Search

With the advancement of text-generation technology, misuse has increasingly challenged academic research sustainability. The Chinese academic community, vast and active with millions of researchers and extensive literature, faces an inevitable generator misuse. However, research on the automatic detection of Chinese academic texts remains scarce, necessitating the thorough exploration of detection methods to support ongoing academic development. This study explored automatic detection technology for Chinese academic texts, focusing on TK2A dataset construction, generation models, and detection methods to assess their practical impact. TK2A covers papers across disciplines, such as computer science, engineering, and medicine, ensuring broad applicability and forming a solid foundation for model training and evaluation. Using advanced natural language processing, models trained on TK2A showed strong performance across disciplines. Rigorous manual evaluation verified their reliability in terms of grammar, semantics, and logic. The study employed the widely adopted BERT model for detection, achieving high accuracy in distinguishing human-written content from AI-generated content on TK2A. This research underscores TK2A’s practical value by offering crucial support to journals with an accuracy exceeding 84%, institutions, and education in swiftly detecting AI-generated content, preventing misconduct, and enhancing academic publication quality.

With the continuous development of wireless technology, automatic modulation recognition plays an increasingly prominent role in military and civilian fields. However, the complex communication environment and diversified strategy bring challenges to modulation signal recognition. Therefore, an automatic modulation signal recognition technique combining truncated migration processing and convolutional neural network is proposed. Then multi-task learning is used to optimize and distinguish easily confused modulated signals. Three datasets, RadioML2016.10A, RadioML2016.10B and RadioML2016.04C are used for experiments. The results show that the modulation signal automatic recognition method proposed in this paper has good recognition accuracy. When the signal-to-noise ratio was 14dB, it reached a maximum of 95.46%. In addition, the Floating Point Operations of the proposed method were 1.71G, the number of parameters was 25636712, and the operation time was 228s, which showed that the method was light in weight and high in calculation efficiency. The optimized recognition technology can effectively distinguish three groups of easily confused signals, and the recognition rate is as high as 100%, and the minimum is not less than 90%. The proposed modulated signal automatic recognition technology has achieved remarkable results in improving the recognition accuracy and processing efficiency, which provides a strong technical support for the development of wireless communication technology, and also provides a new tool for researchers and engineers in related fields.

The spinal cord is an important part of the central nervous system, responsible for transmitting nerve signals throughout the body. The cervical spinal cord contains eight nerve bundles located in the neck region of the spinal cord that transmit to the face and head region. For this reason, in addition to traditional methods of monitoring changes in the spinal cord region in routine clinical practice, spinal cord segmentation using innovative computer-based systems makes an important contribution to the understanding of disease progression. Lesions in the cervical spinal cord can be a symptom of several neurological diseases, especially demyelinating diseases such as multiple sclerosis (MS). The detection of lesions in the spinal cord is particularly important in diseases such as MS, which affect a wide age range and for which early diagnosis is crucial. Therefore, automated segmentation of the spinal cord to quantify spinal cord atrophy is critical for changes in the human spinal cord. In addition to clinical findings, magnetic resonance imaging (MRI) technologies have improved the quality of images for monitoring, diagnosing and determining the treatment protocol for MS lesions in the spinal cord. However, due to the difficulty of scanning the cervical spinal cord region and the occurrence of artefacts during acquisition, it is very difficult to determine the spinal cord boundaries and detect lesions in this region. In this study, we propose a fractal network-based U-Net (FractalSpiNet) deep learning architecture for automatic segmentation of the spinal cord and spinal cord MS lesions from cervical spinal cord MR slices. The developed FractalSpiNet architecture incorporate a fractal network for enhanced feature extraction in MRI scans. In addition, a new dataset of axial plane MR images from the cervical spinal cord of 87 MS patients is first created in the study. Using the proposed FractalSpiNet architecture, the cross-sectional area of the cervical spinal cord was segmented with a Dice Similarity Coefficient (DSC) score of 98.88%, while MS lesions in the cervical spinal cord were detected with a DSC score of 90.90%. These results indicate that FractalSpiNet provides results that close to expert mask for segmentation of cervical spinal cord and MS lesion detection. The experimental studies also compare the results of the proposed FractalSpiNet with the results of state-of-the-art hybrid U-Net models such as base U-Net, Attention U-Net, Residual U-Net, and Attention Residual U-Net. In conclusion, the experimental results demonstrate the effectiveness of our approach in achieving accurate segmentation of cervical spinal cord and MS lesions, outperforming state-of-the-art methods. The proposed FractalSpiNet offers a promising approach for automated segmentation of the cervical spinal cord and MS lesions, potentially aiding in the diagnosis and treatment of neurological disorders.INDEX TERMS Cervical spinal cord, multiple sclerosis, automatic segmentation, fractal networks, U-Net, FractalSpiNet.

Industrial control systems are facing ever-increasing security challenges due to the large-scale access of heterogeneous devices in the open Internet environment. Existing anomaly detection methods are mainly based on the priori knowledge of industrial control protocols (ICPs) whose protocol specifications, communication mechanism, and data format are already known. However, when these knowledge are blank, namely, unknown ICPs, existing methods become powerless to detect the anomaly data. To tackle this challenge, we propose a multigranularity feature automatic marking-based deep learning method to classify unknown ICPs for anomaly detection. First, to obtain the feature sequences without priori knowledge assisting, we propose a multigranularity feature extraction algorithm to extract both byte and half-byte information by fully utilizing the intensive key information in the header field of the application layer. Then, to label the feature sequences for deep learning, we propose a feature automatic marking algorithm that utilizes the inconsistency feature sequences to dynamically update the feature sequence set. With the labeled feature sequences, we employ deep learning with 1-D convolutional neural network and gated recurrent unit to classify the unknown ICPs and realize anomaly detection. Extensive experiments on two public datasets show that both the accuracy and precision of the proposed method reach above 98.4%, which is better than the three benchmark methods.

Previous research on automatic control of high-speed trains in speed limit sections is insufficient. This article proposes a new offline reinforcement learning strategy for automatic tracking of autonomous trains. Firstly, the operating speed and deceleration starting point were determined for different speed limit scenarios. Then, a tracking controller based on the improved offline conservative Q-learning (CQL) algorithm was designed to avoid frequent interaction between the train and the environment. Selected an appropriate policy to implement the CQL algorithm. The data samples were reclassified to increase sample concentration. The value and strategy network structure was redesigned. The state space and action space of tracking trains were limited, and the dimension of state space was increased. A multi-objective reward function was designed to distinguish the tracking process of trains in different sections. The simulation results show that the proposed high-speed railway tracking interval automatic control algorithm is superior to traditional online reinforcement learning methods in terms of safety, comfort, and convergence efficiency.

