Abstract：The traditional Chinese sensitive feature recognition methods in the field of power material supply chain mainly identify the flat feature entities with a single structure, and are difficult to target the nested feature entities with complex structures. These nested feature entities with complex structures have diverse structures and complex levels, so the traditional sensitive feature recognition methods are difficult to accurately identify and analyze. Therefore, this paper proposes a multi-structure Chinese sensitive feature recognition model based on semantic matching, which redefines the sensitive feature entity recognition task as a semantic matching problem. The model adopts a two-tower model structure, encodes the target entity feature categories as cue words, and aligns semantically with the sequence segments in the sentence. This method can effectively identify various feature entity segments including flat and nested structures. The results show that the accuracy, recall rate and F1 value of the SFSM model proposed in this study are better than the current mainstream models and other fusion models.

Abstract：This paper proposes a method for testing the credibility of a grid dispatching AI algorithm based on an exponential hierarchical structure algorithm, aiming to evaluate the model performance and robustness of the grid dispatching AI algorithm in multiple temporal and spatial scenarios. Firstly, this paper defines a set of credibility assessment criteria, covering the algorithm′s accuracy, robustness and generalization capabilities. Then, using the exponential hierarchical structure algorithm embedded with a knowledge graph, the performance of the AI algorithm (such as LSTM) is verified through multi-level temporal divisions. During the testing process, the grid dispatching tasks are divided into different time levels (short-term, medium-term, long-term) and spatial levels (different geographical regions and grid topologies) for testing. In each level, the assessment criteria are weighted and integrated through exponential weights, and ultimately a comprehensive credibility score is generated. Experimental results show that this method can effectively identify the performance and potential risks of the grid dispatching AI algorithm in complex scenarios, providing new technical support for the verification of the safety and robustness of grid dispatching AI.

Abstract：To address the challenges in decision modeling and solution methods for collaborative task allocation, this study investigates multi-agent collaborative task allocation technologies to achieve efficient coordination and dynamic task distribution among multiple agents. A collaborative task allocation method based on deep reinforcement learning is proposed, utilizing the deep Q-network (DQN) algorithm to construct agents and simulate environments, thereby optimizing the efficiency and effectiveness of task allocation. By constructing typical task scenarios and designing detailed decision model training and application schemes, the effectiveness of the proposed algorithm in collaborative tasks is verified. Experimental results demonstrate that the method can balance multi-objective comprehensive task performance, achieving an approximate 17% increase in success probability. The multi-agent collaborative task allocation method based on deep Q-network can effectively improve the efficiency of task allocation.

Abstract：With the rapid development of domestic new energy vehicles, the demand for master MCUs equipped with domestic encryption algorithms has been growing steadily. To meet this demand, this paper designs a universal hardware security module for MCUs based on the SM4 cryptographic algorithm. The module employs the SM4 algorithm as its core engine and supports mainstream modes such as ECB, CBC, CFB, OFB, and CTR. It adopts a standard AHB bus interface, supports DMA transmission, and incorporates low-power design techniques, making it suitable for current mainstream MCU circuit designs. The module was designed and manufactured based on a commercial 40 nm CMOS process. Post-tape-out testing and verification confirm that the SM4 algorithm module functions correctly, operates at a frequency of up to 400 MHz, and holds significant commercial value。

Abstract：In response to the voltage requirements of anti-fuse OTP memory in different working modes, a level shifter unit and block power supply design with Cascode structure is proposed based on the SMIC 28 nm CMOS process. Realize the rapid switching of WL/PL and BL terminal voltage of OTP memory in programming and reading mode. The flip delay of this level shifter unit is less than 2 ns. The block power supply design reduces the impact of capacitor parasitic effects on the read threshold of the OTP memory, ensuring the normal programming and reading functions of OTP memory.

Abstract：This paper presents a novel constant-gm rail-to-rail operational amplifier with enhanced slew rate. The input stage utilizes an improved redundant differential pair incorporating series-transistor and current-mirror techniques to reduce subthreshold current and minimize transconductance variation to only 1.4%. Supported by slew-rate enhancement circuitry, the design improves dynamic response by utilizing MOS voltage characteristics across operating regions. Ahuja compensation is employed to simultaneously optimize power consumption and area while maintaining adequate phase margin. Implemented in a 40 nm CMOS process under a 3.3 V supply, the amplifier supports full rail-to-rail input/output. The slew rates reach 7.52 V/μs (positive) and 7.87 V/μs (negative), improved by 50%. It attains a DC gain of 114 dB, a phase margin of 79°, and a unity-gain bandwidth of 3.7 MHz. It attains a CMRR of 135 dB and PSRR of 90 dB. The proposed op-amp is well-suited for high-precision and low-power analog applications.

