As the informatization process continues to advance, digital natives have become a major component of society. The widespread application of digital technologies in education has transformed their learning styles and knowledge acquisition, leading to a growing phenomenon of technology dependency. Based on grounded theory, this study conducts in-depth interviews with 31 digital natives and implemented three-level coding using NVivo11 to identify the underlying mechanisms that lead to technology dependency in the learning process and construct an attribution model of technology dependency. The model identifies factors contributing to technology dependency at the individual, school, and societal levels, including technological convenience, teacher encouragement, and peer pressure. It also identifies potential cognitive, behavioral, and emotional impacts of technology dependency, including reduced independent learning, weakened critical thinking, and distracted attention. To address these impacts, intervention strategies are proposed at the school and societal levels, including increased awareness and supervision, as well as psychological counseling and intervention, to help digital natives reduce their technology dependency and develop appropriate behavior habits using technology.

A multi-sensor fusion model based on Bidirectional Long Short-Term Memory (BiLSTM) and gated attention mechanism is proposed to address the challenges of heterogeneous data fusion and insufficient real-time performance in multi-sensor human activity recognition. This method captures bidirectional long-term dependencies through BiLSTM and employs gated attention to achieve adaptive weighting of multi-source features, effectively enhancing feature representation capability and fusion efficiency. Experimental results demonstrate that the proposed approach achieves a highest accuracy of 95.3%,in addition, under a comparable lightweight setting, the proposed model exhibits lower parameter count and faster inference latency than the Transformer baseline, thereby achieving a more favorable balance between accuracy and efficiency.Ablation studies further confirm the critical role of BiLSTM and the attention mechanism in improving performance. This research provides a solution that balances recognition accuracy and inference efficiency for behavior recognition in complex scenarios, showing strong potential for practical applications.

To address the importance of global background information and local detail features in rain streak removal, as well as the limitations of deep learning models in handling high-resolution images, we propose ClearMamba, a spatio-temporal state-space model with global context fusion. This model achieves pixel-level feature refinement through spatio-temporal feature modeling and global context integration, significantly enhancing single-image deraining performance. Experiments on datasets such as DID-Data demonstrate the superiority of this algorithm, with a peak signal-to-noise ratio (PSNR) improvement of approximately 4.9% compared to methods like PReNet. Furthermore, we systematically evaluate the empowering potential of deraining models for downstream visual tasks (e.g., object detection), achieving top performance. The proposed method offers a practical solution for optimizing intelligent visual systems in rainy environments.

Online peer assessment provides new theoretical and technical means for the development of critical thinking, and peer evaluation and peer interaction are the main activities implemented by online peer assessment. The article proposes a model of online peer assessment for critical thinking development around peer evaluation and peer interaction, uses social network and epistemic network analysis methods, and explores the development process of critical thinking in online peer assessment from the temporal dimension and role dimension, respectively, using the interactive comments as the analysis content. The findings show that (1) critical thinking shows a progressive and continuous development process in online peer assessment; (2) peer interaction is the main driver for critical thinking development; (3) critical thinking has progressive differences in different stages and groups of online peer assessment; and (4) critical thinking has a significant structural pattern. The article reveals the development process of critical thinking from the perspective of online peer assessment, and provides a theoretical basis and practical reference for improving the quality of online peer assessment and the development level of critical thinking.

Traditional anti-bullying interventions in schools rely primarily on proactive reports from teachers and students, together with periodic inspections and surveillance by administrators. These approaches are untimely reactive and labor-intensive, which makes timely detection and response to bullying incidents difficult. A deep-learning-based campus-bullying monitoring system that provides robust protection for campus security and safeguards the mental health of students. The system integrates YOLOv8 for object detection and 3D ResNet for action recognition as its core algorithms, while leveraging servo-driven dynamic tracking as its enabling hardware technology. By combining modules for object detection, action analysis, and dynamic tracking, we construct a comprehensive and intelligent monitoring framework. By processing and analyzing live surveillance streams in real time, the system can rapidly and accurately identify potential bullying behaviors and promptly issue early-warning notifications so that administrators can intervene without delay. In addition, the system includes the “Glimmer Haven” mini-program, which offers emotional support and psychological solace to victims, alleviating their mental distress. An AI-powered chatbot further assists students and faculty in understanding and addressing campus-bullying issues, thereby fostering a safer and more harmonious school environment.