Computer-aided interpreting(CAI) systems are software applied to one or more stages of interpreting tasks, which can directly promote the work of interpreters and improve the interpreting quality. Until now, studies on CAI in simultaneous interpretation(SI) have been limited, primarily focusing on the design and development of manual extraction models. Moreover, what’s particularly noteworthy is the lack of thorough investigation into the development of fully automated CAI models for extracting terms (SI difficulties) and other related aspects. Based on the experimental research of existing ones, this study puts forward some new methods based on automatic speech recognition(ASR) and develops a CAI system–InterpretSIMPLE with user-friendly interface, which implements automatic retrieval and display of terms (pre-imported), numbers, etc. as well as other functions specialized for conference interpreters. Through the setting of three-line label control and underline panel control, the system realizes the attention allocation and positioning of the source text content at different levels. One-click import of commonly-used Excel glossary gives simple operation with no additional format conversion. Terminologies and numbers are displayed below the corresponding position while displaying the source text, so that interpreters could locate and solve these recognized SI difficulties. Through the “exact matching” or “partial matching” setting, it could meet the personalized requirements of terms matching. The experiment shows that after the system receives text information from Tencent Cloud, the real-time display rate of the pre-imported glossary reaches 98.92%. The research results could provide references for the research and development of in-process automated CAI tools.

The customized automatic transverse robots play important roles in transporting heavy printing material rolls in automatic and intelligent flexible packaging printing workshops. A three-dimensional (3D) real-time visualization monitoring and feedback system of the customized automatic transverse robot is developed by adopting digital twin technology. This system is based on the theoretical framework of the data-driven digital twin model, and the main functions work through the physical layer, digital twin layer, and functional layer. The twin model is established and connected through the parent-child matching relationship of the components and it is driven by the real-time data which are acquired, transmitted and stored through communication protocols, such as OPC UA and RS-485. The physical automatic transverse robot and the virtual world of digital twins are synchronously mapped. The data are transferred from the virtual world twin model to the real world by establishing the application programming interface (API), implementing a feedback regulation mechanism that controls reality with virtuality. Combined with the long short-term memory (LSTM) model, the prediction of the failure trend of automatic transverse robot during the motion process has been achieved. A prototype digital twin system is developed for validation, which completes the closed-loop system of digital twin theory and proves the feasibility and effectiveness of the proposed system.

The contextual understanding ability in complex conversation scenarios has been a challenging issue, and existing methods mostly failed to possess such characteristics. To bridge such gap, this paper formulates a novel composite large language model to investigate such issue. As a result, taking English context as the scene, a Transformer-BERT integrated model-based automatic conversation model is proposed in this work. Firstly, the unidirectional BERT-based automatic conversation model is improved by introducing attention mechanism. It is expected to enhance feature expression for conversation texts by linking context to identify long-difficult sentences. Besides, a bidirectional Transformer encoder is utilized as the input layer before the BERT encoder. Through the two modules, dynamic language training based on English situational conversations can be completed to build the automatic conversation model. The proposed conversation model is further assessed on massive real-world English language context in terms of conversation performance. The experimental results show that compared with traditional rule-based or machine learning methods, the proposal has significantly improved response quality and fluency in English context. It can more accurately understand context, capture subtle semantic differences, and generate more coherent responses.

Studying the structure and function of the heart through the left ventricle is one of the most common methods for diagnosing heart diseases. The automatic segmentation of the left ventricle can be achieved through deep learning techniques, and researchers have conducted a series of explorations in this field. Recently, the segment anything model (SAM) has achieved significant success in the field of natural images, sparking considerable interest among researchers. This has led them to investigate whether SAM can also be successfully applied in the medical imaging domain. The SAM model’s interactive interface enables zero-shot and few-shot learning in the natural image domain, achieving accurate segmentation tasks. However, there are certain limitations in the automatic segmentation of medical images, specifically in the context of natural image cues such as points, boxes, and text prompts. To address this issue, this paper explores the performance of a prompt-free SAM-related model with an attention module for automatic segmentation of the left ventricle in echocardiography, named as SAM-Att. The model employs a low-rank fine-tuning strategy in the upstream, introduces an attention mechanism in the downstream, and successfully accomplishes the automatic segmentation task of the left ventricle with the support of weight files pretrained on the SAM large model. The SAM-Att model achieves dice similarity coefficient (DSC) of 94.49% and hausdorff distance (HD) of 3.505 mm on the test set. The accuracy reaches 98.83%, with precision of 93.65% and recall of 94.77%. A performance comparison of SAM-Att with other SAM-related models (SAM-b, MSA, Sam-CNN, AutoSAM, SAMed) is conducted on the same echocardiography dataset. The results indicate that the left ventricle automatic segmentation achieved the best performance when using SAM-Att.

Chromosome karyotyping is an important yet labor-intensive procedure for diagnosing genetic diseases. Automating such a procedure drastically reduces the manual work of cytologists and increases congenital disease diagnosing precision. Researchers have contributed to chromosome segmentation and classification for decades. However, very few studies integrate the two tasks as a unified, fully automatic procedure or achieved a promising performance. This paper addresses the gap by presenting: 1) A novel chromosome segmentation module named ChrRender, with the idea of rendering the chromosome instances by combining rich global features from the backbone and coarse mask prediction from Mask R-CNN; 2) A devised chromosome classification module named ChrNet4 that pays more attention to channel-wise dependencies from aggregated informative features and calibrating the channel interdependence; 3) An integrated Render-Attention-Architecture to accomplish fully automatic karyotyping with segmentation and classification modules; 4) A strategy for eliminating differences between training data and segmentation output data to be classified. These proposed methods are implemented in three ways on the public Q-band BioImLab dataset and a G-band private dataset. The results indicate promising performance: 1) on the joint karyotyping task, which predicts a karyotype image by first segmenting an original microscopical image, then classifying each segmentation output with a precision of 89.75% and 94.22% on the BioImLab and private dataset, respectively; 2) On the separate task with two datasets, ChrRender obtained AP50 of 96.652% and 96.809% for segmentation, ChrNet4 achieved 95.24% and 94.07% for classification, respectively. The COCO format annotation files of BioImLab used in this paper are available at https://github.com/Alex17swim/BioImLab The study introduces an integrated workflow to predict a karyotyping image from a Microscopical Chromosome Image. With state-of-the-art performance on a public dataset, the proposed Render-Attention-Architecture has accomplished fully automatic chromosome karyotyping.

Context: Facial recognition is one aspect of research that still has broad potential for research and development, especially as a security system for automatic border control. There is a significant continuous need to understand the characteristics of system development by considering system complexity and implementation environmental conditions. Objective: This research aims to provide in-depth insight and assist researchers and practitioners in developing large-scale facial detection systems for automatic border control. It has a high level of complexity that necessitates special attention to several factors such as real-time system, privacy, variations in facial features, quantity of data, model, and implementation environment. Method: This study used a systematic literature review as a research methodology by Kitchenham. The analysis was based on studies published between 2019 and 2023 on using facial recognition in autonomous border control. A systematic analysis of research was conducted by examining 112 scientific studies from 7884 papers in scientific databases. Result: Based on research questions, 12 types of threats are often encountered in ABC face recognition, which can be seen in section IV. The method most widely used is deep learning, especially for detecting emotional features and morphing attacks. Apart from that, most datasets used are private because they require collaboration with organizations and are related to privacy. Three remaining issues are encountered in this research, including face recognition methodology, privacy, and architecture for large-scale development. Future directions: This study suggests two future research topics to enhance achieving desired results in large-scale and complex advancements in a methodical and structured while upholding privacy ethics.

