omitted] • Advocate explainable knowledge-data integration for RUL prediction throughout the life cycle, leveraging multi-sensor data. • Introduce a novel unsupervised learning framework that integrates IGCA and CMSACG to construct HI. • A two-stage prediction framework, fusing MCLF and BR-GDAU, is proposed for predicting the life cycle RUL. • The effectiveness of the proposed framework is validated using both Self-Designed Experiment-Derived and PHM 2012 Prognostic Challenge Bearing Datasets. [ABSTRACT FROM AUTHOR]

Corporate fraud risk detection is a branch of fraud. It may exist in various industries and cause economic problems. Effective identification of corporate fraud can protect the safety of funds for investors in some sense. This paper proposes a classifier model of a fractional-order immune BP neural network based on the self-attention mechanism to improve efficiency. The improved artificial immune algorithm with dynamic region contraction strategy is used to optimize the initialization process of the BP neural network. Furthermore, it combines the self-attention mechanism to design the input layer. Finally, Caputo fractional noncausal calculus is used to optimize the parameter updating process in BP neural network. The experiment results indicate that our model has fast convergence rate and powerful capacity of detection, and performs efficiently in detecting fraud behaviors. [ABSTRACT FROM AUTHOR]

Load Forecast (LF) is an important task in the planning, control and application of public power systems. Accurate Short Term Load Forecast (STLF) is the premise of safe and economical operation of a power system. In the research of short-term power load forecasting, machine learning and deep learning are the most popular methods at present, but there still exists a problem that the single and simple structure of power load forecasting model leads to low accuracy of load forecasting. In order to improve the accuracy of STLF, a Gated Cycle Unit (GRU)-Transformer combined neural network model is proposed. Transformer encoder structure is used as feature extractor to mine the complex mapping relationships between the input features and load. The advantage of self-attention mechanism is used to solve the problem of information loss of long sequences in short-term power load forecasting. At the same time, the multivariate time series model of GRU is used for model training. The experimental results on the power load data set of a certain region in southwest China and Panama City show that the proposed combined model prediction method has higher accuracy than those proposed in other literatures, which further proves its feasibility and superiority. [ABSTRACT FROM AUTHOR]

Traditional recognition methods often lead to problems such as speaker information loss and reduced recognition rates. To address these problems, an Fca-Res2Net speaker recognition model incorporating a self-attentive mechanism is proposed in this paper. First, the model uses the modified mel-frequency cepstral coefficients (MFCCs) as the system feature input and combines the inverse mel-frequency cepstral coefficients (IMFCCs) with the MFCCs as the base input features to extract more representative speech spectral features. On this basis, the difference parameters △MFCC and △IMFCC are fused to fully extract the speech dynamic and static features in the high- and low-frequency bands. Second, frequency channel attention networks (FcaNets) are introduced on top of the baseline model (Res2Net: a new multiscale backbone architecture), and the residual module is used to fuse the shallow and deep speaker features to better obtain the different feature channel weights without increasing the number of parameters. In addition, to better introduce temporal information and capture long-span speech features, the self-attention mechanism is integrated to enhance the long-span modelling of speech features. Finally, the classification output results are identified. Experimental results show that the proposed model improves the recognition rate and robustness of speakers in long speech when compared with the current mainstream speaker recognition methods in the VoxCeleb dataset with sufficient data volume. [ABSTRACT FROM AUTHOR]

Due to the lack of long-range association and spatial location information, fine details and accurate boundaries of complex clothing images cannot always be obtained by using the existing deep learning-based methods. This paper presents a convolutional structure with multi-scale fusion to optimize the step of clothing feature extraction and a self-attention module to capture long-range association information. The structure enables the self-attention mechanism to directly participate in the process of information exchange through the down-scaling projection operation of the multi-scale framework. In addition, the improved self-attention module introduces the extraction of 2-dimensional relative position information to make up for its lack of ability to extract spatial position features from clothing images. The experimental results based on the colorful fashion parsing dataset (CFPD) show that the proposed network structure achieves 53.68% mean intersection over union (mIoU) and has better performance on the clothing parsing task. [ABSTRACT FROM AUTHOR]