Under the background of the "Double High Plan" construction, the disconnect between higher vocational talent cultivation and industrial technology iteration urgently needs to be addressed. Based on the "Five Golds" (elite majors, high-quality courses, outstanding teachers, advanced practice bases, and innovative teaching materials) integrated resource optimization framework, and relying on the reform practice of the Internet of Things major at Suzhou Polytechnic University, this study explores the construction of an educational ecosystem featuring "integration of industry and education, and integration of certificates and curricula". The model aims to achieve deep coupling between professional chains and industrial chains, dynamic synchronization between course content and technological development, and effective integration of faculty capabilities with industry practices. Practical results demonstrate that this approach enhances the alignment of students' occupational competence with enterprise demands, providing a replicable pathway for the digital transformation of higher vocational education. As a foundational project for high-quality development in vocational education, the "Five Golds" initiative requires systematic integration and dynamic adaptation mechanisms to drive vocational education from adaptive reform toward innovative leadership^[1].

The Linux laboratory course is a fundamental core course for computer science and information related majors, characterized by strong practicality and complex operational chains. Traditional evaluation systems mainly rely on students' submitted lab reports, which are result-oriented, lacking process analysis and contextual awareness, making it difficult to comprehensively and dynamically reflect students’ actual abilities and learning status. To this end, this paper proposes an evaluation system for Linux laboratory courses that integrates knowledge graphs with multimodal perception. It integrates key technologies such as code behavior analysis, speech and facial emotion recognition, system operation trajectory extraction, and knowledge graph reasoning. The system takes multimodal inputs—including student behavior logs, oral defense speech, facial expressions, operation flows, and error types—and applies focal contrastive learning and a Neo4j-based knowledge graph model to achieve comprehensive scoring and personalized feedback. Experimental results demonstrate that the proposed system significantly improves evaluation accuracy, consistency, and the stimulation of students’ learning motivation, providing a new paradigm and technological support for teaching reform in Linux and other system-oriented laboratory courses.

With the rapid development of technologies such as cloud computing and artificial intelligence (AI), the education sector is undergoing profound changes. As an essential component of the information technology major, the teaching reform and practice of Linux technology courses are of great significance for cultivating high-quality talents that meet the demands of the new era. This paper focuses on the teaching innovation in the field of information technology and deeply explores the teaching reform paths and practical methods of Linux technology courses in a private cloud environment, empowered by AI technology. By analyzing the existing problems in the current teaching of this course, it expounds on the necessity and feasibility of the cloud-intelligence integrated teaching model. It elaborates on specific measures such as building a teaching environment based on a private cloud platform, optimizing teaching content and resources with AI assistance, and innovating teaching methods and means. Through the analysis of actual teaching cases, it compares the teaching effects before and after the reform, aiming to provide beneficial references and inspirations for improving the teaching quality of Linux technology courses and cultivating high-quality talents that adapt to the demands of the times.

Aiming at the problems existing in the current teaching of database course, such as single teaching mode, disjointed theory and practice, and one-sided evaluation method, this paper puts forward a project-driven teaching mode based on constructivism theory. This model takes "supermarket management system" as the teaching project carrier, and designs and implements the whole process teaching scheme covering project start-up, exploration construction and achievement evaluation, aiming at guiding students to actively construct database knowledge through active exploration and collaborative communication in real situations. Teaching practice shows that this model can effectively improve students' ability of problem solving, teamwork and knowledge transfer, and provides an effective path with both theoretical support and practical value for the teaching reform of database course.

The "Curriculum-Training-Competition-Innovation-Industry"(CTCI²) innovation and entrepreneurship practice teaching model, driven by digital-intelligent empowerment and competitions in higher vocational computer-related disciplines, aims to effectively integrate digital-intelligent technologies, skills competitions, and innovation/entrepreneurship practices. Its goal is to cultivate high-level technical and skilled talents equipped with innovative thinking and precise career orientation, capable of adapting to the new-era industrial environment. This model explores the construction of a five-in-one innovation and entrepreneurship practice teaching framework based on digital-intelligent empowerment and competition-drive, the establishment of a digital-human interactive CTCI² innovation platform, and the development of an integrated educational effectiveness evaluation system. It seeks to propel higher vocational computer education into a new stage characterized by enhanced quality and excellence, value-added empowerment, and ultimately achieve high-quality development in vocational education.