Magnetic actuation is particularly useful due to its deep penetration and safety to control microrobots for biomedical purposes. This paper proposes an automatic navigation algorithm using a magnetic potential field in the form of a Field Free Point (FFP). The FFP is a powerful tool for magnetic actuation because of its efficiency and versatility, and it can also be used in Magnetic Particle Imaging (MPI). The FFP provides efficiency and versatility in imaging and actuation and the ability to perform both functions on a single device without additional hardware. We suggest a novel navigation scheme based on FFP and Magnetic Potential Field (MPF) to achieve automatic navigation. It consists of a modified attractive field with velocity-based disturbance cancellation and a repulsive field to avoid collision. The suggested algorithm, leveraging microrobot movements and effectively addressing disturbance as inertia, is characterized by its capability to reduce deviations. Thus, simulations were conducted to tune the parameters and evaluate the proposed actuation scheme’s effectiveness. Also, the entire algorithm’s performance was verified through experiments in a dynamic environment with moving fluid and obstacles. The results demonstrated the proposed navigation system’s robustness in reaching the micro-robot to its destination even in an environment with a peak fluid velocity of $80 mm/s$ .

This paper presents a novel methodology for enhancing Automatic Speech Recognition (ASR) performance by utilizing contrastive learning to filter synthetic audio data. We address the challenge of incorporating synthetic data into ASR training, especially in scenarios with limited real-world data or unique linguistic characteristics. The method utilizes a contrastive learning model to align representations of synthetic audio and its corresponding text transcripts, enabling the identification and removal of low-quality samples that do not align well semantically. We evaluate the methodology on a medium-resource language across two distinct datasets: a general-domain dataset and a regionally specific dataset characterized by unique pronunciation patterns. Experimental results reveal that the optimal filtering strategy depends on both model capacity and dataset characteristics. Larger models, like Whisper Large V3, particularly benefit from aggressive filtering, while smaller models may not require such stringent filtering, especially on non-normalized text. This work highlights the importance of adjusting synthetic data augmentation and filtering to specific model architectures and target domains. The proposed method, robust and adaptable, enhances ASR performance across diverse language settings. We have open-sourced the entire work, which includes 140 hours of synthetically generated Portuguese speech, as well as the pipeline and parameter settings used to create these samples. Additionally, we provide the fine-tuned Whisper models and the code required to reproduce this research. Our code will be available at https://github.com/my-north-ai/semantic_audio_filtering.

Structured light 3D cameras have generally been the preferred choice for depth imaging in case high accuracy and precision of point clouds need to be captured at relatively short working distances. However, in environments featuring wide variations in surface reflectivity and ambient lighting, they face difficulty obtaining point clouds that sufficiently cover all the surface areas of interest without partial omissions. This happens not only with a single exposure but even with multiple exposures when a limited range of exposure settings is available. Although automatic multi-exposure settings have been explored for use in structured light 3D cameras to address this issue, there is yet to be a rigorous optimal solution for automatically setting multiple exposures to determine the minimum number of exposure settings while maximally covering the ill-conditioned surface of interest. In the current study, we present a method for optimal multi-exposure settings that automatically determines the minimum number of exposure settings for capturing the maximum number of 3D points from the surface of interest under a limited range of available exposure. The proposed method, which applies to any extremity in surface reflectivity and ambient light variations, is based on finding the minimum number of exposures that enable the identification of exposure-intensity slopes for as many pixels in the surface of interest as possible under projected light while removing the effect of ambient light. The proposed method was experimentally shown to have superior performance based on a case study and a performance comparison with a state-of-the-art multi-exposure setting practice.

Automatic essay scoring (AES) utilizes computational methods to evaluate written essays, offering an efficient approach to grading in various educational contexts. Essays can have multiple scoring criteria. Instead of building separate systems for each criterion, it’s better to have a single system capable of scoring multiple criteria. This paper introduces a cost-effective and efficient approach to Arabic automated essay scoring, focusing on the evaluation of grammaticality. The proposed approach can also be used to score essays based on other criteria such as similarity to model answer, organization, and adherence to prompt. The approach leverages the pre-trained AraBART model, employing different parameter-efficient methods. We start with Parameter-Efficient Fine-Tuning (PEFT) which optimizes a minimal set of parameters for each specific criterion. Then we apply additional parameter-efficient strategies such as Model Soup, Multi-Round Inference, and Edit Merging. To validate the effectiveness of our approach, we conduct experiments on multiple datasets. The primary objective is to enhance the grammatical correctness of texts and subsequently assess them based on the number of errors. We evaluate the approach on QALB-2014, QALB-2015, and ZAEBUC datasets. The results show similar and sometimes better performance than full fine-tuning while reducing computational and hosting costs. This efficiency is attributed to the utilization of parameter-efficient strategies that entail minimal additional parameters. By scoring essays based on multiple criteria, our approach offers a versatile and flexible tool for educators and language professionals, facilitating efficient evaluation of Arabic essays.

Recent studies have utilized deep learning (DL) techniques to automatically extract features from synthetic aperture radar (SAR) images, which shows great promise for enhancing the performance of SAR automatic target recognition (ATR). However, our research reveals a previously overlooked issue: SAR images to be recognized include not only the foreground (i.e., the target), but also a certain size of the background area. When a DL-model is trained exclusively on foreground data, its recognition performance is significantly superior to a model trained on original data that includes both foreground and background. This suggests that the presence of background impedes the ability of the DL-model to learn additional semantic information about the target. To address this issue, we construct a structural causal model (SCM) that incorporates the background as a confounder. Based on the constructed SCM, we propose a causal intervention-based regularization method to eliminate the negative impact of background on feature semantic learning and achieve background debiased SAR-ATR. The proposed causal interventional regularizer can be integrated into any existing DL-based SAR-ATR models, mitigating the impact of background interference on the feature extraction and recognition accuracy without affecting the testing speed of these models. Experimental results on the moving and stationary target acquisition and recognition and SAR-AIRcraft-1.0 datasets indicate that the proposed method can enhance the efficiency of existing DL-based methods in a plug-and-play manner.

With the development of the metro system, an increasing number of new metro lines are being constructed. However, the lack of historical data on the new metro lines leads to the challenge of predicting the origin-destination distribution of the whole network. Based on the advantage of Automatic Fare Collection data, this study proposes a novel method to predict origin-destination distribution over metro networks with expansions. Firstly, the disaggregate model is built to predict the origin-destination distributions of the whole network with consideration of passengers’ destination choice behavior. Then, the aggregate model which can capture the historical passenger distribution patterns is developed to correct the origin-destination distribution in existing metro lines. Finally, this study combines the results of the disaggregate model and aggregate model to obtain the whole network origin-destination distribution. This study applies the method to the Guangzhou Metro network, and the results show the superiority of the proposed model. The results can help to improve the operation and management of urban rail transit.

With the increasing prevalence of short videos, viewers are increasingly seeking concise and engaging segments of video content. In soccer matches, despite the fact that the duration of the match often exceeds 90 minutes, it is the brief highlight moments that truly capture the viewers’ attention. However, traditional editing methods are often time-consuming and labor-intensive, making it challenging to meet the immediate demands of contemporary viewers. To address this challenge, this paper presents an innovative automatic editing method for soccer match highlights using multimodal learning (AESHML), which integrates both video and audio features, enabling the precise capture of highlight moments in soccer events through advanced Transformer and long short-term memory (LSTM) multimodal learning models. Utilizing the self-constructed soccer match highlight moments dataset (SHD-114), the AESHML method exhibits outstanding performance, achieving an accuracy of 83.95% and an F1 score of 82.71%. The implementation of the AESHML method not only significantly reduces editing time and labor costs but also enhances editing efficiency, allowing viewers to experience the match more quickly and appreciate its essence. Furthermore, it promotes the dissemination of soccer events, thereby enabling a broader audience to appreciate the allure of the sport.

