In recent years, the application of 3D Gaussian splatting technology in novel view synthesis has made remarkable progress, achieving significant improvements in rendering speed and efficiency. Unlike the previous Neural Radiance Fields (NeRF), which are based on implicit representations, 3D Gaussian splatting models scenes using a set of Gaussian ellipsoids and rasterizes these ellipsoids onto images to achieve efficient rendering. This explicit representation not only enhances rendering efficiency but also provides crucial support for subsequent tasks such as dynamic reconstruction, geometric editing, and physical simulation. This review aims to outline its development trajectory, analyze key technologies, and explore current applications and challenges, providing a comprehensive reference for researchers in related fields. The study further indicates that the next phase of development for 3D Gaussian sputtering technology will shift from pursuing single-point breakthroughs in rendering speed and static quality to addressing system-level challenges such as scalability, dynamic realism, generalization capability, and interactive intelligence. This advancement will provide core support for major application fields including digital twins, mixed reality, and autonomous systems.

With the rapid development of Artificial Intelligence technology, programming education is transforming from traditional grammar instruction to cultivating computational thinking and practical problem-solving abilities. Python, as a widely adopted programming language, urgently needs innovative teaching methods. While project-based learning aligns with the inherent needs of programming education, it still faces challenges in practice, such as limited project selection, high student cognitive load, and simplistic evaluation methods. The emergence of Generative Artificial Intelligence offers a new path to address these issues. Take DeepSeek as an example, this paper systematically constructs a Python project-based teaching model empowered by Generative Artificial Intelligence. Focusing on three stages—project introduction, activity exploration, and results presentation—it elaborates on the collaborative mechanism between teachers, students, and GenAI in six stages: project selection, planning, activity exploration, project creation, results sharing, and activity evaluation. It provides a systematic theoretical framework and instructional design reference for Python project-based teaching supported by generative artificial intelligence, laying the foundation for subsequent empirical research and practical exploration.

Aiming at the problems of high learning difficulty, disconnect between theory and practice, and insufficient interdisciplinary adaptability in Artificial Intelligence general education courses in higher vocational colleges, this study proposes the "Intelligence and Application Integration" dual-driven teaching philosophy. It constructs a "Five-Dimensional Integrated" curriculum framework consisting of "Basic Introduction - Theoretical Support - New Technology Application - Tool Application - Comprehensive Inspiration". Taking intelligent tool empowerment as the practical carrier and intelligent thinking leadership as the cognitive orientation, the study integrates cutting-edge technologies such as large models and agents, and aligns with emerging professional demands like AI trainers. This approach effectively enhances the intelligent thinking level and interdisciplinary application ability of students from different majors, providing a systematic teaching solution to solve the current curriculum dilemmas.

Frequent flood disasters pose a serious threat to people's lives and property safety. Rapid and accurate monitoring of flood inundation areas and their dynamic changes is crucial for disaster emergency response and assessment. To meet this demand, this study uses the flood disaster in Zhuozhou City, Hebei Province, from July to August 2023 as a study case. Utilizing Sentinel-2 multi-temporal optical imagery, we conduct the identification and dynamic change monitoring of floodwater bodies. By integrating spectral reflectance with various water body indices, we construct comprehensive identification features and separately developed Artificial Neural Network (ANN), Random Forest (RF), and Support Vector Machine (SVM) models for water body extraction. The results indicate that the RF model exhibits a water body identification accuracy exceeding 98.84% across different time phases, with a Kappa coefficient greater than 0.97. Furthermore, its stability and efficiency surpass those of the SVM and ANN models. Based on the time-series monitoring results obtained from RF, the complete process of "rapid inundation - slow recession" during the Zhuozhou flood disaster is revealed. During the recession phase, due to topographical and drainage conditions, some areas experience persistent waterlogging. This study combines the accessibility of Sentinel-2 data with the efficiency of the RF model, offering a high-precision and high-efficiency emergency monitoring solution for flood disasters.

