Abstract:Convolutional neural networks (CNN) have gained widespread application in the field of deep learning due to their exceptional visual feature extraction capabilities. However, with increasing model complexity and the stringent demands of edge computing scenarios for computational performance and energy efficiency, hardware acceleration faces significant challenges. In specific application scenarios requiring low-batch inference, deterministic low latency, and difficulty in updating hardware, general-purpose central processing units (CPU) and graphics processing units (GPU) often fall short of the requirements. In this context, CPU-FPGA heterogeneous computing platforms based on field programmable gate arrays (FPGA) offer an effective pathway for implementing low-batch, low-latency CNN inference tasks at the edge. Their advantages lie in customized computational dataflow, reconfigurability, and deterministic low latency. This paper systematically explores CNN compression and computational acceleration methods for deployment on CPU-FPGA platforms from a hardware-software co-design perspective. First, we introduce network compression techniques for reducing model complexity. Second, we present hierarchical hardware optimization strategies for computational units, computational arrays, and data memory access. Third, we describe how an agile development flow based on high-level synthesis (HLS) enables rapid iteration and verification of accelerator systems. Finally, from the three dimensions of complex chip front-end verification, deterministic low-latency computing, and agile algorithm iteration, we explore the technological positioning, development prospects, and challenges of FPGAs in the new era of Artificial Intelligence.

Abstract:Aiming at the problems of low model prediction accuracy caused by the temporality, strong nonlinearity, and high dimensions of water quality data, a water quality prediction model based on the IGJO-TCN-BiGRU-SA has been proposed. Firstly, missing data are repaired through the CSI algorithm, while abnormal data are removed through Boxplot and FFT algorithms to enhance data completeness. The GRA algorithm is used to analyze the correlation of water quality indicators and obtain strongly correlated water quality indicators. Secondly, the BiGRU model is employed as the baseline model to extract latent change features from the data. The TCN algorithm is employed for local feature extraction from the data, enhancing the model’s capability to extract features at different time scales. The SA mechanism is used to adaptively update the feature weights of the BiGRU model, effectively extracting key features from global contextual information. Finally, the IGJO algorithm is employed to optimize the L2 of TCN, the learning rate and neuron number of BiGRU, and the key of the SA mechanism, thereby improving the accuracy of water quality prediction. The historical water quality data from the Huai River Basin were used as samples for experiments. Compared to classic algorithms such as LSTM, GRU, BiGRU and TCN-BiGRU-SA, the IGJO-TCN-BiGRU-SA prediction algorithm proposed in this research achieved reductions of 19.64% to 68.29% in MAE, 25.47% to 75.19% in RMSE, 28.71% to 73.36% in MAPE, and 0.19% to 3.61% in R2. Experimental results indicate that the water quality prediction algorithm based on IGJO-TCN-BiGRU-SA significantly enhances the model’s predictive accuracy.

Abstract:To overcome the issues of energy attenuation in fault features and modal overlap that traditional modal decomposition methods often encounter in high-noise environments, sparse reconstruction theory has been introduced into the field of bearing fault diagnosis. These methods achieve precise approximation of weak fault impacts by identifying the atomic combinations that best match the signal structure within an overcomplete dictionary, thereby mechanistically ensuring the integrity of feature energy. The Bayesian orthogonal matching pursuit (BOMP) algorithm demonstrates advantages in reconstructing and identifying weak features. However, existing methods typically eliminate atoms with contributions below a preset threshold during iteration, discarding the faint yet crucial fault feature energy contained within these atoms. This compromises signal reconstruction accuracy and noise reduction effectiveness. To address this issue, this paper proposes an improved Bayesian orthogonal matching pursuit (IBOMP) model. Its core enhancement lies in optimizing the decision threshold strategy for support set atom selection, aiming to maximize the retention of faint fault feature energy contained within low-contribution atoms. Comparative experiments against classical sparse reconstruction algorithms—including orthogonal matching pursuit (OMP) and Bayesian orthogonal matching pursuit (BOMP)—demonstrate that the proposed IBOMP method more effectively suppresses noise interference and enhances fault feature signals. Validation using the bearing fault dataset from the intelligent diagnosis and expert system laboratory at Nanjing university of aeronautics and astronautics demonstrates that compared to OMP and BOMP algorithms, the proposed IBOMP significantly enhances the frequency energy contribution of bearing rolling element and outer ring fault features—by 22.71%, 22.73%, 46.22%, and 46.52%, respectively.