As an important ocean dynamic process, internal waves (IWs) have a remarkable effect on sediment resuspension and ocean mixing, and also seriously threaten oil rig operation and the navigability of underwater submarines. Synthetic aperture radar (SAR) is an advanced tool widely applied for IW observation due to its all-weather observation capability. However, complex atmospheric and oceanic environments can obstruct the IW structures in SAR single-channel images, and accurate automated extraction still limits the application of SAR-based IW detection from complex oceanic backgrounds. This study proposes a novel automatic extraction method for IWs to address this problem. The main innovations for accurate IW extraction with SAR data are highlighted as follows. 1) Polarimetric SAR images, which contain more information to characterize IWs, are utilized. 2) An attention-based U-Net is proposed to achieve better feature-focusing capabilities of the IW extraction network. 3) A multiscale loss function is proposed to improve the accuracy of the IW extraction network. We create an IW dataset that contains 1008 scenes of Sentinel-1 images and 10 686 IW crests in the Indo-Pacific warm pool. Experimentally, the proposed method achieves good completeness, correctness, and F1 scores of 94.54%, 81.43%, and 87.50%, respectively, which are significantly higher than other methods. Further experiments indicated that the trained IW extraction network shows satisfactory applicability and accuracy for different test areas and various SAR data. The proposed method has great potential applications for performing ocean/global scale monitoring and statistical analysis of IWs.

This paper investigates the automatic generation control in a deregulated environment for three unequal interconnected power systems involving renewable energy sources and electric vehicles. In each area, there are reheat thermal plants combined with various renewable energy sources, including solar photovoltaic plant, wind turbine plant, hydropower plant, biodiesel engine generators, biogas turbine generators, and electric vehicles. Non-linearities, such as generation rate constraints and governor dead bands, are taken into consideration. This paper proposes a new cascade controller called the two degree of freedom Tilt integral Derivative with Filter cascaded with Fractional order Tilt Integral controller using a new Zebra optimization algorithm to optimize controller parameters. Several case studies are done to illustrate the robustness of the proposed controller and optimization. It has been observed that the deviation in frequency’s settling time, peak overshoots, and undershoots are better with including EV by 55%,71.4% and 44.4% in Area1, 55.6%, 62.5% and 14.2% in Area 2, 30.3% and 46.15% in Area 3 than the system without EV.

In the aerospace field, researchers and engineers have been using attitude simulators to test and evaluate spacecraft attitude control algorithms, usually developed at the design stage. Moreover, thanks to the new technological achievements, small satellites can include more complex algorithms in their attitude determination and control systems. In consequence, the development of an attitude simulator that is able to perform different control algorithms with high attitude accuracy and stability is in high demand. In this paper, an automatic balancing method, along with a modified Model Reference Adaptive Control architecture for an uncertain dynamical system with unknown perturbation, was developed to address uncertainty suppression and disturbance rejection for an air-bearing spherical attitude simulator. The main feature of the proposed scheme consists of obtaining estimates of the offset vector generated by an unbalanced mass system in order to reduce it by moving small masses along each principal axis, ensuring stability and improving motion performance with an adaptive controller with input error modification. Simulation results were conducted according to the parameters for a small satellite 1U-class CubeSat platform inside a spherical structure.

Automatic speech recognition (ASR) systems have become increasingly popular in recent years due to their ability to convert spoken language into text. Nonetheless, despite their widespread use, existing speaker-independent ASR systems frequently encounter challenges related to variations in speaking styles, accents, and vocal characteristics, leading to potential recognition inaccuracies. This study delves into the feasibility of personalized ASR systems that adapt to the unique voice attributes of individual speakers, thereby enhancing recognition accuracy. It provides an overview of our methodology, focusing on the design, development, and evaluation of both speaker-independent and personalized ASR systems. The dataset used included diverse speakers selected from three extensive datasets: TedLIUM-3, CommonVoice, and GoogleVoice, demonstrating the capability of our methodology to accommodate various accents and challenges of both natural and synthetic voices. In terms of signal classification and interpretation, the personalized model eclipsed the speaker-independent variant, registering an enhancement of up to ~3% for natural voices and ~10% for synthetic voices in recognition accuracy for individual speakers. Our findings demonstrate that personalized ASR systems can significantly improve the accuracy of speech recognition for individual speakers and highlight the importance of adapting ASR models to individual voices.

The automatic load frequency control (ALFC) is designed based on integral sliding mode control (ISMC) without reaching time for a multi-time delay power system (MTDPS) with the system disturbances and uncertainties. In contrast with a recent method, the time delay of an area control error (ACE) signal and an interconnected signal are deliberated for the ALFC design in the MTDPS. The sliding mode is proposed with a selection of an integral single-phase surface and control rules such that the sliding state variables begin at the surface at the initial moment and converge to zero without reaching time. In addition, Lyapunov-based new linear matrix inequality (LMI) is applied to analyze the entire power system’s frequency steadiness. Moreover, the integral single-phase surface is advanced to improve the performance of the MTDPS. Using the proposed new integral single-phase surface approach, undershoot/overshoot and settling time were reduced compared to the recent (ISMC) approach. The results show the new scheme is highly robust in sliding variable’s fast convergence to zero asymptotical compared to the recently designed SMC. It has no significant interruption in operation, making its application in real power systems possible.

With the rapid advancements in deep learning, Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) has seen significant improvements in performance. However, the effectiveness of even the most advanced deep-learning-based ATR methods is limited by the scarcity of training samples. This challenge has sparked growing interest in SAR ATR under data-constrained conditions in recent years. Most current approaches for SAR ATR with limited data enhance recognition through data augmentation, specialized modules, or contrastive learning-based loss functions. However, effectively utilizing limited supervision signals to identify key features remains a significant challenge that existing methods have not thoroughly addressed. In our research, we introduce a novel coarse-to-fine hierarchical feature learning strategy for SAR ATR with limited data. Starting with a feature extractor that produces multi-level features, we implement a coarse-to-fine gradual feature constraint to optimize each level using limited supervision signals. This approach simplifies parameter search and ensures effective feature utilization from coarse to fine granularity. Additionally, our method enhances the compactness within classes and the separability between classes of features at various levels. This is achieved by capitalizing on the consistency of features across multiple levels, thereby progressively enhancing the features and, in turn, boosting the model's overall performance. To validate our approach, we conducted recognition and comparative experiments on the MSTAR and OpenSARShip datasets. The results demonstrate our method's exceptional performance in limited-sample recognition scenarios. Moreover, ablation studies confirm the robustness of our approach, underscoring its potential in addressing the challenges of SAR ATR with limited data.

