Abstract：Aiming at the problems such as the low generalization ability of traditional methods in road defect detection, which are vulnerable to environmental interference, and the high power consumption and low speed when deploying deep learning models on computing platforms, an acceleration and deployment strategy for the semantic segmentation model based on a low-power FPGA (Field-Programmable Gate Array) platform is proposed. Firstly, a multi-source dataset containing road cracks and potholes is constructed, and data augmentation techniques are used to balance the sample distribution. Secondly, channel pruning is carried out separately for the feature extraction network and the upsampling network of the UNet model. Combined with the quantization technique, the model weights are compressed from FP32 (32-bit floating-point) to INT8 (8-bit integer), further reducing the computational load. Finally, the Vitis AI toolchain is utilized to complete the model quantization and compilation, and the model is deployed to the FPGA platform to fully exert its parallel computing capability. The experimental results show that, on the premise of ensuring that the loss of the mean intersection over union (MIoU) is less than 5%, the inference speed of the optimized model reaches 17 ms. The number of model parameters and the computational load are significantly reduced, and the power consumption is remarkably decreased. This method achieves efficient and low-power road defect detection at the edge side, providing a feasible solution for the automated maintenance evaluation of asphalt roads.

Abstract：With the increasing application of visible light imaging and infrared imaging in various scenarios, integrated design concept is adopted to develop the compact dual-mode imaging and processing platform. The platform employs RK3588 and ZYNQ as main processor and coprocessor, enabling parallel real-time processing of 1080P visible light images and 1 280×1 024 resolution uncooled infrared images. Under the control of software algorithms, it can achieve real-time target tracking and recognition. Through field tests and high-low temperature tests of physical prototypes, the platform demonstrates stable and reliable performance with a compact structure, meeting the requirements of multiple application scenarios such as UMV, USV or UAV .

Abstract：With the rapid development of technologies such as intelligent driving and robots, conventional 2D detection algorithms cannot meet the requirements of environmental perception in these scenarios, and 3D target detection is required to obtain accurate environmental information. However, the current mainstream 3D target detection models based on multi-source data fusion rely on high-computing and high-power platforms, and are difficult to implement on low-performance embedded platforms. In response to these problems, a method for implementing multi-source fusion 3D target detection on a low-power FPGA platform is proposed. By fusing the LiDAR point cloud and camera image data, the lack of point cloud feature information is compensated to achieve higher accuracy and detection stability. At the same time, combined with the characteristics of the FPGA platform, the fused features are screened and processed, and the model is compressed in combination with a quantization strategy. After experiments, the fusion method significantly improves the accuracy of small objects, and the quantized model runs successfully on the end-side FPGA platform with an average 3D accuracy loss of less than 3%.

Abstract： In order to meet the diversified requirements of airborne display, this paper proposes a video conversion and overlay technology based on FPGA. The technology takes FPGA as the core, and is equipped with peripheral circuits such as decoding circuit and signal conversion circuit, which can realize the conversion of XGA and PAL analog video signals to RGB digital video signals, and overlay display with digital image signals. It has strong versatility and flexibility. The experimental results show that the video conversion and overlay technology can meet the requirements of stable, reliable and highly integrated picture display of airborne displays, and has high application value.

Abstract：To address the increasingly high time cost of traditional automatic routing algorithms in the design process of printed circuit boards (PCB), a routing generation method based on generative adversarial networks is proposed. This method transforms the PCB routing problem into an image generation problem. By discarding the noise input of the generator, introducing multi-level residual connections, and adding a custom loss function, the method can learn the physical features of the routing problem, and the accuracy of the prediction results is improved. The network is trained and tested on a self-built PCB routing dataset. Experimental results show that compared to the traditional A* search algorithm, the proposed generative method reduces routing time by 50%. This method provides a new solution for PCB routing problems, helping to reduce the complexity and time cost of routing tasks.