Aiming at the core pain points of current car owners, such as scattered maintenance records and lack of targeted reminders, and combining with the diversified development trend of China's automobile aftermarket in 2024, an intelligent record and recommendation system for automobile maintenance is designed and implemented. With intelligent recording and accurate reminders as the core, the system constructs a three-layer architecture consisting of a perception layer, a data layer, and an application layer: The perception layer collects image documents, integrates YOLOv8 to locate and correct document areas, and uses OCR (Optical Character Recognition) technology to extract VIN (Vehicle Identification Number) codes. The data layer adopts MySQL master-slave architecture, Redis cache, and Alibaba Cloud OSS (Object Storage Service) to store information such as vehicle files and maintenance records. The application layer generates accurate reminders based on three-dimensional factors and dynamically iterates the system message reminder strategy in combination with user feedback.Test results show that the recognition accuracy of key information in system documents reaches 99.5%, and the record query response time is less than 1 second. The system can effectively reduce manual entry operations, solve the problems of low efficiency of traditional maintenance management and lack of targeted reminders, and provide car owners with a lightweight and highly user-friendly one-stop maintenance management service.

Focusing on the prediction of customers' new car model purchase behavior, this paper carries out a comparative analysis on the performance of four classification algorithms: Logistic Regression (LR), K-Nearest Neighbors (KNN), Gaussian Naive Bayes (GaussianNB), and Support Vector Machine (SVM). Based on data related to car consumption, following data preprocessing, hyperparameter tuning is performed for each model. For Logistic Regression, the regularization parameters and penalty terms are optimized. For the KNN algorithm, the number of neighbors is adjusted, and an appropriate distance measurement method is selected. For Gaussian Naive Bayes, the adaptability of the feature independence assumption is verified, and corresponding fine-tuning is carried out. Optimization is performed on key parameters of SVM, including the kernel function and regularization coefficient. After training and testing evaluation, the SVM demonstrates outstanding comprehensive performance after tuning, with an AUC reaching 0.965 and a prediction accuracy of 93%. Both KNN and Gaussian Naive Bayes achieve an AUC of 0.963 and an accuracy of 93%, while Logistic Regression has an AUC of 0.955 and an accuracy of 91%. The research shows that SVM has significant advantages in fitting nonlinear data, is suitable for the car consumption prediction scenario, and provides powerful data support and decision basis for enterprises to develop targeted marketing strategies.

On-site observation in the marine environment is required to obtain real and reliable data for research on marine microorganisms. This article presents an automatic seawater sample collection and distribution system for ships based on the NI CompactRIO. The system is deployed on scientific research vessels and establishes connections with instruments such as the winch system at the stern of the vessel and microbial flow image analyzer, under the control of the management system, achieving fully automatic or semi-automatic extraction, filtration, transportation, and automatic cleaning of the entire pipeline of seawater samples. The system can also monitor various water quality parameters such as temperature, salinity, dissolved oxygen, conductivity, chlorophyll, etc. online and upload to the system for storage. The system is designed based on a reliable and stable embedded hardware system, which is conducive to long-term operation in marine environments, maximizing the preservation of the original appearance of seawater samples, and ensuring the authenticity and effectiveness of analysis and detection data.

In response to the disconnection between traditional electronic technology experiment teaching content and complex engineering problems, this paper restructures the experimental courses in a task-chain-based hierarchical manner to enhance students' practical abilities and comprehensive professional competencies. During the restructuring process, the curriculum design adopts the perspective of engineers' professional qualities and technical skills, promoting the deep integration of experimental content with engineering demands. The experimental content has been reorganized into five progressive levels: basic experimental projects, compound experimental projects, comprehensive experimental projects, AI-enhanced innovative experimental projects, and real competition questions from electronic design contests. Simultaneously, the course promotes innovation and reform in teaching methods and evaluation systems, moving away from a single assessment model toward a multidimensional framework that integrates theoretical knowledge, practical skills, innovative thinking, and teamwork. Through systematic reform, the electronic technology experiment course not only enhances students' hands-on abilities, but also consistently strengthens their comprehensive competencies in analyzing and solving engineering problems. Ultimately, these efforts lay a solid foundation for cultivating high-quality, interdisciplinary, and application-oriented talent in private universities, effectively bridging the gap between talent development and industry needs.