The deep neural network has achieved remarkable recognition performance in synthetic aperture radar (SAR) automatic target recognition (ATR) by extracting the discriminative features from massive SAR images. Due to the sensitivity of SAR image to the observation aspect, the multiview ATR method could enhance the robustness of feature representation and improve the recognition performance. However, existing multiview ATR methods suffer from increasing complex structure and heavy computation when the number of input images grows. An adaptive multiview fusion network based on image attention (IA-AMF-Net) compatible with variable number of input images is proposed for SAR ATR in this article. In IA-AMF-Net, first, the depthwise separable convolution is employed to extract the classification features from multiple SAR input images in parallel with the lightweight attribute. Second, the channel feature weight vector of each image is generated and concatenated by applying the squeeze-and-excitation operation to the extracted features. The image attention weights for feature fusion are calculated through softmax normalizing the concatenated channel feature weights of input images. At last, the extracted features from multiview SAR images are fused by the obtained image attention weights. The dimension of fused feature keeps constant regardless of the number of input images, and the attention to the classification features of interested images is enhanced. Experimental results on the moving and stationary target acquisition and recognition dataset show that IA-AMF-Net achieves superior recognition performance under various operating conditions with fewer parameters and lower computational load compared to the other networks.

Traffic signal control is an important part of intelligent transportation. Efficient traffic signal control strategies not only alleviate traffic congestion, improve vehicle traffic efficiency, but also reduce exhaust pollution during the waiting period. The traditional multiple intersection signal light control method are generally faced with passive control and can not adapt to complex changes. In view of this, the study first combines reinforcement learning and Markov decision to construct a signal control problem model for multiple intersections. Secondly, deep Q-learning networks are introduced for decision solving, and multi-head attention mechanisms and graph convolutional networks are further introduced for optimization and improvement. Finally, a spatial lightweight model for multiple intersection adaptive graph convolution is proposed. The experimental results show that the maximum throughput of the multi-intersection adaptive control model can reach 195 vehicles/min, the shortest queue length is 4.31 vehicles, the shortest average delay is 1.23 minutes, the shortest average travel time is 2.07 minutes, and the attention value of adjacent intersections is close to 80%. The average green wave time under this model is 4 minutes, and the longest green wave section can reach 5 minutes. It can be seen that the new multi-intersection signal adaptive control model proposed in this study has advantages and feasibility compared with similar methods in various indexes, and can provide new ideas and methods for the further development of traffic signal automatic control technology.

This study addresses the challenges of effectively reflecting maritime traffic flows in Korea’s MSP. Maritime traffic intensity zones were identified through a spatio-temporal density analysis and the application of standardization techniques. This study proposes unified extraction standards for maritime traffic intensity zones and provides quantitative indicators for MSP based on the occupancy rates of Automatic Identification System (AIS) data. The results of the spatio-temporal density analysis were polygonized and represented at 10% intervals. A suitable maritime traffic intensity zone was then selected based on the occupancy rates of ship counts and ship points within each polygon, followed by a statistical review. Consequently, the maritime traffic intensity zone with an upper 20% density was identified, and the zone exhibiting an upper 10% density was designated as the primary maritime traffic intensity zone. This approach facilitates improved navigation and port area reflections, as anticipated from the implementation of unified extraction standards and methods for maritime traffic intensity zones established by local governments in Korea.

A sewer pipe system is an indispensable part of a modern city’s infrastructure. However, sewer pipes gradually deteriorate over time and develop defects such as cracks, deposits, and infiltration, which seriously affect their performance. Consequently, routine inspections are crucial for assessing the condition of sewer pipes and ensuring their proper operation. Closed-circuit television (CCTV) is frequently utilized for sewer pipe inspection. Previously, the interpretation of sewer pipe CCTV images relied on professional technologists and traditional computer vision techniques to identify and localize defects. They are inefficient and poorly robust. In this paper, we propose an efficient and robust automatic interpretation technique based on deep learning to identify and localize defects in sewer pipe CCTV images in real time. Specifically, we propose an improved you only look once version 5 large 6 (YOLO v5l6) network based on the original YOLO v5l6 network. The improved YOLO v5l6 network is based on improvements in the following two aspects: (1) We concatenate a convolutional block attention module (CBAM) module into the Bottleneck block of the Cross Stage Partial Bottleneck with 3 Convolutions (C3) module at the end of the original YOLO v5l6 backbone to create a C3CBAM module. (2) Moreover, we concatenate the CBAM into the end of four C3 modules of the original YOLO v5l6 neck to create four C3_CBAM_Att modules. With these improvements, the improved YOLO v5l6 network can significantly reinforce the backbone feature extraction and neck feature fusion to rapidly and accurately detect multiscale, multiclass defects in the presence of interference.

In this study, a method that automatically generates regular expressions using genetic algorithms is designed to extract relevant images on web pages. Data extraction, which is usually done with web scrapers, can also be done with regular expressions. The complexity of regular expressions and the fact that they require expert knowledge make their writing difficult. With this study, a regular expression is automatically created to obtain relevant images of news content on websites. With the principle of genetic algorithms, the survival of the good and the elimination of the bad, a regular expression that can reach the most relevant image is produced. Thus, instead of a time-consuming and error-prone method such as creating the appropriate pattern for each site with web scraper tools, automatic regular expression generation using genetic algorithm methods can be used as a better method. A data set containing text-based related and irrelevant images from 200 websites collected from 58 countries was used in the study. There are 22,682 relevant images among 635,015 image data in the dataset. With the method developed using the genetic algorithm, the rate of accessing the relevant images by regular expressions produced by only looking at the relevant image data is approximately 98.49%.

Automatic Speech Recognition (ASR) systems have improved and eased how humans interact with devices. ASR system converts an acoustic waveform into the relevant text form. Modern ASR inculcates deep neural networks (DNNs) to provide faster and better results. As the use of DNN continues to expand, there is a need for examination against various adversarial attacks. Adversarial attacks are synthetic samples crafted carefully by adding particular noise to legitimate examples. They are imperceptible, yet they prove catastrophic to DNNs. Recently, adversarial attacks on ASRs have increased but previous surveys lack generalization of the different methods used for attacking ASR, and the scope of the study is narrowed to a particular application, making it difficult to determine the relationships and trade-offs between the attack techniques. Therefore, this survey provides a taxonomy illustrating the classification of the adversarial attacks on ASR based on their characteristics and behavior. Additionally, we have analyzed the existing methods for generating adversarial attacks and presented their comparative analysis. We have clearly drawn the outline to indicate the efficiency of the adversarial techniques, and based on the lacunae found in the existing studies, we have stated the future scope.

In deep learning, various optimizers use a starting value for the learning rate. In general, this value is the one defined by default in the optimizer or chosen by the user after searching for a value giving the best performances of the studied neural network. Finding this value in the case of complex models with large-scale data, either by fine-tuning it manually or by applying a grid search algorithm requires several training sessions, which is penalizing in terms of learning time. To solve this problem, we propose an efficient optimizer using an automatic setting of the learning rate value. Learning then requires a single training session. The proposed optimizer is based on the inexact linear search of Armijo’s rule that we adapt for neural networks. We have experimented with this optimizer for several types of loss functions such as mean squared error, mean absolute error, and cross-entropy. Different types of models such as regression and classification are studied. Also, the robustness of this optimizer against the noisy labels is verified. Empirical results demonstrate that the proposed optimizer works well in practice, it outperforms other state-of-the-art optimizers that use default learning rate in terms of accuracy in addition to the time saving when searching for the optimal value of the learning rate.