Abstract：The logging-while-drilling nuclear magnetic resonance (LWD NMR) instrument can provide in-situ formation porosity and permeability measurements unaffected by mud invasion, but it poses severe challenges to adaptability under extreme conditions and system integration and reliability. Currently, the LWD NMR instrument technology is mainly held by foreign energy giants, and domestic LWD NMR instruments are still in the catching-up stage. To improve the development efficiency of LWD NMR logging instruments, simplify the instrument testing steps, and shorten the project cycle, a data acquisition and processing software based on LabVIEW language was designed and implemented according to the characteristics of LWD NMR logging instruments. This software is based on the RS-485 serial interface and a custom logging command transmission and data upload communication protocol, and it realizes basic functions such as LWD NMR sequence parameter configuration, data transmission and reception, data processing and display. Experimental results show that in the actual instrument testing, the nuclear magnetic spin echo signal of the water tank solution can be accurately obtained, the system's response time for sending and receiving data can be completed within 16 milliseconds, and the inverted NMR T2 spectrum is basically consistent with the laboratory measurement.

Abstract：Sonar is a highly prevalent underwater detection technology. In multi-channel sonar echo signal acquisition, the accuracy of individual channels, phase and amplitude consistency among channels, and crosstalk can all compromise the final detection outcomes, leading to significant discrepancies between the calculated results and the actual targets, thereby severely impacting subsequent decision-making. To address these issues, this paper employs Direct Digital Synthesis (DDS) technology to generate a common-source clock, ensuring strict synchronization across all acquisition channels and reducing displacement errors of moving targets caused by timing mismatches. Furthermore, to mitigate transmission delays associated with large volumes of raw acquired data, high-speed serial protocols such as Serial RapidIO (SRIO) are utilized to enhance transmission bandwidth, achieving real-time detection capabilities. The feasibility of the proposed approach is ultimately validated through practical engineering design.

Abstract：To enhance the accuracy and real-time performance of frequency offset estimation in wireless communication systems, this paper proposes a frequency offset estimation algorithm that integrates autocorrelation and pilot-assisted methods, and implements and optimizes the algorithm on the Xilinx Zynq-7000 SoC platform. Firstly, based on the frequency offset mathematical model and AWGN channel analysis, a theoretical framework of the impact of frequency offset on the received signal is constructed, and the spectral drift characteristics are verified through MATLAB simulation. In terms of algorithm design, the autocorrelation method is used for coarse frequency offset estimation, while the pilot-assisted method is employed for fine adjustment. The two methods are fused to form a highly robust hybrid algorithm. Simulation results show that this method maintains good estimation accuracy even in low SNR environments, and its BER and MSE performance are superior to traditional single algorithms. Subsequently, the algorithm is modularly implemented on the Zynq FPGA, covering core modules such as autocorrelation calculation, pilot synchronization, and frequency offset compensation, through resource optimization strategies, the logic utilization and power consumption are effectively reduced. The comprehensive simulation and on-board verification results demonstrate that the proposed algorithm achieves high accuracy while keeping the system logic resource occupancy rate within 35%, meeting the comprehensive requirements of real-time performance, resources, and power consumption in practical communication scenarios. This research provides a feasible solution for the design of high-performance and deployable frequency offset estimation and compensation systems.

Abstract：This paper investigates a malicious traffic identification technique based on an improved self-organizing competitive neural network. It designs a data feature engineering module for network traffic, constructs an S_Kohonen neural network model for network malicious traffic identification, and explores key issues in parameter optimization. By introducing the integral area mean maximum, a fast iterative parameter selection method is proposed. Integrated utilization of network traffic identification feature processing modules, unsupervised and supervised learning neural network mathematical models, the algorithm validation for malicious traffic identification is conducted. The results demonstrate that the parameter selection scheme proposed in this paper enables the neural model to effectively accomplish the task of identifying malicious network traffic. Under the condition of achieving an identification accuracy of no less than 98% in critical network systems, the model can be quickly applied.

Abstract：In response to the series of problems of inaccurate broccoli identification caused by occlusion and complex light in farmland environments, a field broccoli identification model based on YOLOv11 is proposed. Firstly， this model introduces the Deformable Large Kernel Attention(DLKA) attention mechanism, integrates the attention mechanism into the C3K2 module, forming a new module C3K2_DLKA to replace the C3K2 module in the original backbone network. Secondly, the advantages of the unique inverted bottleneck structure of Mobile Inverted Bottleneck Convolution(MBConv ) are incorporated into the detection head of YOLOv11, and the detection head in the original network is replaced with the Detect_MBConv detection head. In the object detection task, it can not only identify the target more accurately, but also respond quickly with limited computing resources. Finally, the loss function is replaced with Shape-IoU to enhance the convergence and stability of the model. Experiments show that the improved model has improved accuracy by 4.8%, the recall by 3.8%, and the mAP value has increased by 1.9%, reflecting the effectiveness of the improved algorithm.