To solve the long detection time in traditional soil nutrient testing, the shortcomings of transmission distance limitation in currently available portable equipment, and the susceptibility to detection precision degradation caused by temperature interference, a cloud-based portable sensor for the detection of basic soil nutrients is introduced in this paper. The sensor uses STM32F407 as the main control chip and employs a 485 type, five-pin, four-parameter soil sensor, which can measure the temperature, moisture, pH value and NPK at the same time, and realizes the data acquisition of the four parameters through the Modbus RTU protocol. The FS-MCore-F800 series highly integrated 4G Cat.1 module is adopted to implement real-time remote transmission of detection data and remote monitoring functions. Through testing, the device offers extensive signal coverage and stable communication under weak signal conditions. It supports integrated detection of multiple sensors via RS-485, while reducing moisture measurement errors through a random forest-based temperature compensation model. The cloud platform enables multi-terminal access, and the device features excellent operability, presenting broad promotion prospects in smart agriculture.

Addressing the prominent issues in traditional Python Programming courses, such as dynamic adjustment, practical feedback, and teaching precision, this paper constructs a "1+1+N" teaching AI intelligent platform. Guided by the educational philosophy of "integrating intelligence and education, tripartite collaboration", and based on the "one body, two wings" architecture, the platform achieves precise support and dynamic optimization throughout the teaching process. Practical results demonstrate that this reform significantly enhances students' programming skills, learning engagement, and course satisfaction, promotes the transformation of teachers' roles towards advanced education, and establishes a data-driven, teacher-student collaborative teaching ecosystem. This provides a replicable practical pathway for curriculum reform in applied undergraduate institutions.

In response to the demand for highly skilled talents in the new generation of information technology industry, and to address issues such as the disconnection between industry and education, the singularity of evaluation methods, and insufficient integration of ideological and political education in computer courses in higher vocational education, a classroom teaching model featuring "dual-line guidance, three-stage progression, six E's linkage, and multi-dimensional evaluation" has been constructed, using the Windows Server operating system management course as the carrier. This model integrates value shaping and skill cultivation, aligns with job standards based on the three-stage abilities of "network construction, network management, and network maintenance", creates real work scenarios through the "six E's" teaching process, and establishes a multi-subject whole-process evaluation system covering knowledge, abilities, and qualities. Practice has shown that this model effectively synchronizes course content with industrial technology, integrates teaching processes with work tasks, and coordinates educational effectiveness with social services, demonstrating good application effectiveness and promotion value.

To address the limitations of traditional micro-landscape systems, which rely on manual maintenance and lack ecological self-regulatory capabilities, an intelligent micro-landscape control system based on the K210 chip has been designed. This system incorporates adaptive environmental regulation and human-machine affective interaction functionalities. By integrating sensors for real-time monitoring of temperature, humidity, and light intensity, it autonomously controls misting and lighting to maintain ecological equilibrium within the micro-landscape. Utilizing Wi‑Fi technology, the system enables remote transmission and storage of environmental data. Innovatively, the KPU + YOLO‑K210 model has been incorporated to achieve facial expression recognition, allowing the system to perceive user emotions and dynamically adjust the ambient atmosphere, thereby delivering personalized experiences such as stress relief, aesthetic enhancement, and anti‑fatigue effects. Experimental results demonstrate that the YOLO‑K210 model achieves a recognition accuracy of 0.874, with all system functions operating stably and reliably. These findings validate the feasibility of establishing bidirectional "environment-human" interaction in intelligent micro-landscape systems and provide a referential framework for their systematic design.

Case-based practical teaching methods are widely used in the teaching of Artificial Intelligence Technology courses, but there are common problems such as low student participation, serious "toy-like" of cases, and disconnection from actual talent demand. To address these issues, this paper explores a project-case-guided practical teaching method oriented by talent market demand and aimed at promoting the subjective initiative of students. Based on typical and currently mainstream artificial intelligence technologies, the method integrates demands and resources from multiple sources, including scientific research projects, national Class-A competitions, and enterprises, to design a three-level, full-process project case system of "basic-advanced-practical". Through phased and step-by-step guidance of cases, it enables students to independently learn theoretical and practical knowledge of the course, enhancing their hands-on abilities while improving their competitiveness in industry employment.