This article aims to explore an effective evaluation path that adapts to the educational goals of vocational colleges that combine moral and technical education. It organically integrates the CIPP evaluation model with the concept of "dual core of morality and technology", and constructs a three-dimensional evaluation system that covers professional qualities, skills, and career development capabilities. Take the course of Responsive Web Development for example, it conducts evaluations through four stages: background, input, process, and outcome, and combines blended teaching to achieve full-cycle data collection and dynamic optimization. Practice shows that this system can enhance students' technical abilities and professional qualities, providing theoretical and practical paths for the implementation of "dual core of morality and technology" in professional courses in vocational colleges.

In the digital era, the problem of data transmission security has become increasingly prominent, and traditional encryption methods have certain limitations. In order to improve the security of data transmission, this paper proposes an encryption and decryption method based on AES algorithm. Firstly, AES algorithm is used to generate data key, which is efficient, secure and widely used. During encryption, byte substitution and column are mixed, the former is nonlinear byte substitution through S-box, and the latter is linear transformation of data matrix column to enhance the degree of data confusion. At the same time, the artificial bee colony algorithm is used to optimize the confusion sequence composed of data and key, and give full play to its global search and adaptive ability to improve the encryption security and complexity. Experiments are carried out under different transmission bandwidths. The results show that compared with the existing methods, this method can resist brute force attack more effectively, the number of successful attacks is significantly reduced, and the data transmission security is guaranteed. After applying this method, the encryption and decryption time is shorter and the memory usage is reduced.

Rain in the acquired images can cause problems such as loss of details and blurry textures, which is not conducive to the analysis and research of subsequent computer vision tasks. To remove rain stripe from images and obtain clean background images, a hybrid CNN-Transformer single image deraining network based on attention mechanism is proposed. Firstly, shallow feature extraction is performed by the first 5 layers of Resnet18. Then, the high and low frequency rain stripe detection module is introduced to generate rain stripe attention maps to promote the subsequent network focus on rain stripes by Laplacian operator and global processor. Besides, an improved hybrid CNN-Transformer module is formed to fully extract image features by incorporating channel and spatial attention into the traditional Transformer structure. Finally, feature reconstruction is achieved through pixel sampling to obtain deraining images. Compared with other mainstream deraining methods on common datasets, the results show that the proposed network achieves better quantitative indicators and visual effects, confirming the effectiveness of the proposed method.

This study focuses on the research of AIGC (Artificial Intelligence Generated Content)-empowered collaborative learning, analyzing the evolutionary process of collaborative learning paradigms from traditional collaborative learning, computer-supported collaborative learning to AIGC-driven generative collaborative learning. It elaborates on the core elements of the AIGC-driven generative collaborative learning paradigm, including the role reconstruction of teacher-student agents, the supporting system of environmental agents, and in-depth interactions under agent collaboration. The study proposes its implementation path, which forms a technical supporting system through four mechanisms: self-guidance, collaborative creation, path evolution, and swarm intelligence. Meanwhile, it points out the structural breakthroughs of AIGC-driven generative collaborative learning, such as developing learners' meta-collaboration abilities, constructing a collaborative knowledge co-creation network and cross-domain collaboration ecosystem, and promoting the intelligent evolution of collaborative roles and dynamic optimization of group decision-making. The research indicates that AIGC propels collaborative learning towards the stage of smart education, providing a practical blueprint for the transformation of educational paradigms and pointing out the direction for the application and innovative development of educational technology, which is conducive to cultivating innovative and collaborative talents.

The parameters of the PWM wave generated by the existing methods, such as frequency and duty cycle, have limitations in terms of adjustment flexibility, making it difficult to meet the requirements of actual circuits for accuracy and flexibility. To address this issue, this paper proposes a new parameterized and modular PWM wave design scheme based on Field Programmable Gate Array (FPGA). This scheme introduces a counter inside the FPGA and achieves high-precision adjustable duty cycle and frequency of the PWM wave through parameter soft control. Moreover, it introduces delay triggers and logic operations inside the FPGA to configure the dead time. Take the PWM wave control of the Buck circuit for example, the FPGA collects the output current parameters in real time, generates feedback signals through a comparator, and adjusts the duty cycle of the PWM wave according to the feedback signal status, thereby achieving closed-loop control of the output current of the Buck topology. Through simulation verification by Quartus software and actual circuit testing, it can be achieved that the adjustable frequency range is 1 Hz to 50 MHz, with a step precision of 5 Hz at a frequency of 50 KHz; the adjustable duty cycle range is 0% to 100%, with a step precision of 0.1% at a frequency of 50 KHz; the minimum dead time is 20 ns, with a step progress of 20 ns; and the ripple coefficient of the Buck constant current circuit is 1.8%.