Abstract：Body size parameters are critical indicators in pig breeding. A non-contact pig body size measurement method using dual view point cloud registration was proposed to address the isssues of single parameter, complex equipment, and limited processing of large-scale point cloud data exsited in current pig body size measurement methods. Firstly, we constructed a data acquisition system with two Kinect DK depth cameras to collect bilateral point cloud data of pig and then we preprocessed the data using an improved Local Outlier Probability (LoOP) filtering algorithm and a multi-level feature extraction method for point cloud simplification. Secondly, the dual view point cloud registration was completed by combining coarse registration and fine registration algrithms. Finally, We integrated normal vector point cloud with the Alpha Shapes algorithm to extract pig contour features of pig, achieving non-contact measurement of multiple body size parameters. The experimental results showed that the average relative errors of pig body length, body height, hip height, body width, abdominal width, hip width, chest circumference, and abdominal circumference were 1.28%, 0.88%, 1.97%, 2.71%, 2.83%, 3.71%, 2.03%, and 2.17%, respectively. The overall average relative error and absolute error were 2.20% and 1.04 cm, respectively. The method in this study provides an accurate, non-invasive solution for multi-parameter measurement of pig, offering technical support for efficient breeding selection in pig farming.

Abstract：Current brain tumor segmentation models typically concatenate multi-modal MRI images along the channel dimension, failing to fully exploit their multi-modal information. To address this limitation, this paper proposes MGF-UX Net, a model based on 3D UX-Net that integrates a multi-modal grouped feature processing module. The model consists of two key components: a modality feature extraction module and a modality feature fusion module. The feature extraction module employs specialized networks tailored to the characteristics of four MRI modalities to maximize the retention of effective features. The feature fusion module, guided by prior knowledge, groups the four modalities into two sets and applies a Modality-Grouped Cross-Attention (MGCA) mechanism for deep feature fusion. Experimental results on the BraTS2021 dataset demonstrate that MGF-UX Net achieves an average DSC of 91.01% and an HD95 of 4.975 mm, outperforming existing models in brain tumor segmentation tasks.

Abstract：With the continuous development of the network and the continuous improvement of data bandwidth, higher requirements have been put forward for key indicators such as switching capacity, forwarding delay, and memory utilization in network devices. This design adopts a switching architecture based on crossbar multi-port cache sharing to achieve the storage, scheduling, sorting, and forwarding of messages. This design uses a shared cache architecture and defines data packets with descriptors, achieving shared queue cache for multiple queues under multiple ports. Each port supports 8 priority queues, realizing orderly scheduling and reducing the blocking problem of multiple queues. In the cache management of multiple SRAMs, a paged linked list method is adopted to achieve dynamic shared cache, improving data access efficiency and reducing forwarding delay. This design has completed FPGA prototype verification on the XCZU3EG verification platform of Xilinx Company, achieving functional verification and performance testing of the multi-port structure. The average port rate reaches 6.08 Gb/s, and the switching delay is controllable.

Abstract： This paper presents a radio frequency front-end integrated circuit design based on a standard 0.18 μm CMOS process, specifically tailored for dual-mode navigation systems supporting GPS L1 and Beidou B1. The design emphasizes optimized key performance metrics for both the low-noise amplifier (LNA) and down-conversion mixer. For the LNA architecture, this study innovatively employs an inductive source degeneration technique, effectively reducing the noise figure by connecting external capacitors in parallel with gate source capacitors. Regarding the down-conversion mixer module, a structurally advantageous passive single-balanced topology is adopted, significantly improving linearity metrics and suppressing 1/f noise interference while maintaining conversion gain. The test results show that the scheme achieves a 60 dB link gain and a 2 dB noise figure at a voltage of 1.8 V, with an overall power consumption of only 22 mW. In the system integration test, when the input power is -100 dBm, the output intermediate frequency power reaches 3.15 dBm.