In the C language knowledge question answering system, text similarity calculation is very important, but traditional methods, such as cosine similarity and editing distance, have problems such as inability to deeply understand semantics and difficulty in dealing with language complexity. To this end, this study proposes a C-knowledge text similarity calculation model (BMHA) that fuses bidirectional long short-term memory network (BiLSTM) and attention mechanism. The model takes advantage of the context-aware ability and attention mechanism of BiLSTM to focus on key information to dig deep into text semantics. Experimentally, on the platform built based on the TensorFlow framework, the F1 value of the BMHA model in the semantic matching task is 92.1%, which is 19.7% and 24.5% higher than that of the traditional LSTM and CNN benchmark models, respectively. It shows that its performance is excellent, which provides a new method for text similarity calculation in the field of programming education, and is expected to promote the development of related fields.

To address the issues of feature redundancy and noise interference caused by modal heterogeneity in multimodal sentiment analysis, this paper proposes an adaptive collaborative perception model based on multi-task joint learning (AdaSP-MTL). The AdaSP-MTL framework decouples unimodal and multimodal sentiment tasks and constructs a heterogeneous feature disentanglement space to enhance modality-specific modeling. A dynamic gated fusion module based on inter-modal cosine similarity is designed to adaptively allocate fusion weights, thereby suppressing cross-modal noise propagation. Furthermore, the model integrates low-level local features extracted by CNNs with high-level sequential features captured by BiGRUs to form multi-scale sentiment representations. Experimental evaluations on the SIMS, MOSI, and MOSEI datasets demonstrate that AdaSP-MTL comprehensively outperforms existing baseline models in sentiment classification tasks, validating its advantages in effectively modeling modal heterogeneity, suppressing noise interference, and constructing robust sentiment representations.

Evaluating students' vocational abilities is a key point in assessing the quality of talent cultivation in New Engineering disciplines in higher vocational colleges. However, at present, there are relatively few studies on the vocational ability assessment of New Engineering students. Based on the COMET capability model and typical work tasks, this study establishes a vocational ability assessment system for students majoring in New Engineering. A total of 244 students from a certain vocational college are selected as the assessment subjects. Based on three ability levels and 24 ability indicators, evaluation indicators and scoring standards are formulated. Open-ended comprehensive test questions, background situation questionnaires and answering motivation questionnaires are designed. The assessment results show that 80.4% of the students have reached the process ability level, while only 5.3% have reached the design ability level. Students lack environmental protection awareness and innovation awareness when completing tasks. This research will provide a scientific reference for the reform of talent cultivation models in colleges and universities, and help them optimize their curriculum systems, teaching models and evaluation systems.

This paper constructs an intelligent teaching evaluation system for electrical and electronic experimental course group based on the "6E" instructional model, integrating evaluation mechanisms into the five teaching stages - Elicit, Engage, Explore, Explain, and Elaborate, to form a closed-loop feedback structure. Then an intelligent evaluation platform for the "Three Electrical Courses" experimental curriculum cluster has been developed to support process-oriented and personalized evaluation. Practice has shown that the system effectively enhances students' comprehensive abilities and innovation competence, while providing an effective evaluation pathway and decision-making support for the reform of experimental teaching.

To address the issues of strong subjectivity, delayed feedback, and single dimension in traditional classroom teaching evaluation, and to meet the personalized and capability-oriented needs of new engineering education, this paper constructs a multi-modal intelligent classroom teaching evaluation model driven by whole-process collaboration. This model collects data from visual, auditory, operational, and behavioral aspects through a distributed sensing network. After preprocessing and feature extraction, it achieves collaborative analysis of multi-source data relying on a dynamic hierarchical and adaptive fusion framework. The model is practiced in the Java Programming Technology course, and the results show that the recognition accuracy of the multi-modal fusion model is highly consistent with teachers' manual evaluation. It can achieve full-dimensional evaluation from teaching theory to practice, providing new ideas for intelligent evaluation of programming courses in new engineering education.