Network intrusion detection is a crucial approach in the field of cybersecurity, with anomaly traffic detection being a key technique in intrusion detection. To address the issues of high false alarm rates and low efficiency in traditional detection models, this paper proposes an anomaly traffic detection model based on MA-GRUCNN. XGBoost is used for feature dimensionality reduction, and the reduced-dimensional data is then fed into the MA-GRUCNN model. CNN is employed to extract high-dimensional features from the traffic data, an attention mechanism is used to capture global dependencies, and GRU is utilized to capture long-term dependencies in the time series. Experimental results on the NSL-KDD dataset demonstrate that the proposed method outperforms other approaches in terms of detection accuracy, precision, recall, and F1 score, achieving a detection accuracy up to 97.45%.

This study proposes a human posture action detection system based on image processing and machine learning, designed to automate fitness exercise counting and eliminate errors associated with manual tracking. The framework begins by capturing and decomposing test videos into sequential motion frames using OpenCV. The acquired images undergo preprocessing for enhancement, followed by the detection and normalization of 33 skeletal keypoints via the Mediapipe library. The normalized keypoint coordinates are then transformed into embedding vectors as model input. A K-Nearest Neighbors (KNN) algorithm compares these vectors against a reference database to classify and count exercise repetitions. Temporal smoothing refines the classification output, ensuring robust action recognition and accurate repetition counting. The system culminates in a user-friendly UI that visually presents the processed video stream alongside real-time analytics. By addressing the limitations of conventional counting methods, this research delivers an efficient, integrated solution for automated posture recognition and repetition tracking. Its practical implementation offers significant societal and economic value by enhancing workout efficiency and promoting health-conscious fitness practices.

Community detection has become a key issue in the field of complex network analysis, which essentially involves dividing the network into groups or communities, where nodes within the same community exhibit more dense connections than nodes outside their community. However, most existing label propagation methods are not stable enough. To this end, this paper proposes a community detection algorithm based on node importance. Firstly, node importance is calculated based on neighbor similarity. Then, label update order is rearranged based on node importance. Finally, label propagation is performed by combining node importance and label update order. The experimental results show that the proposed algorithm has more stable community partitioning results, and its modularity and standard normalization indicators have good performance.

Based on the Outcome-based Education (OBE) concept, this study addresses key challenges in Computer Introduction teaching in private colleges and universities, such as significant student heterogeneity, strong practical preferences, and insufficient learning autonomy, then constructs a "Three-in-One" curriculum reform framework. The framework integrates: (1) a three-tier teaching organization mechanism ("standard instruction-tiered training-personalized tutoring") to unify core knowledge standardization and personalized learning paths; (2) a trinity task system ("professional cognition-vocational skills-research methods") that combines tool application with disciplinary thinking; (3) a three-phase motivation system ("task-driven-team interaction-flipped certification") to activate learning engagement. Teaching practice demonstrates that this framework significantly enhances students' knowledge mastery, tool application skills, and career planning awareness. The study provides a replicable reform approach for introductory computer courses in private universities, offering practical guidance for cultivating application-oriented computing professionals.

With the rapid iteration of artificial intelligence technology, multi-agent modeling and simulation technology (MAMS) has developed rapidly and been successfully applied in various fields. This article firstly introduces the concept, technical advantages, research process, and application status of MAMS. It reviews the development of hybrid modeling and simulation combining multi-agent and system dynamics, modeling and simulation of multi-agent reinforcement learning, and modeling and simulation of large-scale multi-agent systems. A comparative analysis is conducted on the general multi-agent modeling and simulation platform. Finally, the existing MAMS platforms and multi-agent technology and tool platforms in the era of large models are introduced. This article systematically summarizes the current research status of MAMS, helping scholars quickly understand the development of MAMS in the AI era, summarizing the advantages and disadvantages of different MAMS platforms, and pointing out the direction for further development of MAMS.

Traditional speaker recognition methods based on convolutional neural networks (CNNs) often fail to capture long-term temporal dependencies and full-frequency information, limiting their robustness in complex acoustic environments. To address this, we propose LTEB-CAM++, a hybrid architecture combining a Context-Aware Masking Network (CAM++) with a Lightweight Transformer Encoder Block (LTEB). The LTEB module, inserted between CAM++'s FCM and D-TDNN layers, leverages Nyström-based self-attention to efficiently model global speech contexts (up to 10s), while the D-TDNN module extracts fine-grained local features. Trained on fused MFCC-FBANK features, LTEB-CAM++ reduces EERand minDCF by 7.98% and 12.58%, respectively, on CN-Celeb versus the CAM++ baseline, demonstrating superior efficiency and discriminability.