Abstract：Aiming at the shortcomings of traditional security protection for power monitoring systems in attack detection, traceability and unknown threat response, this study constructs a power monitoring decision-making system by integrating knowledge graph and large model technologies. Multi-source heterogeneous data is processed through standardized processes, entity recognition and relationship extraction are used to build a knowledge graph, and a special large model for network security is combined to achieve intelligent threat detection and analysis. The system has core functions such as threat detection and attack traceability, and can monitor in real time, locate accurately and provide operation and maintenance suggestions. Practices show that its attack detection accuracy, unknown attack recognition ability and traceability efficiency are better than traditional technologies, with an average vulnerability detection accuracy of 95.5%. It improves the system security and decision-making intelligence level, providing technical support for the digital transformation of the power industry.

Abstract：To address antenna impedance variations, a high-efficiency and load-insensitive power amplifier operating in the 8~12 GHz band is designed using 0.25 μm GaN HEMT technology. By employing an asymmetric impedance matching network structure, the design achieves an optimal balance among high efficiency, compact size and good load-insensitive characteristics. Under continuous-wave test conditions, the chip demonstrates a saturated output power of 42.1~43.3 dBm and power-added efficiency (PAE) of 46.8%~55.5% across the 8~12 GHz frequency band. When subjected to load mismatch (VSWR=3:1), the mean variation of output power within the frequency band is 2.2 dBm.

Abstract：To realize real-time integration of sample detection data and traceability processing of patient diagnostic information in the in vitro diagnostic instrument assembly line, a multi-channel RFID read/write system was developed for the transmission track and reagent tray detection of the in vitro diagnostic immunoassay module. The system comprises a five-channel RF antenna multiplexing interface and an micro control unit. The host computer communicates via serial communication and supports functionalities such as RF tag data read/write operations, adjustable transmission gain, key protection, audio-visual alarm, and remote serial online firmware upgrades. This design effectively addresses the limitations of traditional systems, which suffer from redundant detection mechanisms due to the need for separate read/write control units for each RF antenna and the challenges of RF communication within metal tracks. Furthermore, the system meets the communication requirements of two types of fixed-distance antennas used for internal transmission track and reagent tray detection in the in vitro diagnostic module. Experimental results indicate that after antenna optimization within the metal track, the impedance matching reaches 50.718 Ω, with a standing wave ratio of 1.11. By adjusting the transmit and receive power settings of the RF chip through software, the read distance for samples equipped with S50 tags reaches 26 mm, achieving a 100% read success rate, thereby satisfying the data read/write requirements of the in vitro diagnostic production line.

Abstract：Stepping frequency ground penetrating radar (GPR) has been widely developed and used in shallow-surface high-resolution detection due to its ultra-wideband advantage. However, the general stepping-frequency GPR hardware implementation often adopts a multi-board coordination scheme, and is mostly designed as a single transmitter and single receiver mode, which cannot simultaneously meet the modern detection requirements such as high detection efficiency and highly integrated system. In response to the above needs, this paper designs a multi-transceiver stepping-frequency ground penetrating radar array system based on RFSoC, and realizes a set of stepping-frequency GPR system that highlights the characteristics of high-resolution and high-detection efficiency, and has the function of transceiver control of antenna arrays. The hardware platform of this system is RFSoC-47DR digital-analogue mixed-signal processing board, the array antenna of this system consists of 23 ultra-wideband antennas of the same type designed as a 11-transmitter and 12-receiver structure, and through the design of the clock link, transceiver logic control link, RF switch control link, high-speed storage of data link and other modules, the system completes the system for all the antennas to transmit and receive. This system is able to image 22 B-sweeps in one test during the real test, which greatly improves the efficiency of detection. Through the test verification of several real-world scenarios such as stone pavement, asphalt tarmac, etc., this system design is able to carry out effective detection and acquire information such as targets and structures under the surface of common roads.