In the era of LLMs, artificial intelligence technology permeates various fields, promotes industrial upgrading, and also puts forward the need for reform and innovation in practical teaching of computer courses in vocational colleges. In response to the problems of vague goals, scattered paths, and disconnection from job positions in the current integration of AI into teaching, this article uses literature research, investigation, and case analysis methods to sort out the current situation, construct a three-level goal practice teaching optimization path of "cognition-application-innovation", and clarify its connotation, characteristics, and practical methods. The results show that this path can enhance students' AI application ability and innovative thinking. The excellent rate of computer practice course assessment in pilot classes has significantly improved, and the number of awards in AI skill competitions has significantly increased. The conclusion indicates that this path can accurately meet the talent training needs in the era of large models, provide an implementation framework for improving teaching quality, and enrich the research results of the integration of AI and vocational education.

In view of the requirement that efficiency and reusability for signal processing modules in software-defined radio (SDR) systems, this paper designs and implements a reconfigurable signal processing platform based on DSP and FPGA. The platform adopts a layered architecture, enabling flexible waveform configuration and efficient management through hardware resource isolation and modular design. Waveform software loading is performed on the DSP side, where the FPGA configuration file is loaded using the Slave SelectMap method, combined with EDMA (Enhanced Direct Memory Access) based data transfer to accelerate the system reconfiguration process. Test results demonstrate that the proposed design achieves excellent real-time performance, stability, and scalability, providing valuable insights for the design of signal processing platforms in SDR systems.

For the problem that finite element modeling mostly uses half manual, even pure manual to conduct vertebra segmentation, this paper proposes an automatic segmentation algorithm. This algorithm firstly marks seed points in segmentation region with the multi-resolution composite feature vector based point landmark algorithm, then grows the segmented regions using the method of region growing, finally achieves the automatic segmentation. Experimental validation using 120 spinal MRI images demonstrates that the proposed method achieves an average similarity coefficient of 90.02% for segmentation results, with a maximum of 95.28%. The segmentation process requires only 5.0 minutes, representing over a tenfold improvement in efficiency compared to manual segmentation and significantly outperforming traditional graph theory-based segmentation methods and U-Net algorithm. Conclusions indicate that the proposed algorithm enables efficient and accurate automatic segmentation of vertebral bodies, demonstrating good stability and application potential. It provides a reliable foundation for subsequent 3D reconstruction and finite element analysis.

Under the backdrop of the rapid development of government informatization, the security of government cloud platforms is of paramount importance. Based on the Technical Specification for Cryptography and Related Security of Certificate Authentication System, this study focuses on the PaaS and IaaS layers of government clouds and constructs a multi-level and multi-mode cryptographic resource service system through systematic transformation. At the IaaS layer, dedicated cryptographic servers and virtualization technology are adopted to achieve distributed cryptographic services. At the PaaS layer, comprehensive cryptographic services such as SM2, SM3, and SM4 are provided based on a microservice architecture to ensure the integrity, confidentiality, and availability of data. By integrating technologies such as signature verification, SSL/TLS, and electronic signatures, a comprehensive security system covering transmission encryption, data verification, tamper-proofing, and non-repudiation is established. This research significantly enhances the security of government cloud platforms, realizes the manageability and controllability of data throughout its lifecycle, and provides all-round encryption protection for sensitive information in government clouds. In the future, with the wide application of commercial cryptography technology and the rapid development of the cryptography industry, this study provides an important reference for the security construction of government cloud platforms and points out the direction for the innovation and application of cryptography technology.

Aiming at the converging waves of Artificial Intelligence, the Internet of Things, and big data, the course Principles and Applications of Embedded Systems is confronted with rapidly evolving knowledge, high practical thresholds, and fragmented teaching resources. Under the new paradigm of AI-empowered smart classrooms, cultivating self-directed, high-level, innovative talent has become the primary mission of universities. We therefore propose an AI-enabled blended-teaching reform and exploration for embedded courses: an embedded-AI curriculum is created and an AI-empowered embedded-course system is constructed. Online-offline blended instructional designs are developed in which AI agents permeate the entire learning chain—pre-class diagnosis, in-class interaction, and post-class extension—dynamically generating "one thousand faces for one thousand students" learning paths and realizing true individualized instruction. Simultaneously, a multi-data-driven, diversified learning-assessment model is introduced to accurately profile knowledge and form an iterative, interpretable “AI+Embedded” teaching ecosystem, offering a new roadmap for transcending traditional blended-teaching models.