Under the promotion of the information and innovation strategy, in order to cultivate highly skilled talents that meet with the needs of the Internet of Things industry, this article takes the courses of C Language Programming and Microcontroller Technology and Application in the Internet of Things Application Technology major of vocational colleges as the research object, and uses literature research, case analysis, and practical verification methods to explore the integrated teaching path of the two courses. The research aims to form an integrated teaching system by integrating teaching content, innovating teaching methods, strengthening practical activities, building a teaching team, and constructing a teaching platform based on the Hi3861 chip. Practical case studies will be presented in conjunction with serial communication programming. The results show the significant improvement in students' comprehensive practical abilities, the increase in the number of awards in subject competitions, and enhanced interest in learning and recognition of domestic technology. Research shows that the integration of two courses in teaching can effectively connect knowledge systems, improve teaching quality, and provide practical references for the reform of relevant professional courses in vocational colleges under the background of information and innovation.

With the increasingly widespread application of voice evidence in judicial cases, grassroots public security organs are facing practical difficulties such as a shortage of voice testing professionals and disconnect between technical capabilities and case needs. As the main battlefield for cultivating public security talents, public security colleges undertake the dual tasks of serving practical needs and leading technological innovation. In order to adapt to the new situation, face new challenges, and better cultivate students' awareness of using voice to serve cases and their ability to use voice testing methods to serve cases, public security colleges and universities invite public security practical experts, judicial appraisal experts, and technical experts in the speech recognition industry to build a set of Forensic Speech Analysis course teaching system oriented towards case needs, focusing on teaching content, teaching resources, practical teaching, teaching methods, teaching assessment system, and teacher team. This paper aims to cultivate high-quality applied voice inspection professionals who understand the profession, can apply, can innovate and have professional ethics.

To address the issue of low detection accuracy in smart classroom behavior detection due to dense student populations, mutual occlusion, and large differences in target scales, an improved YOLO11-based classroom behavior detection model, ADU-YOLO11, is proposed. Firstly, the partial convolutional downsampling layers are replaced with the Adown downsampling module to reduce model complexity and minimize the loss of key feature information. Secondly, the dynamic detection head DyHead (Dynamic Head) is adopted, which enhances detection capabilities through scale, spatial, and task-aware attention. Finally, the UIoU (Unified-IoU) loss function is used to optimize the training convergence speed and improve the regression accuracy of the predicted bounding boxes. Experimental results show that ADU-YOLO11 achieves mAP50 and mAP50-95 improvement of 1.1% and 1.6% respectively, along with a precision increase of 2.7% and a recall increase of 0.6% compared to the original YOLO11n on the STBD-08 student-teacher behavior dataset. Moreover, it outperforms other object detection algorithms, demonstrating its effectiveness and superiority in smart classroom behavior detection.

The surface defects of steel rails are difficult to detect due to irregular shapes, scale differences, and background complexity, and the existing YOLOv5s model has limitations such as insufficient bounding box positioning accuracy and weak multi-scale feature extraction ability. This article proposes an improved detection method for YOLOv5s. Firstly, to address the issue of positioning deviation, the GIoU loss function is used to enhance the robustness of bounding box regression; Secondly, embedding CBAM attention mechanism to enhance the focus of defect area features; Finally, to address multi-scale defect detection, the BiFPN structure is introduced to achieve bidirectional weighted fusion. The experiment shows that the improved model can be applied to the dataset of surface defects on steel rails mAP@0.5 reaching 81.8%, an increase of 5.3% compared to the benchmark model, significantly improving the reliability of detection under complex working conditions and providing an effective solution for surface defect detection of rails.