Financial risk management has always been a critical issue; banks, debt issuers, and government officials all need credit ratings in order to make intelligent financial decisions. Most of the existing studies on corporate credit rating prediction utilize financial statement features as their input data. Credit rating is closely related to credit risk. However, very few studies consider credit risk elements, such as credit systematic risk / beta and Credit Default Swap (CDS) spread data in credit rating prediction. Furthermore, the application of generative adversarial learning for corporate credit rating prediction was rarely investigated. In this work, a novel generative adversarial network (GAN), Self-Attention Recurrent Conditional GAN (SAR-CGAN) for corporate credit rating prediction is proposed. The proposed model takes advantage of Conditional GAN and Recurrent GAN to improve prediction performance. The financial statement features and corporates' CDS spread-related features: credit systematic risk / beta and quarter mean of CDS spread are used as input features. The proposed model adopts long short-term memory networks (LSTM) based on self-attention to process historical data and generate corporate credit rating. We improve the recurrent-based GAN model by modifying the network structure, in which the self-multihead attention layer is added to capture the weighted importance of the time series data. Moreover, a data sampling strategy is designed to alleviate the overfitting issue and enhance the effectiveness of the proposed GAN model. The experimental results indicate that the proposed model performs better than other state-of-art models on the applied datasets. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Harbin Institute of Technology. Social Sciences Edition / Haerbin Gongye Daxue Xuebao. Shehui Kexue Ban is the property of Harbin Institute of Technology and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Even though the Vision Transformer leverages the self-attention mechanism to capture long-range dependencies, showing significant potential in medical image segmentation, the limited annotations in the image dataset make it difficult for the Transformer model to extract different global features, resulting in attention collapse and generating similar or identical attention maps. Previous studies have attempted to solve the problem by integrating convolutional neural layers into Transformer-based architectures. However, improper integration may lead to the inability of the model to effectively capture local and global information in both spatial and channel dimensions. To address the above issue, we propose a hybrid architecture using the Dual Channel-Spatial Self-Attention Transformer and CNN Synergy Network (DTC-SUNETR) for medical image segmentation. Specifically, we redesigned the self-attention mechanism. A novel Channel-Spatial Self-Attention (CSSA) block is introduced to integrate the enhanced channel and spatial self-attention mechanism to capture the global relationship and local structure among image features. This helps the model to more comprehensively understand the inter-dependencies between different channels and capture the relationships between different pixels, thus enhancing the feature representation of the corresponding dimensions. Simultaneously, it also improves the overall computational efficiency of the network. Extensive experiments on four different medical image segmentation datasets, including Synapse, ACDC, Brain Tumor, and Lung Tumor, demonstrate the superiority of the proposed DTC-SUNETR over state-of-the-art methods. • DTC-SUNETR can effectively integrate the benefits of both CNNs and Transformers. • CSSA block is adopted to capture the interdependence of spatial and channel information. • DTC-SUNETR can comprehensively characterize the feature relationships using CSA and SSA. • The experimental results on four medical image segmentation datasets demonstrate the superiority of DTC-SUNETR. [ABSTRACT FROM AUTHOR]

Sentiment analysis of customer feedback is a pivotal component of Natural Language Processing (NLP), enabling businesses to gauge consumer emotions towards their products and services. The inherent variability of language introduces substantial challenges in the analysis of unstructured customer data. Sentiment analysis techniques generally fall into two categories: traditional methods, which involve the extraction of hand-crafted features and the application of machine learning algorithms, and end-to-end deep learning approaches that process raw data, transforming them layer by layer into high-level textual features. Each method presents its own balance between simplicity and speed versus analytical power and flexibility, and they all grapple with the need for extensive labeled data and computational resources. To address these challenges, this paper presents an innovative hybrid model that amalgamates RoBERTa, ResNeXt, BiLSTM, and self-attention mechanisms. This integration capitalizes on the collective strengths of these models to surmount the limitations inherent in each. The proposed model proves to be a powerful and efficient tool for sentiment analysis, demonstrating proficiency across various data types and tasks. We undertake a comprehensive evaluation of the proposed model's accuracy and training efficiency using four benchmark datasets. Our investigation also explores the impact of different similarity measures and convolutional neural network architectures on the model's performance. The results affirm that our model not only achieves high accuracy but also significantly reduces training time compared to RoBERTa's fine-tuning. Furthermore, the model exhibits exceptional domain adaptability, particularly when fine-tuned on the IMDb dataset following initial training on the Yelp Review Full dataset. The practicality of the proposed model is underscored by its reduced computational complexity and its adeptness at navigating semantic and syntactic nuances. • A novel hybrid sentiment model using RoBERTa, ResNeXt, BiLSTM, and self-attention. • Outperforms other methods on two datasets; matches RoBERTa base on two others. • Fewer parameters and faster fine-tuning compared to RoBERTa base. • Analysis of document similarity methods and text feature extractors. • TS-SS is viable for low-dimensional features but not high-dimensional ones. [ABSTRACT FROM AUTHOR]