Abstract：With the development of the third generation of wide bandgap semiconductor materials, new power switches such as GaN and SiC have been widely used, and the fast switching speed of the new power tubes makes the transient voltage dV/dt of the driver switch up to 200 V/ns, resulting in the false opening of the power tube. In order to improve the Common Mode Transient Immunity (CMTI) of isolated drive, a simple high-CMTI codec circuit is designed, which adopts OOK (On-Off Krying) modulation structure, the coding circuit uses a negative transconductance oscillator structure to encode the high potential of the signal into a differential oscillation signal, and the decoding circuit uses a common-mode point lifting circuit to stabilize the common-mode point of the differential signal at VDD/2. The input signal can be restored only by means of a comparator. The simulation results show that the speed of the input signal of the codec circuit can reach 143 Mb/s, the frequency of the differential oscillation signal is 644 MHz, and the typical transmission delay is 1.2 ns, which is used in an isolated driver chip, and the measured CMTI can reach up to 400 V/ns.

Abstract： The electric drive system of new energy vehicles usually uses a resolver to obtain the position of the motor rotor, but its circuit is complex and the chips are expensive, resulting in high costs. The traditional second-order phase-locked loop has steady-state errors when accelerating and tracking the rotor. In actual working conditions, the loss of one phase of the input sine and cosine often leads to resolver faults in the soft ware decoding. For this, an angle observer based on the third-order Angle observer is proposed to improve the position decoding accuracy. Meanwhile, the phase-locked loop algorithm was adopted to track the remaining signal during the decoding fault and recover the lost signal to complete the decoding after the fault. Finally, the feasibility and effectiveness of the algorithm were proved through MATLAB simulation. The DSADC module of Infineon TC387 was adopted for actual verification of the decoding effect. The results show that the system achieves stable tracking of acceleration and can maintain a good decoding effect after the absence of a certain signal.

Abstract：This paper proposes a high-precision real-time attitude measurement system based on the TMS320F28375S Digital Signal Processor (DSP). By fusing six-axis inertial data from a triaxial fluxgate sensor and Micro-Electro-Mechanical System (MEMS) accelerometer with an adaptive Kalman filtering algorithm, the system achieves high-precision measurement with ±0.1° pitch angle and ±0.5° yaw angle accuracy. The system employs a Δ-Σ analog-to-digital converter (ADS131M04) to achieve 128 kSPS synchronous sampling rate, while utilizing the DSP's built-in Trigonometric Math Unit (TMU) to reduce the attitude calculation cycle to 0.8 ms. Experimental results demonstrate that under simulated dynamic interference conditions, the system attains a 1 kHz attitude update rate with power consumption below 1.5 W, significantly outperforming conventional solutions. The system meets the real-time control requirements for mobile platforms such as Unmanned Aerial Vehicles (UAVs) and Automated Guided Vehicles (AGVs).

Abstract：With the rapid development of space information technology, the demand for multicast communication has gradually extended to satellite networks. However, the traditional multicast key management scheme based on ground devices faces performance bottlenecks and security risks in the on-board environment. In view of the limited computing resources of satellite nodes, this research proposes a scheme of unsearchable balance plus logical key hierarchy (UBP-LKH) tree based on logical key hierarchy (LKH) tree structure adjustment. By designing a double threshold trigger mechanism, the lower limit reconstruction (tree compression) and upper limit reconstruction (tree expansion) are implemented by using the idle time of computing resources of the multicast key management system (MKMS), and the key tree hierarchy is dynamically optimized to reduce the encryption delay. Experiments show that the encryption efficiency of this scheme is improved by up to 20% compared with the classical LKH tree in the small and medium-sized multicast scale (5~127 nodes), and the optimization range is negatively correlated with the number of members with the tree height. This research provides a scalable multicast key management mode for the on-board environment, which has both resource adaptability and practical application value.