This paper addresses the training issues associated with neural network-based automatic speech recognition (ASR) under noise conditions. In particular, conventional joint training approaches for a pipeline comprising speech enhancement (SE) and end-to-end ASR model surfer from a conflicting problem and a frame mismatched alignment problem because of different goals and different frame structures for ASR and SE. To mitigate such problems, a knowledge distillation (KD)-based training approach is proposed by interpreting the ASR and SE models in the pipeline as teacher and student models, respectively. In the proposed KD-based training approach, the ASR model is first trained using a training dataset, and then, acoustic tokens are generated via K-means clustering using the latent vectors of the ASR encoder. Thereafter, KD-based training of the SE model is performed using the generated acoustic tokens. The performance of the SE and ASR models is evaluated on two different databases, noisy LibriSpeech and CHiME-4, which correspond to simulated and real-world noise conditions, respectively. The experimental results show that the proposed KD-based training approach yields a lower character error rate (CER) and word error rate (WER) on the two datasets than conventional joint training approaches, including multi-condition training. The results also show that the speech quality scores of the SE model trained using the proposed training approach are higher than those of SE models trained using conventional training approaches. Moreover, the noise reduction scores of the proposed training approach are higher than those of conventional joint training approaches but slightly lower than those of the standalone-SE training approach. Finally, an ablation study is conducted to examine the contribution of different combinations of loss functions in the proposed training approach to SE and ASR performance. The results show that the combination of all loss functions yields the lowest CER and WER and that tokenizer loss contributes more to SE and ASR performance improvement than ASR encoder loss.

Automatic Modulation Classification (AMC) is an essential technology that is widely applied into various communications scenarios. In recent years, many Machine Learning and Deep-Learning methods have been introduced into AMC, and a lot of them apply different approaches to eliminate interference in complex Multiple-Input and Multiple-Output (MIMO) signals and improve classification performance. However, in practical communication systems, the perfect elimination of MIMO signal interference is impossible, and therefore classification performance suffers. In this paper, we propose a new AMC algorithm for MIMO system based on mutual information (MI) features extraction, which does not require a large amount of training data and the elimination of MIMO signal interference. In this approach, features based on mutual information are extracted using In-Phase and Quadrature (IQ) constellation diagrams of MIMO signals, which have not been explored previously. Our method can be effective since mutual information considers the interdependencies among variables and measures how much information about one variable reduces uncertainty about another, providing a valuable perspective for extracting higher-level and interesting features from the data. The effectiveness of our method is evaluated on several model and real-world datasets, and its applicability is proven.

Automatic modulation recognition (AMR) is an essential topic of cognitive radio, which is of great significance for the analysis of wireless signals and is one of the current research hotspots. Traditional AMR approaches predominantly utilize raw in-phase/quadrature symbols (I/Q), amplitude/phase (A/P), or pre-processed data (e.g., high-order cumulates, spectrum images, or constellation diagrams) as inputs for the recognition model. However, it is difficult to achieve superior performance with only a single type of data as input. This paper proposes a novel multi-channel hybrid learning framework that integrates convolutional layers, Long Short-Term Memory (LSTM) layers, fully connected layers and classification layers. The model is built for modeling spatial-temporal correlations from four signal cues (including I/Q signals, A/P signals, I, and Q signals), which aims to explore various differences and leverage the complements from multiple data-form. Two functions employed during the data conversion process further enhance the non-linear representational capacity of the model, thereby boosting the recognition accuracy of the model. Experimental results demonstrate that the proposed framework effectively addresses the classification challenges of QAM16 and QAM64. For the RAML2016A dataset, our model achieves an impressive recognition accuracy of 95% at an SNR of 0 dB. Extensive experiments indicate that the proposed framework outperforms other current networks in terms of recognition accuracy.

Synthetic aperture radar automatic target recognition (SAR-ATR) models based on deep neural networks (DNNs) are vulnerable to attacks of adversarial examples. Universal adversarial attack algorithms can help evaluate and improve the robustness of the SAR-ATR models and have become a research hotspot. However, current universal adversarial attack algorithms have limitations. First, considering the difficulty in obtaining information on the attacking SAR-ATR models, there is an urgent need to design a universal adversarial attack algorithm under a black-box scenario. Second, given the difficulty of acquiring synthetic aperture radar images, the effectiveness of attacks under small-sample conditions requires improvement. To address these limitations, this study proposed a black-box universal adversarial attack algorithm: transferable universal adversarial network (TUAN). Based on the idea of the generative adversarial network, we implemented the game of generator and attenuator to improve the transferability of universal adversarial perturbation (UAP). We designed loss functions for the generator and the attenuator, respectively, which can effectively improve the success rate of black-box attacks and the stealthiness of attacks. In addition, U-Net was used as a network structure of the generator and the attenuator to fully learn the distribution of examples, thereby enhancing the attack success rate under small-sample conditions. The TUAN attained a higher black-box attack success rate and superior stealthiness than up-to-date UAP algorithms in non-targeted and targeted attacks.

The performance of a deep learning-based synthetic aperture radar (SAR) automatic target recognition (ATR) model largely relies on the scale and quality of training samples. However, it is time-consuming and expensive to collect sufficient data in practice. Although generative adversarial network (GAN) provides a way for SAR target image generation, existing GAN-based methods cannot confirm what features the generator learns, thus they struggle in generating precise SAR target images. In this article, we propose an angle transformation GAN (ATGAN) that can generate azimuth-controllable SAR target images while preserving the target details. The key idea of our ATGAN is to reframe the generation task from the perspective of image-to-image translation. To this end, ATGAN consists of two modules, a coarse-to-fine generator that aims to learn the angle transformation in the deep feature space, and then, apply it to manipulate the representation of an input SAR target image to generate a new one, while a spectral-normalized patch discriminator that tries to estimate the probability that an input SAR target image is real rather than fake using a patch-averaged strategy. Combining with spatial transformer and adversarial training paradigm, ATGAN can generate precise SAR target images for ATR. Extensive experiments verify the effectiveness of the proposed ATGAN, and our method outperforms the state-of-the-art method qualitatively and quantitatively.

Acupuncture can be effective in relieving pain and reducing drug dependence and abuse caused by chronic pain. However, the cumulative effect of acupuncture requires patients to undergo multiple treatments, which limits its use due to time and economic costs. Currently, existing acupuncture devices include acupoint identification and acupoint stimulation devices. The identification and stimulation functions are separated for these two classes of devices. They were unable to locate the acupoint automatically and required a professional acupuncturist to perform the procedure. Although acupuncture robots that can simultaneously identify the acupoints and initiate treatment are under development, the technology is not yet mature that it is currently not available for general use. To address these issues, we developed an automatic system that combines the identification of acupoints and the administration of treatment. The system is designed to assist users without any knowledge of acupuncture in the management of chronic pain at home. We conducted tests to evaluate the performance of the system in terms of resistance detection accuracy, electroacupuncture accuracy, and acupoint identification test. The results demonstrate that the system significantly improves the accuracy of locating acupoints, indicating its suitability for daily home use as an effective approach to managing chronic pain.