Accurate water demand forecasting can help understand water usage dynamics, which has a potential application in water saving and demand management. Despite extensive research, most exiting methods cannot capture long-range dependency of water demand and maintain high-accuracy in multi-horizon forecasting continuously. To address these issues, the focus of this research is primarily on the temporal model for multi-horizon water demand forecasting using deep learning. This research proposes a signal enhancement strategy and develops an attention-based deep learning model. Firstly, the Fourier transform is used for sparse approximation of water demand data, which helps represent the signal more compactly. To enhance the performance of multi-horizon forecasting model, the complex water demand time series is decomposed into trend and seasonal components. Then, an attention mechanism is utilized to learn temporal dependencies within the data, providing assists for multi-horizon water demand forecasting. Furthermore, a comparative study is carried out between the proposed model and the state-of-the-art methods using the water demand data from four usage scenarios. Experimental results suggest that the present model has the ability to produce accurate and robust forecasting, especially for multi-horizon water demand. Therefore, this research has the potential to enhance water resource management in a socioecological context. • Stable and efficient multi-horizon forecasting of water demand was achieved. • Fourier transform was employed for noise reduction and compact representation. • A hierarchical structure was proposed to extract periodic and trend pattern. • Attention-based framework was applied to capture long-range dependencies. • This approach can be a promising method for sustainable water resource management. [ABSTRACT FROM AUTHOR]

Keywords are useful in natural language tasks. However, it is a challenge task to extraction keywords from short texts. In which the model may be subject to impaction of topic dependence and poor text organization structure. To resolve this limitation, we propose a keywords generation model ADGCN of short texts based on graph-to-sequence learning. The model to jointly short texts contextual feature and positional feature based adaptation for this task. We learn domain-invariant feature representations by using graph-building feature and node topic feature space, and jointly perform linear generate feature in framework of keywords decoding. Experiment results on real social datasets demonstrate that our proposed model achieves impressive empirical performance on relevance, information and coherence. Besides, the proposed ADGCN also outperforms the state-of-the-arts on public KP20k dataset. The experiments testify that the model can generate the topic keywords of short texts and effectively alleviate the influence of data disturbance. [ABSTRACT FROM AUTHOR]

Building extraction with high spatial resolution images becomes an important research in the field of computer vision for urban-related applications. Due to the rich detailed information and complex texture features presented in high spatial resolution images, the distribution of buildings is non-proportional and their difference of scales is obvious. General methods often provide confusion results with other ground objects. In this paper, a building extraction framework based on deep residual neural network with a self-attention mechanism is proposed. This mechanism contains two parts: one is the spatial attention module, which is used to aggregate and relate the local and global features at each position (short and long distance context information) of buildings; the other is channel attention module, in which the representation of comprehensive features (includes color, texture, geometric and high-level semantic feature) are improved. The combination of the dual attention modules makes buildings can be extracted from the complex backgrounds. The effectiveness of our method is validated by the experiments counted on a wide range high spatial resolution image, i.e., Jilin-1 Gaofen 02A imagery. Compared with some state-of-the-art segmentation methods, i.e., DeepLab-v3+, PSPNet, and PSANet algorithms, the proposed dual attention network-based method achieved high accuracy and intersection-over-union for extraction performance and show finest recognition integrity of buildings. [ABSTRACT FROM AUTHOR]

Quantitative analysis of tumor immune microenvironment (TIME) in immunohistochemical (IHC) tissue microarray (TMA) sections is crucial in diagnosis and treatment recommendations for cancer patients. Nuclei segmentation and classification are the prerequisites for the TIME quantification, but it still lacks of robust nuclear quantification models used for IHC histological slides. In this paper, we design an approach for simultaneously segmenting and classifying cell nuclei in multiplex IHC TMA sections. The large TMA tissue core is first divided into a set of small overlapping patches, where cell nuclei are then simultaneously segmented and classified by using our multi-task learning model. The model has one feature encoder with cascaded separable-ResUnit blocks, and three decoder branches that incorporate the Self-Attention modules and DenseUnit blocks to perform nuclear segmentation, classification and distance map regression, respectively. After processing all patches, the weighted loss map and vote mechanism are applied to seamlessly stitch patch-level predictions to form the tissue core level results. We finally exploit generalized Laplacian of Gaussian (gLoG) filters based algorithm to post-process segmentation results to further split overlapping cell nuclei. Quantitative evaluations have been performed on a IHC stained histological image dataset with 9725 manually identified cell nuclei and a public H&E stained dataset (CoNSep), which show that our model outperforms state-of-the-art nuclei segmentation and classification models. The qualitative evaluations on TMA sections show the potential of using our approach in clinical applications. • We design an integrated technical pipeline for nuclei segmentation and classification in IHC TMA sections, which includes patch-level predictions, seamless patch stitching and post-processing. • We design a multi-task learning model for nuclei simultaneous segmentation and classification. The network has an encoder–decoder architecture, where the encoder consists of cascaded separable ResUnit blocks, and three decoder branches perform nuclear segmentation, classification, distance map regression, respectively. • We perform experiments on self-built IHC dataset and another public H&E dataset, which show that our SRSA-Net outperforms other state-of-the-art comparative models. [ABSTRACT FROM AUTHOR]