The accurate prediction of the groundwater level is of vital importance for water resource management and disaster prevention and control. In view of the complex characteristics of the groundwater level time series, such as nonlinearity, multi-scale fluctuations and long-range dependence, traditional prediction models often fail to achieve satisfactory results. To improve the prediction accuracy of complex groundwater level time series, this paper proposes a prediction framework that combines specific data preprocessing with the CNN-Informer hybrid model. This framework first performs data cleaning, wavelet denoising and an adaptive resampling algorithm combining extremum and data volatility of the data to optimize the original high-frequency data. Subsequently, the CNN-Informer model is constructed. Local features are extracted using CNN, long-range dependencies are captured by the Informer encoder, and comprehensive predictions are made through the weighted fusion mechanism. Based on the verification of the measured data from Well Su 18 in Dantu, Jiangsu Province, the proposed CNN-Informer model significantly outperforms baseline models such as LSTM, CNN and individual Informer in key indicators such as MAE, RMSE and NSE, demonstrating its high accuracy and robustness in multi-step prediction. This research provides an effective new approach for the precise prediction of groundwater levels.

Addressing the challenges faced by general domain named entity recognition methods, which struggle to identify specialized terms and security entities within the cybersecurity domain, and suffer from insufficient feature extraction leading to low accuracy in cybersecurity entity recognition, this paper introduces a new model named Res-Inception Bi-LSTM-CRF (RIBIC). The RIBIC model leverages a Res-Inception Network to extract multi-granularity features, thereby capturing a richer set of feature information. Furthermore, an in-house cybersecurity domain-specific dictionary is developed, and a dictionary-based matching correction algorithm is incorporated to enhance the precision of entity recognition. The experimental results indicate that on two publicly available threat intelligence datasets, the F1 scores achieved are 94.09% and 83.91%, representing improvements of 15.02% and 15.72% over the baseline models, respectively. These findings robustly validate the effectiveness of the proposed method for named entity recognition in the threat intelligence domain.

Anomaly detection in borehole strain data as earthquake precursors is of great significance for earthquake prediction. This paper takes the borehole strain time series provided by the Jiangsu Earthquake Agency as experimental data. Firstly, a preprocessing method combining trend separation and residual denoising is used to effectively extract the long-term trends and critical fluctuation information from the data, thereby reducing noise interference. Subsequently, an LSTM-Attention model is designed and its performance is compared with three other models: LSTM, CNN, and AutoEncoder. The comparison results indicate that the LSTM-Attention model outperforms the other models in terms of MSE, MAE, and RMSE. Finally, this model is used for anomaly detection. Experimental validation demonstrates that the LSTM-Attention model can accurately capture anomalies in the data for the task of anomaly detection in borehole strain precursors of earthquakes.

This paper proposes an unsupervised log anomaly detection method based on an improved variational autoencoder generative adversarial network (VAE-GAN) to address the issues of instability and interdependence in log sequence data. The proposed model combines the advantages of GAN and VAE by embedding the temporal convolutional network module into the encoder, decoder, and discriminator, effectively capturing the distribution of log sequence data and optimizing the sequence mapping in the latent space, thereby achieving high-precision reconstruction of normal log sequences. The model continuously improves the reconstruction ability of the variational autoencoder through adversarial training mechanism, enabling it to identify abnormal patterns in the log more accurately. The experimental results show that compared with other unsupervised methods, this method has better performance on public log datasets.

This paper addresses the pain points of traditional campus Q&A systems, such as "information silos", response delays, and differentiated needs, by proposing and implementing an intelligent Q&A platform that integrates Large Language Model with Retrieval-Augmented Generation (RAG). Through constructing a hierarchical system architecture, the platform combines three core modules - knowledge base construction, hybrid retrieval, and intelligent generation - to achieve deep integration of knowledge retrieval and intelligent Q&A. Data demonstrates that the platform achieves 92.2% intent recognition accuracy and 89.9% answer accuracy for high-frequency queries like course schedules and campus policies, while supporting multi-role services and dynamic knowledge base updates.

Aiming at the problems of poor independent learning ability of students, single teaching resources, unilateral passive acceptance of knowledge by students, insufficient combination of theoretical and practical classes, insufficient combination of teaching methods and new technologies in traditional teaching of object-oriented programming (C++) course, this teaching innovation proposes a blended teaching plus large model-assisted teaching program, and describes the process of program implementation in detail. The program implements the teaching concept of "student-centered", adopts the teaching mode of “online + offline”, and carries out detailed teaching design and practice of the teaching process from the four dimensions of pre-course, in-course, post-course, and evaluation and feedback, so as to realize all-round enhancement of students' abilities. At the same time, it adopts the large model to assist the teaching implementation, realizing the transformation from "knowledge transmission" to "ability cultivation". The practical results show that the program can realize students' personalized learning, improve students' independent learning ability, increase teacher-student interaction, and enhance the quality of teaching while greatly improving students' social competitiveness.