Abstract：To improve the imaging quality of spaceborne SAR, this study focuses on addressing the velocity spatial variation characteristics among different range cells during imaging processing. It investigates high-precision equivalent velocity calculation methods to achieve accurate focusing of targets in each range cell and optimizes the equivalent velocity calculation workflow. Firstly, based on the range equation, Doppler equation, and Earth ellipsoid model equation of spaceborne SAR, geometric positioning of targets at different range gate sampling positions at the imaging center time is performed. The traditional hyperbolic model is adopted to calculate the slant range history of targets positioned by different range gates, and a first-level rough estimation of equivalent velocity is completed according to imaging time and slant range variations. Subsequently, secondary estimation research is conducted on two types of equivalent velocity estimation strategies: one based on image statistical features (maximum contrast method) and the other on echo signal features (fractional Fourier transform method). After completing the accurate estimation of equivalent velocity, the coupling relationship between processing efficiency and estimation accuracy in different calculation methods is optimized, and a full-process framework integrating secondary rough estimation and precise estimation of equivalent velocity is finally constructed and completed. Finally, by comparing the processing results of point target echo simulation data in strip modes generated based on spaceborne imaging parameters, quantitative analysis is carried out focusing on key indicators such as point target resolution, integrated sidelobe ratio, peak sidelobe ratio, and algorithm time consumption, thereby completin

Abstract：Aiming at the scheduling problem of multifunction phased array radar, a self-adaptive scheduling algorithm based on the overall planning concept is proposed. This algorithm divides the entire scanning area into sectors and dynamically calculates the scheduling strategy to be used based on the number of targets in each sector and the different tasks. When the resource load of dense target sectors is insufficient, the scheduling tasks of subsequent sectors are recalculated, and the excess resources of idle sectors are automatically allocated, greatly improving time utilization and the success rate of dense target scheduling. The simulation data analysis results show that the algorithm proposed in this paper is effective.

Abstract：To address the low development efficiency of FPGA signal processing in primary radar systems, a graphical design method based on System Generator is proposed. This approach constructs visual models for algorithms such as digital beamforming, down-conversion, and pulse compression, enabling full-process development from algorithm design to hardware deployment. Tests on a Xilinx Virtex-7-based radar processing board demonstrate stable output of correct pulse-compressed signals, enabling continuous tracking of drones within an 8 km range when combined with DSP. This method shortens development cycles, reduces resource usage, and optimizes routing. The study provides an efficient development paradigm for radar signal processing systems, suitable for rapid algorithm iteration and engineering applications.

Abstract：The aerospace electric propulsion power supply has strict requirements for high efficiency and reliability. The boost full-bridge LLC resonant converter has become the mainstream solution in this field due to its high efficiency, simple topology and high reliability. Under the condition of meeting the preset gain, further improving the energy conversion efficiency through parameter optimization is one of the core issues to be solved in the engineering design stage. Focusing on the efficiency optimization of the converter, this paper conducts a full-process study: deriving the critical mQ curve equation for realizing soft switching under the condition of a determined gain, and establishing mathematical loss models of each link; verifying the correctness of the models and optimizing parameters through PSpice simulation; building an engineering prototype based on the optimized parameters. The deviation between the measured results, simulation results, and calculation results is ≤ 5%, which confirms the accuracy of the models. The loss optimization method formed in this paper provides theoretical and verification support for the engineering design of this type of converter, and helps to improve the performance of the aerospace electric propulsion power supply.

Abstract：The key parameters of the tobacco redrying process are crucial to the final product quality. The traditional parameter control method, which relies on experience, suffers from poor stability and response hysteresis, making it difficult to meet the demands of modern industrial production. To address this, this paper constructs an intelligent prediction model for tobacco redrying process parameters based on the XGBoost machine learning algorithm. Firstly, the redrying process flow and key quality indicators are analyzed to establish an efficient data acquisition and feature engineering method. Subsequently, the XGBoost model parameters are optimized and combined with a feature selection strategy to improve prediction accuracy and model robustness. The experimental results show that this method has high accuracy and stability in process parameter prediction. Compared with traditional methods and other machine learning models, it significantly enhances the intelligent level of the redrying process. This research provides effective support for the digital and intelligent optimization of the tobacco processing procedure, contributing to improving product quality and production efficiency.

Abstract：The method for image captioning based on frozen large language models breaks through the limitations of traditional models in utilizing external knowledge, but it tends to overly rely on textual priors, leading to insufficient use of visual features and hallucination issues in descriptions. To address this, this paper proposes a semantic alignment-adaptive image captioning method, which constructs a circulation path between visual and textual data elements through a visual compression module and a semantic routing module, achieving precise alignment and efficient interaction of cross-modal semantics. Experimental results on benchmark datasets such as MSCOCO, Flickr30k, and NoCaps show that this method can effectively promote the value transfer of multimodal data elements while maintaining a low number of trainable parameters, reaching current state-of-the-art performance and providing a reliable solution for industrial applications.