• Development of a powerful Deep Learning model (DenseNet) for automatically classifying spinal osteosarcoma and giant cell tumors in medical images. • Integration of a self-attention mechanism and multi-scale feature map extraction to enhance feature extraction capabilities. • Use of Grad-CAM for improved visualization of tumor regions during predictions. • Significant support for orthopedic physicians in accurate diagnostic classification, aiding in treatment and care plan development. • Acknowledges the need for a larger dataset to improve model performance and its applicability in diverse clinical settings. This study aims to explore an optimized deep-learning model for automatically classifying spinal osteosarcoma and giant cell tumors. In particular, it aims to provide a reliable method for distinguishing between these challenging diagnoses in medical imaging. This research employs an optimized DenseNet model with a self-attention mechanism to enhance feature extraction capabilities and reduce misclassification in differentiating spinal osteosarcoma and giant cell tumors. The model utilizes multi-scale feature map extraction for improved classification accuracy. The paper delves into the practical use of Gradient-weighted Class Activation Mapping (Grad-CAM) for enhancing medical image classification, specifically focusing on its application in diagnosing spinal osteosarcoma and giant cell tumors. The results demonstrate that the implementation of Grad-CAM visualization techniques has improved the performance of the deep learning model, resulting in an overall accuracy of 85.61%. Visualizations of images for these medical conditions using Grad-CAM, with corresponding class activation maps that indicate the tumor regions where the model focuses during predictions. The model achieves an overall accuracy of 80% or higher, with sensitivity exceeding 80% and specificity surpassing 80%. The average area under the curve AUC for spinal osteosarcoma and giant cell tumors is 0.814 and 0.882, respectively. The model significantly supports orthopedics physicians in developing treatment and care plans. The DenseNet-based automatic classification model accurately distinguishes spinal osteosarcoma from giant cell tumors. This study contributes to medical image analysis, providing a valuable tool for clinicians in accurate diagnostic classification. Future efforts will focus on expanding the dataset and refining the algorithm to enhance the model's applicability in diverse clinical settings. [ABSTRACT FROM AUTHOR]

With the proliferation of radio frequency (RF) systems and radar applications, Electronic Warfare (EW) is receiving increasing importance. The analysis of the radar signals is a critical EW task that decides the nature of counter employments. In an Electronic Support (ES) system, the challenge is to detect hostile radiation sources efficiently and trigger a counter response. Detection of types of Pulse Repetition Interval (PRI) modulation of radar signal significantly facilitates the manifestation of RF emitters during recognition which is difficult in a dense EW environment. Recent developments in artificial intelligence (AI) methods suggest that this emerging technology can be effective for such purposes. In this direction, an automatic approach for recognizing several kinds of complex PRI modulation based on Continuous Wavelet Transform (CWT) and a combination of the vanilla Convolutional Neural Network (CNN), a multi-head self-attention (MHSA) mechanism and the popular Long Short-Term Memory (LSTM) is proposed. The CWT is used to decompose the PRI modulation sequence and obtain different time–frequency components. Further, aided by the proposed CNN-MHSA-LSTM combination, the features extracted from the CWT 2D-scalograms are used to execute PRI modulation discrimination. In this method, the vanilla CNN is employed for the extraction of deep features to figure out the class details while capturing the spatial attributes. Thereafter, to improve the discriminative power of the entire framework a MHSA mechanism is used. The temporal attributes are acquired by the LSTM which works in concert with the CNN for executing the detection of the PRI classes based on the extracted features. Also to assess the effectiveness of the proposed method, three models based on ResNet, popular CNN and SqueezeNet are implemented for benchmark comparison in terms of overall performance and complexity. The simulation results show that the proposed method enhances performance and achieves robustness in the noise-filled and imperfect channel knowledge environment. The best recognition accuracy is 98.3% with 50% spurious pulses in the environment which fluctuates with imperfect channel knowledge cases. [ABSTRACT FROM AUTHOR]