Abstract：To fully utilize edge computing power and achieve real-time and efficient garbage recognition and detection, a lightweight garbage detection model is proposed. The model uses ShuffleNetv2 as the feature extraction network, reducing computational complexity through channel rearrangement and depthwise separable convolution while preserving important information. And the Ghost network is adopted to improve the C3 module, reduce computational burden and improve fusion efficiency. To further reduce the number of model parameters, the computational efficiency is optimized by reducing the depth of the model. A garbage classification mechanism based on cumulative voting has been designed. When the number of garbage type recognitions reaches a set threshold, classification is carried out, and the results are transmitted through a serial port and linked with the control system. The experimental results show that the improved model reduces memory usage by 71.4%, accuracy loss by only 0.16%, accelerates inference speed, significantly reduces energy consumption, and ensures efficient transmission of classification results.

Abstract：Network probing and scanning is an essential method for discovering network assets, with HTTP/HTTPS protocols representing a significant proportion of the discovery results and serving as a key source for identifying Internet assets. As the network environment becomes increasingly complex, the variety and volume of assets utilizing the HTTP/HTTPS protocol have grown rapidly, which poses challenges for traditional network asset identification methods based on fingerprinting rules. These conventional approaches suffer from low recognition efficiency and poor adaptability, making them inadequate for identifying HTTP/HTTPS protocol assets. Therefore, this paper proposes a novel method for discovering HTTP/HTTPS protocol assets. The approach processes HTTP/HTTPS response data through an automated rule generator, performs pre-filtering of the raw data based on term frequency statistics and similarity information, and applies a text encoding model to encode the HTTP/HTTPS response body and fuse the features. By integrating an unsupervised clustering algorithm, this method enables the discovery of HTTP/HTTPS protocol assets. Experimental results show that the proposed method significantly improves the efficiency of HTTP/HTTPS protocol asset discovery, accelerates asset labeling, and enables the discovery of unknown assets without prior knowledge.

Abstract：With the development of Internet of Things technology, the performance of the single-chip microcontroller has been upgraded and the functions have become rich, and it has become feasible to use the single-chip microcontroller to create a web server and use the browser as a client to access. Drawing on its ideas, a mobile phone instrument design method that does not require software installation based on embedded Web server + browser architecture was proposed. First, the JavaScript language is used to design the commonly used mobile phone instrument elements as a 50KB library that can be embedded in the storage system of single-chip microcontroller. Then, on the basis of it, the C-style mobile phone instrument HTML web page generation function is formed. Finally, through the single-chip microcomputer web server, the encapsulated C-style mobile phone instrument HTML web page generation function is converted into the HTML web page supported by the mobile phone browser for display and user operation. It realizes the function of installing the instrument software on the lower computer, and accessing the instrument interface with zero installation and zero configuration on the client.

Abstract：In order to meet the high-precision design requirements of broadband microwave wireless communication systems, a five-bit digital controlled attenuator with low additional phase shift in the 12~30 GHz range was designed using GaAs 0.15 μm PHEMT technology in this paper. This attenuator can achieve an attenuation range of 0 dB to 15.5 dB with a minimum step of 0.5 dB by cascading five different attenuation bits. Among them, the 0.5 dB attenuation bit adopts a simplified T-type attenuation structure, while the 1 dB, 2 dB, 4 dB and 8 dB attenuation bits adopt a Lange coupler-based reflective structure. In the reflective attenuator structure, the stacked switch transistor technology and capacitor compensation technology are adopted to effectively reduce the additional phase shift within the ultra-wide frequency band and improve the attenuation percosopn. Electromagnetic simulation shows that within the 12~30GHz frequency band, the insertion loss of the reference state is less than 3.8 dB, the root mean square (RMS) error of the 32-state attenuation of the attenuator is less than 0.5 dB, its typical value is only 0.3 dB, the additional phase shift of the 32-state attenuator is less than ±3°, and the input and output return losses at the ports are both greater than 12 dB. The circuit size is only 2.95 mm×1.2 mm.