To enhance the teaching effectiveness of Web Front-End Development in secondary vocational schools and meet the industry's demand for innovative technical and skilled talents, this study introduces the SCS Maker Teaching Method into teaching practice. Employing a controlled experimental method, two secondary vocational classes with initially equivalent academic levels are selected: the experimental class implements the teaching design based on the SCS Maker Teaching Method, while the control class adopts traditional teaching methods. The research results show that students in the experimental class significantly outperform their counterparts in the control class in terms of classroom participation rate, the proportion of independently developed innovative interactive functions, and the integrity of project codes. It is concluded that the SCS Maker Teaching Method can effectively stimulate secondary vocational students' interest in learning, significantly improve their practical and innovative abilities, and strengthen their team collaboration spirit and professional literacy, thus providing an effective practical path for the reform of Web Front-End Development in secondary vocational education.

The ''Posts, Courses, Competitions, and Certificates'' comprehensive education model is the focus and feature of the current reform of vocational education, which is a kind of teaching and training model that integrates the work position, educational curriculum, skills competition and vocational skills certificate in depth. Take the course of Microcontroller Technology and Application for example, based on the ''Posts, Courses, Competitions, and Certificates'' model, this paper integrates the requirements of job ability, combines the vocational skills competition and skill level certificate, reconstructs the course content and the ability standard. Based on the ''Posts, Courses, Competitions, and Certificates'' education model, the course is integrated with the requirements of job competence, combined with the vocational skills competition and the skill level certificate. Piloting the reformed model, the practice shows that the reform and practice of the course meet the requirements of industrial development ability, and the students' ability in theoretical knowledge and practical skills meets the expectation, which improves the students' vocational competence.

With the rapid development of digital and intelligent e-commerce, the live streaming e-commerce industry is facing multiple challenges such as fragmented user demands, intensified market competition, and insufficient marketing efficiency. Based on user behavior data, this study constructs an extended RFM-QPC-V user characteristic model, and uses the K-Means clustering algorithm to generate user portraits, accurately depicting users' consumption preferences and value hierarchies. On this basis, a gradient boosting model is adopted to design a dynamic voucher delivery strategy, and the model performance is optimized through the SMOTE oversampling technique and the ENN under-sampling method. Experimental results show that the recall rate of this model reaches 0.7 and the AUC-ROC value reaches 0.88. It performs excellently in accurately identifying high value users and improving the conversion rate, providing effective technical support and decision - making reference for live streaming e-commerce. In addition, this study proposes a layered distribution and regional marketing strategy, designing differentiated coupon schemes for different customer groups, aiming to empower the efficient development of the smart digital e-commerce industry.

In the educational domain, processing and analyzing large-scale datasets are crucial for enhancing teaching quality and student learning outcomes. This study proposes a novel hybrid ensemble learning framework incorporating ordered feature processing—the Ordered Adaptive Extra-Forest (OAEF) model. Leveraging the UCI Machine Learning Repository's Portuguese higher education evaluation dataset, we establish a dual-channel predictive architecture through feature space decomposition: A Regularized Extreme Machine Tree (REMT) constructs probability distribution predictions for ordered features, while Extra-Trees algorithm models unordered features in parallel. Predictive outcomes are integrated via a dynamic weighted fusion strategy. Comparative experiments against benchmark models (Random Forest, Support Vector Machine, etc.) demonstrate OAEF's superior performance, achieving accuracy (98.73%), precision (98.66%), recall (98.43%), and F1-score (98.13%), with respective improvements of 0.72%, 0.56%, 0.49%, and 0.26% over conventional methods.

At present, domestic civil aviation aircraft has l implemented PED (Portable Electronic Device) startup and internet access restrictions. While this measure provides passengers with a better service experience, there are also potential risks to the operation of onboard aviation electronic devices. In order to better study the electromagnetic interference threat of PED radiation to aviation electronic devices, a design method for a PED simulator is proposed. This simulator can simulate the real scenarios of cabin passengers using electronic devices such as mobile phones and PADs, generate radiation signals identical to those of actual devices, and realize centralized control and management. It eliminates the need for frequent manual operations, making it more convenient to carry out PED interference testing experiments. The simulator has been tested on a domestic civil aviation aircraft. By comparing with the interference data of real PEDs, it is found that the interference generated by the simulator is basically consistent with the real signals. Therefore, this achievement can be widely applied to PED interference testing scenarios of civil aviation aircraft.