Stealthy FDIA (False Data Injection Attack) is a serious cyber threat that can modify state estimation of smart grid through maliciously altering the measurement data, but can't be detected by traditional bad data detection system in smart grid. There exist two weakpoints for numerous deep neural networks (DNNs) based data-driven schemes against FDIA. First, they mainly focus on detecting the presence of FDIA, but fail to localize the specific bus/nodes affected. Second, their performance is not sufficiently desirable under small attack, i.e., anomalies caused by the attack closely resemble normal data. To address the above issues, this paper proposes an effective locational FDIA detection framework based on data reconstruction, AT-GVAE, which seamlessly integrates the variational autoencoders (VAEs) and generative adversarial network paradigm. Specifically, our contributions are threefold. First, the proposed AT-GVAE framework is novelly composed of two main modules: the generative VAE_G and discriminative VAE_D that both play dual roles: reconstruct data from jointly learning distributions of data and latent feature space, and meanwhile play Minmax game with adversarial way. Second, multiple-layer gated recurrent units (GRUs) are utilized as the basic structure of the encoder and decoders in both VAEs, to characterize the temporal correlations of measurement data sequence. Additionally, the self-attention mechanism is used to enhance the expressive ability of GRU based VAEs. Then, the anomaly score for each busbar in the smart grid is determined by comparing the residual between the observed measurement and the outputs of VAE_G and VAE_D, enabling the localization of FDIA. Finally, thorough experiments on multiple power systems with series data of total 3456 measurements demonstrate that, in terms of multiple typical metrics, including false alarm rate vs. detection probability, AUC-ROC, and AUC-PR, our proposed framework outperforms the state-of-the-art GRU, VAE and adversarial training based FDIA detection schemes. • Propose a novel locational FDIA detection framework, seamlessly integrating the VAEs and GAN paradigm. • The generative and discriminative VAEs both have dual roles: reconstruct data and play minmax game with adversarial way. • Multiple-layer gated recurrent units are utilized as the basic structure of the encoder and decoders in both VAEs. • Demonstrate our proposal outperforms SOTA VAE and adversarial training based FDIA schemes on multiple power systems. [ABSTRACT FROM AUTHOR]

With the explosive growth of reviews on e-commerce websites, consumers usually have trouble finding helpful information in massive reviews. Therefore, predicting review helpfulness has drawn the attention of researchers in the research community. Previous works have focused on identifying the helpfulness of reviews based on content, but we found that the helpfulness of a review is determined by the content of reviews and the matching degree between items and reviews. To meet this challenge, this paper proposes a deep matching model based on a feature-aware external memory network (DMMN) to predict review helpfulness. Specifically, the DMMN contains a memory matrix that can store typical features of items via memory reading/writing operations. We then derive item features by properly reading the memory matrix. Meanwhile, a self-attention mechanism is adopted to obtain review features. Finally, the helpfulness score of the review is acquired by the matching degree between item features and review features. Experiments conducted on Amazon datasets, compared with the baseline models, the DMMN obtains an about 1–2% performance improvement on the eight datasets. The experimental results demonstrate that our efforts to predict helpful reviews are feasible and effective. • Review helpfulness prediction is treated as a matching process, considering review content versatility and item consistency. • An external memory network is introduced to extract and store features of items from reviews. • A deep matching model based on feature-aware external memory network is proposed to perform review helpfulness prediction. [ABSTRACT FROM AUTHOR]

Due to the instability and complex distribution of electroencephalography (EEG) signals and the great cross-subject variations, exploiting valuable and discriminative emotional information from EEG is still a significant challenging issue. In this paper, we propose Bi-Stream Multilayer Perceptron – Self-Attention Mixer (BiSMSM), a novel model for EEG-based emotion recognition. The proposed model consists of two streams: the spatial stream and the temporal stream. BiSMSM jointly captures the useful information from temporal, spatial, local and global angles, aiming to encode more discriminative features describing emotions. The spatial stream focuses on the spatial information, while the temporal stream concentrates on the correlation information in the time domain. The structures of the two streams are similar, either of which contains a Multilayer Perceptron (MLP) based module to extract the regional in-channel and cross-channel information. The MLP-based module is followed by a self-attention mechanism module to explore the global signal correlations. Finally, the subject-independent experiments on the public benchmark datasets DEAP and DREAMER have demonstrated the advantage of our model over the related advanced approaches. Specifically, BiSMSM obtains an accuracy of 63.10% for valence classification and 61.89% for aoursal classification on DEAP, and 61.88% for valence classification and 64.25% for arousal classification on DREAMER. • A novel and effective model, BiSMSM, is proposed for EEG-based emotion recognition. • BiSMSM can capture the useful information from temporal, spatial, local and global angles. • In BiSMSM, MLP-SA Mixer is constructed by the MLPBlocks and the self-attention mechanism module. • Experiments on benchmarks DEAP and DREAMER have demonstrated the advantage of BiSMSM. [ABSTRACT FROM AUTHOR]