This study focuses on the automatic generation control (AGC) system, which is crucial for maintaining balance between power generation and demand in power systems. The implementation of the AGC system needs to be more precise due to the increasing uncertainty surrounding renewable energy sources (RESs) and changes in demand. The objective of this study is to investigate the AGC functions in a two-area hybrid power system that combines a PV system with a reheat thermal system. To improve system performance, we utilize a proportional-integral (PI) controller. We utilized a recently developed optimization method, RIME, for tuning controller parameters. This technique has not been studied before in AGC processes. Furthermore, the optimization procedure utilizes a modified version of the integral of time-multiplied absolute error (ITAE) objective function. The study compares the performance of the RIME-tuned PI controller under various scenarios, including changes in thermal system load, load variations in both areas, and robustness considerations, with well-known techniques in the literature, such as the black widow optimization algorithm (BWOA), the salp swarm algorithm (SSA), the shuffled frog leaping algorithm (SFLA), the firefly algorithm (FA) and the genetic algorithm (GA). Our comparative study demonstrates that the proposed controller outperforms state-of-the-art approaches in terms of overshoot values and damping durations for both system frequency and tie-line power changes. The study provides valuable information on the effectiveness of the RIME-tuned PI controller in controlling AGC processes in complex hybrid power systems.

Sentiment analysis represents the systematic procedure of independently discerning polarity inherent in a textual document. A multitude of sectors can derive substantial advantages from this specialized domain. Conducting sentiment analysis (SA) involves various phases, with the initial step being the annotation process, which is often time-consuming and laborious. Within this framework, there exists a notable scarcity of existing research works. The complexity of this task becomes more difficult when analyzing texts in ‘Darija’, a form of the Moroccan dialect (MD). In our research endeavors, we introduced a novel automatic annotation methodology designed explicitly for sentiment analysis within the Moroccan dialect. A pivotal aspect of our contribution is the refinement of the stacking approach, utilizing a weighted voting technique for enhanced predictive accuracy. Our advanced method starts with the training of various neural network models across six unique MD datasets. The selection of these neural network architectures was underpinned by a comprehensive grid search procedure. Conclusively, it was discerned that models predicated on Recurrent Neural Networks (RNNs) outperformed others. Subsequent to this, we deployed an augmented stacking model, grounded in the aforementioned weighted voting technique. This model leverages the predictions generated by the neural networks as inputs. It then employs the mode of these inputs as an output, which feeds directly into a meta-classifier, which in turn produces the coefficients. These coefficients are then multiplicatively combined with the initial neural network predictions to derive the finale outputs. To evaluate the efficiency of our proposed methodology in annotating the six datasets, each dataset was isolated as a test while the remaining five served as training sets. Consequently, within the set of six datasets, the annotation results of three datasets have outperformed the established standards, attaining agreement rate percentages of 87.54% for MSAC, 91.25% for FB, 85.10% for MSDA, and 83.60% for MSTD, all of which represent new achievements in the literature.

Recent advances in deep learning (DL) have led many contemporary automatic modulation classification (AMC) techniques to use deep networks in classifying the modulation type of incoming signals at the receiver. However, current DL-based methods face scalability challenges, particularly when encountering unseen modulations or input signals from environments not present during model training, making them less suitable for real-world applications like software-defined radio devices. In this paper, we introduce a scalable AMC scheme that provides flexibility for new modulations and adaptability to input signals with diverse configurations. We propose the Meta-Transformer, a meta-learning framework based on few-shot learning (FSL) to acquire general knowledge and a learning method for AMC tasks. This approach empowers the model to identify new unseen modulations using only a very small number of samples, eliminating the need for complete model retraining. Furthermore, we enhance the scalability of the classifier by leveraging main-sub transformer-based encoders, enabling efficient processing of input signals with diverse setups. Extensive evaluations demonstrate that the proposed AMC method outperforms existing techniques across all signal-to-noise ratios (SNRs) on RadioML2018.01A. The source code and pre-trained models are released at https://github.com/cheeseBG/meta-transformer-amc.

The significance of automatic text summarization (ATS) lies in its task of distilling textual information into a condensed yet meaningful structure that preserves the core message of the original content. This summary generated by ATS plays a crucial role in simplifying the processing of textual information, as it captures the primary ideas of the source text while eliminating lengthy and irrelevant textual components. At present, the landscape of ATS is enriched with a multitude of innovative approaches, with a notable focus on optimization-based methods. These optimization-driven ATS techniques have introduced new perspectives, illuminating the field with their heightened accuracy in terms of metrics like ROUGE scores. Notably, their performance closely rivals other cutting-edge approaches, including various methodologies within the realm of machine learning and deep learning. The review presented in this paper delves into recent advancements in extractive ATS, centering mainly on the optimization-based approach. Through this exploration, the paper underscores the gains and trade-offs associated with adopting optimization-based ATS compared to other strategies, specifically with the application of real-time ATS. This review serves as a compass, pointing towards potential future directions that the optimization-based ATS approaches should consider traversing to enhance the field further.

Synthetic aperture radar (SAR) images often lack sufficient attention to target features, inadequately express target feature information, and neglect phase information in conventional Convolutional Neural Network (CNN) recognition methods. These limitations lead to reduced recognition accuracy and slower processing speeds, critical drawbacks for SAR Automatic Target Recognition (ATR) systems. To overcome these challenges, this paper proposes the multi-scale complex-valued feature attention CNN (MsCvFA-CNN) for SAR ATR. MsCvFA-CNN is a model specifically designed for amplitude and phase information of SAR images. A novel complex-valued attention module (CAM) is proposed in this work to focus on the amplitude and phase characteristics of the target separately. By decoupling the amplitude and phase features, the CAM reduces the training time of the network, while preserving the relevant information. Furthermore, the MsCvFA-CNN employs multiple branches for feature extraction with different kernel sizes, which are then combined with CAM in the fusion stage to improve the network's representation of target features. The proposed MsCvFA-CNN is evaluated on both the complex-valued moving and stationary target acquisition and recognition (MSTAR) dataset, as well as the more challenging dataset for urban interpretation (OpenSARUrban). The results demonstrate that it outperforms traditional networks in terms of recognition accuracy and computational efficiency. Specifically, the use of complex-valued networks results in a 2.23% improvement in recognition accuracy compared to traditional real-valued networks. When CAM is added, the network's accuracy is further improved by 3.21%, and the number of epochs required to achieve the highest accuracy is reduced by nearly half.

Automated diagnosis, as a temporary medical supplement, has gained significant attention in research in recent years. Existing methods employ sequence generation approaches to inquire about symptoms and diagnose diseases. However, these methods ignore the fact that: 1) doctors utilize their past experience and similar cases to aid in diagnosis in real-world scenarios; 2) doctors inquire about key symptoms that serve as vital diagnostic evidence within limited conversations. To address these issues, we propose an end-to-end model KDPoG. Firstly, in addition to use the symptom and attribute embedding, we propose patient-oriented graph enhanced representation learning, which is built by a patient-oriented graph and learned with heterogeneous graph convolution networks. Furthermore, based on the encoder built with gated attention units, we propose knowledge-guided attention mechanism learning, which incorporates conditional probabilities of co-occurrence between symptom pairs. Finally, we utilize two linear layers as the classification module to achieve symptom probing and disease diagnosis. We conduct extensive experiments on four public datasets, which demonstrate that our proposed model outperforms the state-of-the-art methods. We achieve an average absolute improvement of over 2% in disease diagnosis accuracy. Particularly, on the Muzhi-10 dataset, we observe an absolute improvement of over 14.7% in symptom recall rate.