Abstract：Based on unmatched plastic-package GaAs die and SMT (Surface Mounted Technology) hybrid integration process, a low-cost Ku band 2-stage LNA (Low Noise Amplifier) was realized on single-layer double-side high-frequency PCB (Printed Circuit Board). A global-optimization method was adopted in the amplifier design, involving iterative co-optimization across multiple design dimensions such as waveguide-probe transition structures, optimal noise and conjugate matching networks, bias circuitry, and stability enhancement modules. According to the testing data, the LNA works from 10.7 GHz to 12.7 GHz, noise figure is less than 0.82 dB (including waveguide-coaxial-microstrip transition), small-signal gain is above 17 dB, fluctuation of gain in band is under +/-0.45 dB, VSWR for input/output port is less than 1.7. The area of LNA is 24 mm×17 mm, and cost of production is less than 60 CNY. The LNA is suitable for the radio frequency front end in Ku-band satellite communication ground equipment.

Abstract：With the continuous advancement of airport expansion projects, the issue of interference from obstacles around the airport with the electromagnetic environment of the Instrument Landing System (ILS) has become increasingly prominent. Among these, the reflected and diffracted signals caused by obstacles have particularly significant interference with the Localizer (LOC) radiation field, severely affecting the navigation stability and precision of the LOC system. To address this problem, this paper systematically analyzes the impact mechanisms of reflected and diffracted signals from obstacles on the LOC radiation field based on the ray-tracing algorithm. By establishing a dual-frequency 12-element LOC array model and combining it with typical scenarios of obstacles near airport runways, a signal propagation model is constructed, with a focus on studying the interference characteristics of reflected and diffracted signals on the Difference in Depth of Modulation (DDM). The research results indicate that in engineering scenarios where dedicated obstacles must be constructed at specific locations, the signal components causing significant interference to the LOC radiation field can be effectively suppressed by optimizing the spatial geometric relationship between the obstacle's reflective surface and the runway. This research provides a new technical approach to solving the multipath interference problem in the LOC radiation field of specific areas in practical engineering applications.

Abstract：Radar target recognition is a critical technology for environmental situational awareness. With the rapid development of artificial intelligence, deep learning has become a mainstream research approach in radar target perception. This paper first introduces a signal enhancement method based on an encoder-decoder network, which integrates the physical model of radar targets to achieve accurate feature preservation and signal recovery for noisy signals. Furthermore, we propose a radar target recognition method based on a Self-attention 1d Convolutional Neural Network (S-1dCNN). This method utilizes a one-dimensional convolutional neural network to learn features from radar target echo signals, effectively extracting structural information of target support regions in the range dimension. By incorporating a self-attention mechanism, the proposed approach focuses on features related to target structural information, enhancing the representational capability of learned features and thereby improving the model's recognition performance. Experimental results on measured echo data from five categories of civil aircraft targets demonstrate that the signal enhancement model can denoise signals while preserving target structural characteristics. When the enhanced signals are input into the S-1dCNN model, high recognition accuracy is achieved. Additionally, a MATLAB-based graphical user interface (GUI) is designed for radar target recognition, encompassing modules such as data preprocessing and recognition algorithms to enable multifunctional interactive operations. This software translates abstract algorithms into visual workflows, significantly enhancing the effectiveness of radar target recognition in educational contexts.

Abstract：Radar systems are confronted with diverse and dynamic jamming signals. Traditional methods for jamming signal recognition exhibit significant performance limitations in complex scenarios. Therefore, this paper proposes a recognition method for radar active jamming signals based on multi-domain features. Firstly, multi-dimensional features of radar active jamming signals are designed to serve as the basis for signal differentiation. Subsequently, utilizing the advantages of principal component analysis in reducing dimensional complexity, the most representative and discriminative feature subset is extracted. Finally, by integrating the Random Forest model, complex relationships within the feature data are automatically mined to achieve the classification of radar jamming signals. The experiment indicates that the proposed method achieves high recognition accuracy, making it suitable for practical complex scenarios in radar active jamming signal recognition.