Sleep is crucial for human health. Automatic sleep stage classification based on polysomnography (PSG) is meaningful for the diagnosis of sleep diseases, which has attracted extensive attention recently. Most existed works learn the representations of PSG based on the conventional grid data structure, ignoring the topological channel correlations among different body regions in non-Euclidean space. To this end, we propose a deep learning model named SAGSleepNet, which transforms PSG into a robust graph structure by neural networks and self-attention mechanism, utilizing graph convolution network (GCN) and bidirectional gated recurrent unit (BiGRU) to capture epoch-level topological channel correlations and sequence-level temporal sleep transitions of PSG, respectively. The final learned representation is fed into a softmax layer to train an end-to-end model for automatic sleep staging. Experimental results on three public datasets (i.e., SleepEDFx, UCD, CAP) show that SAGSleepNet obtains the best classification performance compared with several baselines, including accuracy of 0.807 ,0.765, and 0.746, F1-score of 0.744, 0.750, and 0.695, and Cohen's Kappa (κ) of 0.721, 0.690, and 0.660. In general, our work provides a reasonable graph structure to model PSG epoch, and makes contribution to explore implicit channel-wise information of PSG by deep learning techniques. [Display omitted] • Automatic sleep staging based on polysomnography is meaningful for sleep medicine. • Topological channel correlations of polysomnography is close to body correlations. • Transforming polysomnography to robust graph structure by self-attention mechanism. • Graph convolution network effectively learns the topological information of graph. • The proposed hierarchical deep neural networks improves automatic sleep staging. [ABSTRACT FROM AUTHOR]

Combining transformers and convolutional neural networks is considered one of the most important directions for tackling medical image segmentation problems. To learn the long-range dependencies and local contexts, previous approaches embedded a convolutional layer into feedforward neural network inside the transformer block. However, a common issue is the instability during training since large differences in amplitude across layers by pre-layer normalization. Furthermore, multi-scale features were directly fused using the transformer from the encoder to decoder, which could disrupt valuable information for segmentation. To address these concerns, we propose A dvanced T rans Former (ATFormer), a novel hybrid architecture that combines convolutional neural networks and transformers for medical image segmentation. First, the traditional transformer block has been refined into an Advanced Transformer Block, which adopts post-layer normalization to obtain mild activation values and employs the scaled cosine attention with shifted window for accurate spatial information. Second, the Progressive Guided Fusion module is introduced to make multi-scale features more discriminative while reducing the computational complexity. Experimental results on the ACDC, COVID-19 CT-Seg, and Tumor datasets demonstrate the significant advantage of ATFormer over existing methods that rely solely on convolutional neural networks, transformers, or their combination. [Display omitted] • Propose ATFormer with a hybrid CNN and Transformer for medical image segmentation. • Present novel transformer block with post-LN and the scaled cosine attention. • Develop PGF module to avoid semantic gaps and balance accuracy and efficiency. • Contrast experiments on three different segmentation tasks prove the superiority. [ABSTRACT FROM AUTHOR]

The accurate identification of rice varieties using rapid and nondestructive hyperspectral technology is of practical significance for rice cultivation and agricultural production. This paper proposes a convolutional neural network classification model based on a self-attention mechanism (self-attention-1D-CNN) to improve accuracy in distinguishing between crop species in fields using canopy spectral information. After experimental materials were planted in the research area, portable equipment was used to collect the canopy hyperspectral data for rice during the booting stage. Five preprocessing methods and three extraction methods were used to process the data. A comparison of the classification accuracy of different classification models showed that the self-attention-1D-CNN proposed in this study achieved the best classification with an accuracy of 99.93%. The research demonstrated the feasibility of using hyperspectral technology for the fine classification of rice varieties, and the feasibility of using the CNN model as a potential classification method for near-ground crop monitoring and classification. • A Self-attention-1D-CNN model was proposed for fine classification with the accuracy of 99.93%. • A high-precision, multi-variety classification task was achieved on rice canopy scale. • Near-ground crop classification was completed using only spectral reflectance information. • The feasibility of hyperspectral technique (rapid, nondestructive) for rice variety classification was explored. [ABSTRACT FROM AUTHOR]