To address the issues of large size, heavy weight, and complex operation inherent in existing air pressure generator, this article designs a portable numerical control air pressure generator based on STC. The device consists of modules such as the STC8H8K64U microcontroller, a 370 motor pump with an encoder, the MCP_H10 series pressure sensor, and the ADS1115 analog-to-digital conversion chip. The MCP_H10 pressure sensor is used for real-time detection of the current output air pressure value and compares it with the preset control air pressure value. The main control uses a PID algorithm to control the air pump motor to produce a stable air pressure output. The system can display the generated air pressure value on an LCD screen, a host computer, and a mobile phone client. It features miniaturization, portability, and low cost. Tests show that the average relative error of the device in generating air pressure from 0kPa to 10kPa is 3.78%, and within the range of 10kPa to 100kPa, the average relative error is 0.45%. It can be used for portable air pressure generation in outdoor settings.

Illegal opium poppy cultivation poses a serious threat to social security and is explicitly prohibited by law in China. Current poppy patrols primarily rely on drone imagery combined with manual identification, which suffers from low efficiency, susceptibility to fatigue, and high professional requirements. To address these issues, this paper designs an intelligent patrol and detection system based on a multi-rotor drone and the Jetson Orin NX edge computing platform. This system integrates high-definition image acquisition, edge AI real-time inference, a 5G communication module, and web visualization into a cohesive unit. The system utilizes an object detection model on the onboard edge computing platform to analyze images in real-time. The results are then uploaded via the 5G communication module to a Web interface for review and processing by ground personnel. Practical application demonstrates that this system significantly enhances the automation level, response speed, and detection accuracy of poppy patrols, providing efficient and reliable technical support for narcotics control efforts.

In response to the common lack of backend monitoring data in research institutions and the difficulty in directly applying industry-standard estimation models, this paper proposes a historical data-based model for estimating digital infrastructure resource demand. The model categorizes business systems into four types (general-integrated, operating-management, R&D trial, and knowledge-management systems), defines key resource variables and system structures, and incorporates parameters such as time intervals, project approval rates, funding trends, and digital transformation phase coefficients. It further establishes estimation methods—mean estimation and maximum-value estimation—suitable for varying levels of data availability. The model aims to provide a scientific and actionable basis for infrastructure resource planning in research institutions lacking real-time monitoring capabilities.

To address the issues of scarce annotated samples and insufficient feature representation capability in sign language recognition tasks, this study proposes a pre-training method for sign language visual features based on self-supervised learning. By constructing a spatial-temporal joint augmentation mechanism and a contrastive learning framework, spatiotemporal feature representation is optimized under unsupervised conditions. A dual-branch encoding architecture integrating C3D and Transformer is designed to balance local motion dynamics and global semantic dependencies. An occlusion reconstruction auxiliary task is introduced to enhance the model’s ability to recover key motion regions and missing information. After pre-training, a frozen transfer strategy is adopted to migrate features to downstream sign language recognition tasks, combined with BiLSTM-based temporal modeling and CTC-based classification decoding to build the overall system. Experimental results on the RWTH-PHOENIX-Weather 2014T dataset demonstrate a recognition accuracy of 85.7% and a character error rate reduced to 10.3%, significantly outperforming traditional transfer learning approaches and validating the effectiveness and generalization advantages of the proposed self-supervised pre-training mechanism in low-annotation resource scenarios.

Drawing on the teaching practices of low-code application development courses in application-oriented universities, this study analyzes both their positive significance and existing challenges. To address critical challenges in course positioning, instructional depth, platform selection, and the balanced integration of AIGC, this paper introduces a new teaching paradigm for low-code education. This paradigm is designed to foster thinking cultivation at its core through a tiered and diversified implementation approach. The paradigm differentiates between the training objectives of computer science and non-computer science majors, with tailored emphases and distinct instructional trajectories. Moreover, the study examines the prospects of low-code application development under the disruptive impact of large-scale AI models, advancing the viewpoint of symbiosis and a collaborative development framework between the two. The research provides both a theoretical foundation and a replicable practical scheme for the construction of low-code curricula in application-oriented universities.