The problem of closed branches and leaves in orchards can cause satellite navigation systems to experience star loss, leading to safety issues such as collisions and loss of control. To address this issue, this study develops a multi-mode navigation control system for satellite navigation and sensor tracking, aiming to achieve multi-mode intelligent path judgment and ensure the accuracy and safety of operations. This method utilizes point cloud iterative nearest point algorithm LiDAR technology to gain 3D information of the surrounding environment when satellite signals are unavailable. Through perception and analysis of the environment, real-time path planning and obstacle avoidance decision-making are achieved. At the same time, through micro-inertial navigation technology, accurate perception and control of the position and attitude of navigation equipment are achieved, thereby achieving precise control of the operation process. According to relevant experimental data, the experimental distance accuracy based on the vehicle mounted daytime environment dataset was 84.88%, and the duration accuracy was 97.63%. Accordingly, the proposed system is capable of accurately determining the current position and heading, as well as making path planning and obstacle avoidance decisions based on real-time environmental information. This ensures the accuracy and safety of operations. The research method exhibits higher stability and reliability in orchard operations, providing an effective solution for the accuracy and safety of orchard operations.

Objective: Sleep monitoring has extensively utilized electroencephalogram (EEG) data collected from the scalp, yielding very large data repositories and well-trained analysis models. Yet, this wealth of data is lacking for emerging, less intrusive modalities, such as ear-EEG.Methods and procedures: The current study seeks to harness the abundance of open-source scalp EEG datasets by applying models pre-trained on data, either directly or with minimal fine-tuning; this is achieved in the context of effective sleep analysis from ear-EEG data that was recorded using a single in-ear electrode, referenced to the ipsilateral mastoid, and developed in-house as described in our previous work. Unlike previous studies, our research uniquely focuses on an older cohort (17 subjects aged 65-83, mean age 71.8 years, some with health conditions), and employs LightGBM for transfer learning, diverging from previous deep learning approaches. Results: Results show that the initial accuracy of the pre-trained model on ear-EEG was 70.1%, but fine-tuning the model with ear-EEG data improved its classification accuracy to 73.7%. The fine-tuned model exhibited a statistically significant improvement (p < 0.05, dependent t-test) for 10 out of the 13 participants, as reflected by an enhanced average Cohen’s kappa score (a statistical measure of inter-rater agreement for categorical items) of 0.639, indicating a stronger agreement between automated and expert classifications of sleep stages. Comparative SHAP value analysis revealed a shift in feature importance for the N3 sleep stage, underscoring the effectiveness of the fine-tuning process.Conclusion: Our findings underscore the potential of fine-tuning pre-trained scalp EEG models on ear-EEG data to enhance classification accuracy, particularly within an older population and using feature-based methods for transfer learning. This approach presents a promising avenue for ear-EEG analysis in sleep studies, offering new insights into the applicability of transfer learning across different populations and computational techniques.Clinical impact: An enhanced ear-EEG method could be pivotal in remote monitoring settings, allowing for continuous, non-invasive sleep quality assessment in elderly patients with conditions like dementia or sleep apnea.

The computer-based Automated Essay Scoring (AES) system automatically marks or scores student replies by considering relevant criteria. The methodology systematically categorises writing quality and can increase operational effectiveness in academic and major commercial institutions. To study the projected score, AES relies on extracting numerous aspects from the student’s response, including grammatical and textural information. However, the recovered features may result in dimensionality reduction and a challenging-to-understand feature selection procedure. As the number of parameters rises, the model also demands a large cost for processing and training the data. However, these problems worsen the accuracy of score prediction and widen the gap between actual and anticipated results. This study suggested the Fox-optimised Long Short-Term Memory-based Augmented Language Model (FLSTM-ALM) as a solution to these problems for giving successful training to text features; the model uses an augmented learning paradigm. The retrieval score was then analysed and generated using a neural knowledge encoder and retriever. The neural model successfully classifies the output based on this score. The best features are chosen using the Fox optimisation algorithm based on the food-searching category. This choice of parameters solves the exploration and optimisation issues with document classification. The performance of the optimised AES system was assessed using the two datasets, ASAP and ETS, and it demonstrated a high accuracy of 98.92% and a low error rate of 0.096%. Dimensionality reduction can thus be fixed by optimising the FLSTM-ALM model with an appropriate meta-heuristic method, such as the FOX algorithm, which raises the predicted accuracy, recall, and f1 score for the AES model.

To address the issue of measurement errors in ultrasonic water meters caused by various interfering factors during the flow calibration process, this study proposes a highly automated and intelligent ultrasonic water meter flow calibration system design. Firstly, by analyzing and studying the working principle of the water flow calibration device, and considering the actual flow calibration range, the calibration method that combines the master meter method and the dynamic weighing method is determined. To enhance the automation level of the flow calibration system, the flow calibration control algorithm is thoroughly analyzed and designed. Finally, a flow correction algorithm based on linear interpolation was designed and utilized for water meter flow calibration. Test data show that after the flow correction, the value error of the ultrasonic water meter in the “high area” is not more than 2%, in the “low area” it is not more than 3%, and the repeatability is less than 0.05%. Compared with manual calibration, the flow calibration system designed in this paper enables high-precision automated flow calibration of ultrasonic water meters, significantly enhancing detection efficiency and accuracy.

In this article, we explore a cutting-edge concept known as class incremental learning (CIL) in novel category discovery for synthetic aperture radar (SAR) targets (CNTs). This innovative task involves the challenge of identifying categories within unlabeled datasets by utilizing a provided labeled dataset as reference. In contrast to the conventional category discover approaches, our method introduces novel categories without relying on old labeled classes and effectively mitigates the issue of catastrophic forgetting. Specifically, to reduce the bias of the established categories toward unknown ones, CNT extracts representational information via self-supervised learning, gleaned directly from the SAR data itself to facilitate generalization. To retain the model's competence in classifying previously acquired knowledge, we employ a dual strategy incorporating the rehearsal of base category feature prototypes and the application of knowledge distillation. Our methodology integrates multiview and pseudolabeling strategies. In addition, we introduce a novel approach that focuses on enhancing the discernibility of class spaces. This strategy primarily ensures distinct separation of the unlabeled classes from base class prototypes, and imposes stringent constraints on the internal relationships among individual samples and their corresponding perspectives. To the best of our knowledge, this is the first study on category discovery in the CIL scenario. The experimental results show that our method significantly improves the performance on SAR images compared to the previous optimal method, which indicates the effectiveness of our method.

Annotated corpora are indispensable tools to train computational models in Artificial Intelligence and Natural Language Processing. However, manual annotation is a costly, arduous, and time-consuming task, especially when the annotation is semantically complex. To address the problem, this work applies a methodology for semi-automatic annotation of datasets based on the Human-in-the-Loop paradigm. The methodology supports the building of a resource, that benefits from a fine-grained annotation, to aid in the detection of Spanish violent messages sourced from social media (Twitter/X). After implementing the proposed methodology for semi-automatic violence annotation, a high quality resource was obtained (hereafter referred to as VILLANOS). The methodology consists of annotating the dataset incrementally, which delivers an increase in annotator efficiency, thereby validating the suitability of the proposal. Annotation time was reduced by 52% compared to manual annotation and performance, by training a model with the VILLANOS dataset, obtains an $F_{1}$ of 85.2%. These results demonstrate the efficiency and effectiveness of the methodology, evidencing its validity.