• Spatial correlation is elaborately quantified from data and hydraulic perspectives. • Spatial correlation provides weights for aggregating neighboring sensors' data. • Neighborhood information aggregation extracts richer input features for prediction. • Weighting strategy can enhance performance especially for multi-step prediction. • Integrating priori hydraulic knowledge into machine learning makes better forecast. Graph convolutional network (GCN) has been used to capture spatial correlation between multiple sensors for better performance of short-term water demand forecasting, which is essential for the implementation of smart water such as optimal scheduling and anomaly detection. However, the GCN assumes every sensor's importance to be the same and describes the spatial correlation purely from a data perspective. To resolve the two issues that affect prediction accuracy, this study proposes a weighting strategy and develops a spatial correlation-based GCN (SCGCN) prediction model. Self-attention mechanism is used to comprehensively analyze water demand data of every sensor and flow resistance between sensors (i.e., head loss along pipes), generating spatial correlation coefficients with the consideration of hydraulics. Then the coefficients are used as weights to aggregate neighboring sensors' data, extracting accurate input features for the SCGCN model. The study utilized real monitoring data to develop the SCGCN model and compare it with a traditional artificial neural network (ANN) model. Results show that the SCGCN model can outperform the ANN model, especially for multi-step prediction. The root mean square errors of 30, 45, and 60 min multi-step prediction cases are reduced by 4.4% to 9.2% for different sensors. [ABSTRACT FROM AUTHOR]

Artificial intelligence has attracted great attentions in the field of traffic flow forecasting. Graph neural network (GNN), as one of the most popular methods, capable of processing non-Euclidean graph-structured data, which makes it suitable for traffic forecasting problems. However, existing GNN based methods highly rely on fixed coefficients calculated from spatial information, while practical changes of lanes, surrounding environment, and infrastructure conditions may lead to dynamic spatial dependences. Therefore, this paper proposes a dynamic graph attention network model (TransGAT), which includes two phases. Firstly, Node Embedding algorithm based on dynamic attention mechanism is proposed to capture spatial dependency information. Then, the temporal features captured by the Gated Temporal Convolutional Network (TCN) is added by using Skip Connection so as to construct a spatio-temporal capture module. TransGAT model is further validated in the two real datasets, METR-LA and PEMS-BAY. The results show that TransGAT model shows better performances in the short-term (15 min), medium-term (30 min) and long-term (60 min) forecasting. Compared with the baseline method, the forecasting error is reduced by about 13%. • A combined Gated Temporal Convolutional Network and Node Embedding with dynamic attention mechanism is proposed. • The proposed model requires less prior knowledge and can be easily extended to other spatio-temporal forecasting tasks. • The model improves the prediction capabilities for short, medium and long-term traffic flow forecasting. [ABSTRACT FROM AUTHOR]

• We build a new benchmark handwashing action dataset. • We apply self-attention convolutional neural network for handwashing recognition. • The proposed approach is able to achieve more accurate recognition results. Automatic hand hygiene monitoring is an effective and necessary practice in healthcare. The main objective is to automate the quality control of the hand hygiene process by recognizing handwashing actions. This paper proposes a new computer vision-based hand hygiene monitoring approach, with the aim to recognize handwashing actions from the input videos. The proposed approach consists of the following three modules. First, in the feature extraction module , a backbone convolutional neural network (CNN) model (i.e., a ResNet model) is used to extract features from each frame of the input video. Second, in the feature aggregation module , the self-attention mechanism is introduced to form six masked self-attention blocks that are stacked to aggregate features from multiple frames. Third, in the feature classification module, the self-attentive aggregated feature representation is used for handwashing action recognition. The proposed approach outperforms a few state-of-the-art handwashing action recognition approaches. The proposed approach is able to perform effective and accurate handwashing action recognition for automatic hand hygiene monitoring, as verified by three benchmark datasets. [ABSTRACT FROM AUTHOR]

1 A novel quality-related process monitoring method named Self-attention-based Multi-block regression fusion Neural Network (SMNN) is proposed. 2 The strategy of pre-training and fine-tuning is used to extract sub-block features. 3 The paper considers the correlation between sub-blocks in the modeling process. For plant-wide process with multiple operation units, local-global modeling is an efficient method to achieve quality-related fault detection. However, most of algorithms based on local-global modeling ignore the correlation between sub-blocks. This will result in poor performance of the extracted global quality-related features. This paper focus on the correlation between sub-blocks and proposes Self-attention-based Multi-block regression fusion Neural Network (SMNN) to achieve efficient quality-related fault detection for nonlinear multi-unit process. Firstly, to focus on quality-related information, the key variables are selected. Then, to extract quality-related features in each sub-block, a pre-training approach is used, i.e. a deep neural network-based regression network between process variables and quality variables is constructed in each sub-block. Secondly, considering the correlation between the sub-blocks, self-attention mechanism is used to integrate the quality-related feature from each block. With the help of an additional regression network, the quality-related features of sub-blocks are fine-tuned and the global features are extracted. Finally, quality-related statistic is constructed to detect faults. The proposed method shows good performance in Tennessee-Eastman process, which demonstrates the effectiveness of the method. It also shows that considering the potential relationships between sub-blocks during model construction helps in the extraction of global features. [Display omitted] [ABSTRACT FROM AUTHOR]