5G broadcasting technology is one of the key development directions of 5G networks, offering television and mobile users a more diversified and high-bandwidth broadcast television experience. By analyzing the current development status of 5G broadcasting, it is possible to identify development directions and business models for traditional telecom operators in the commercial application of 5G broadcasting. Traditional telecom operators can explore multiple development directions in the 5G broadcasting field, including spectrum refarming, integrating traditional live broadcasting with self-media live streaming, exploring innovative formats, building an "all-media" platform, and supporting and developing 5G core technologies, in order to provide users with a diversified, high-bandwidth broadcast television experience. Through in-depth analysis of development paths, the transformation of traditional telecom operators' business models can be promoted.

To address issues such as incomplete semantic representation, significant semantic variation, and low classification accuracy for low-resource languages in multilingual text classification, thereby enhancing the overall accuracy of multilingual text classification, we propose a dual-channel model based on adaptive gated feature fusion, employing two parallel channels to process text: one channel encodes text via an InfoXLM embedding layer to preserve original language-specific features and semantic information; the other channel integrates adversarial training after InfoXLM embedding, using FGM to perturb inputs to eliminate language-specific biases and enhance the universality of cross-lingual semantic representations. The hidden states from both channels undergo dynamic weighted fusion via a gating module, enabling information exchange between channels. Experiments on multilingual toxicity detection and sentiment datasets demonstrate that the proposed method outperforms comparison models across all evaluation metrics. This approach effectively enhances multilingual text classification accuracy and improves classification performance for low-resource languages.

Aiming at the problems of inaccurate user intention recognition and low efficiency of information interaction in campus network information retrieval system in high concurrency scenarios, this paper proposes a dynamic retrieval method of campus network information based on improved BERT (Bidirectional Encoder Representations from Transformers) model. This method first utilizes an enhanced BERT model for pre-training language representations, combined with classification algorithms to optimize user intent recognition. Next, a dynamic information interaction module is designed to construct a retrieval model based on the identified user needs and organize information according to the retrieval results. Furthermore, an N-gram statistical language model is introduced for grammar correction and natural language generation, improving the fluency and standardization of the output text. Experiments show that the system achieves an information interaction accuracy of 97.5% under high-concurrency conditions, significantly outperforming traditional retrieval methods and enabling efficient and stable dynamic information retrieval on campus networks.

Recently, scribble-based weakly-supervised learning methods have shown great potential in medical image segmentation, achieving promising results at a low annotation cost. However, the limited supervisory signals often lead models to overemphasize the main body of the target while overlooking its precise boundaries, consequently resulting in significant segmentation errors along the edges. To tackle this issue, we propose a Weak-Spatially-Agile-Mamba-UNet (WSAM-UNet) model. This architecture integrates a convolutional U-Net sub-model for local feature extraction, a dynamic adaptive attention U-Net sub-model for context-aware pixel perception, and a Visual Mamba U-Net sub-model for global contextual modeling. Extensive experiments on the ACDC dataset demonstrate that our WSAM-UNet achieves an overlap rate of 0.893 19 with the ground truth, while reducing the Hausdorff Distance and Average Surface Distance to 4.186 5 and 1.035 2, respectively. Compared to several strong baselines, WSAM-UNet achieves an optimal balance between overall segmentation accuracy and boundary error control.

With the rapid development of flexible electronics, wearable devices, and implantable medical electronics, traditional rigid filters can no longer meet the requirements of new electronic systems due to their inability to adapt to dynamic bending conditions and poor spatial compatibility. To address the above issues, a flexible LC filter is proposed. The proposed structure is implemented on a low-cost flexible printed circuit board (FPCB). The introduction of coupled capacitors significantly reduces the device size and loss; in addition, the design of the double-helical inductor reduces the required area and loss, greatly improves the integrability, and enhances the bending reliability. Furthermore, two resonant networks (R1 and R2) form a second-order filtering response through coupled capacitors, thereby constructing the flexible filter and achieving the goals of high spatial compatibility and low loss. Finally, through fabrication and testing, the proposed flexible filter has a center frequency of 11.4 MHz, a 3 dB fractional bandwidth of 19.93%, and an actual insertion loss of 1.52 dB at the center frequency.