Recent years have witnessed growing interests in online incremental learning. However, there are three major challenges in this area. The first major difficulty is concept drift, that is, the probability distribution in the streaming data would change as the data arrives. The second major difficulty is catastrophic forgetting, that is, forgetting what we have learned before when learning new knowledge. The last one we often ignore is the learning of the latent representation. Only good latent representation can improve the prediction accuracy of the model. Our research builds on this observation and attempts to overcome these difficulties. To this end, we propose an Adaptive Online Incremental Learning for evolving data streams (AOIL). We use auto-encoder with the memory module, on the one hand, we obtained the latent features of the input, on the other hand, according to the reconstruction loss of the auto-encoder with memory module, we could successfully detect the existence of concept drift and trigger the update mechanism, adjust the model parameters in time. In addition, we divide features, which are derived from the activation of the hidden layers, into two parts, which are used to extract the common and private features respectively. By means of this approach, the model could learn the private features of the new coming instances, but do not forget what we have learned in the past (shared features), which reduces the occurrence of catastrophic forgetting. At the same time, to get the fusion feature vector we use the self-attention mechanism to effectively fuse the extracted features, which further improved the latent representation learning. Moreover, in order to further improve the robustness of the algorithm, we add the de-noising auto-encoder to original framework. Finally, we conduct extensive experiments on different datasets, and show that the proposed AOIL gets promising results and outperforms other state-of-the-art methods. • We construct the autoencoder with memory module to detect the concept drift. • Our approach extracts effective hidden information and uses self-attention to fuse them. • Our method reduces the problem of catastrophic forgetting by extracting shared and private features of input. [ABSTRACT FROM AUTHOR]

To date, deep learning techniques, especially the combination of convolutional neural networks and recurrent neural networks with the attention mechanism, have been the state-of-the-art solutions for processing relation extraction and classification tasks. However, the neural network model constructed by this method cannot make full use of the labeled entities and their positional information in the relation classification, or even performs poorly on the small sample dataset. To address these issues, this paper proposes an entity relationship classification model BG-SAC, which combines BiGRU, Self-Attention mechanism and Capsule Networks. BG-SAC primarily uses BiGRU to obtain sentence sequential information and context-based semantic information, and then is coupled with the Self-Attention mechanism to get the correlation between words. Capsule Networks are used to acquire the positional information of entities. Eventually the probability that entities belong to a certain relationship category is calculated through the length of a capsule, so as to determine the relationship between entities and realize the classification of entity relationship. The experimental results show that the proposed model can effectively capture the word positional information and improve the classification effect with small sample datasets. • A relation classification model is formed by combining self-attention mechanism and capsule network. • The model captures the location information of entities and highlight the entity pair information. • Even on a small annotated dataset, the model performs well in relation classification. • Compared with the comparative models, the proposed model improves the accuracy of classification. [ABSTRACT FROM AUTHOR]

The biomedical literature contains a sufficient number of chemical-protein interactions (CPIs). Automatic extraction of CPI is a crucial task in the biomedical domain, which has excellent benefits for precision medicine, drug discovery and basic biomedical research. In this study, we propose a novel model, BERT-based attention-guided capsule networks (BERT-Att-Capsule), for CPI extraction. Specifically, the approach first employs BERT (Bidirectional Encoder Representations from Transformers) to capture the long-range dependencies and bidirectional contextual information of input tokens. Then, the aggregation is regarded as a routing problem for how to pass messages from source capsule nodes to target capsule nodes. This process enables capsule networks to determine what and how much information need to be transferred, as well as to identify sophisticated and interleaved features. Afterwards, the multi-head attention is applied to guide the model to learn different contribution weights of capsule networks obtained by the dynamic routing. We evaluate our model on the CHEMPROT corpus. Our approach is superior in performance as compared with other state-of-the-art methods. Experimental results show that our approach can adequately capture the long-range dependencies and bidirectional contextual information of input tokens, obtain more fine-grained aggregation information through attention-guided capsule networks, and therefore improve the performance. [ABSTRACT FROM AUTHOR]