Abstract:Pantograph-catenary pantograph-catenary arc faults pose a serious threat to the safe operation of trains, and the accurate identification of such faults is crucial for ensuring the reliability of railway power supply systems. Aiming at the problems that traditional methods have low recognition accuracy under complex working conditions and are difficult to balance the extraction of temporal and spatial features, this paper designs a pantograph-catenary arc fault detection method based on the Markov transition field （MTF） and an improved dream optimization algorithm. First, the original current signal is denoised by robust empirical mode decomposition. The signal-to-noise ratio of the denoised signals under four working conditions is all above 36 dB, and the correlation coefficient is greater than 0.985, which preserves the fault transient characteristics to the greatest extent. Then, the one-dimensional time-series signal is mapped into a two-dimensional feature matrix using MTF, and a dual-branch architecture combining DarkNet19 and Gated Recurrent Unit is constructed to extract deep visual features and temporal dynamic features respectively, realizing multi-modal information complementarity. Furthermore, the improved dream optimization algorithm is introduced to adaptively optimize key parameters such as learning rate and the number of GRU neurons. Meanwhile, the multi-head self-attention mechanism is optimized, and redundant information is suppressed by L1 sparse regularization and an adaptive dynamic scaling factor for redundancy. In addition, transfer learning is integrated to break through the limitation of scarce fault samples. Experimental results show that the proposed model achieves a fault detection accuracy of 99.58% under four different working conditions, with precision and recall both exceeding 99.5%.Compared with six comparative models including SVM, 1D-CNN and ResNet, the accuracy is improved by 3.75%~6.67%.In the anti-noise interference experiment, the detection accuracy of the denoised signal is 6.71% higher than that of the noisy signal. It exhibits stronger robustness and provides an effective technical solution for the accurate diagnosis of pantograph-catenary arc faults.

Abstract:Aiming at the difficulty of existing extended state observers in balancing the suppression capabilities for both periodic and aperiodic disturbances, particularly the insufficient observation and compensation accuracy for specific frequency harmonics caused by nonlinearities in voltage source inverters, a current harmonic suppression method based on a quasi-resonant extended state observer (QRESO) was proposed. By embedding a quasi-resonant controller into the extended state observer (ESO) architecture, a QRESO with both low-frequency disturbance observation and high-frequency harmonic suppression capabilities was constructed, and disturbance feedforward compensation was implemented within a model-free predictive current control framework. The proposed method can accurately extract 6th-order harmonic disturbances through the synergistic action of low-frequency integral and high-frequency resonant modules while maintaining system stability. Experimental results demonstrate that compared with traditional extended state observers, under dead time conditions of 1, 2 and 6 μs, the QRESO reduces the 6th-order harmonic of q-axis current by 95%, 97% and 97% respectively, while decreasing the 5th and 7th harmonics of phase current to below 1%, it significantly improves the current waveform quality and total harmonic distortion, verifying the superior performance of this method in improving current quality.

Abstract:With the advancement of industrial electrification and automation, electric motors, as the core components of power systems, directly influence the overall performance and reliability of the system. Traditional PI controllers often fail to meet high-dynamic control demands due to their slow response speed and poor disturbance rejection capability. To address this issue, this paper proposes a speed control method for brushless DC motors that integrates an enhanced Harris hawks optimization (EHHO) algorithm with a PI controller. The proposed method incorporates opposition-based learning, adaptive weights, nonlinear energy factors, and embedded smoothing techniques, significantly improving the optimization search capability and accuracy while enhancing the control system’s adaptability to dynamic operating conditions. Through Simulink simulations and hardware experiments based on a GD32F103RCT6 control platform, the performance of EHHO-PI is compared with that of traditional PI control in various complex conditions, including sudden acceleration, sudden load, and variable speed/load scenarios. The results demonstrate that the EHHO-PI controller outperforms the conventional methods in terms of speed response, disturbance rejection, and regulation accuracy, showing strong robustness and practical applicability. This control strategy provides an effective optimization solution for motor control technology with broad engineering application prospects.

Abstract:To address inter-dimensional coupling errors in six-axis force sensors, a decoupling method based on an improved grey wolf optimization (IGWO) algorithm combined with a back propagation (BP) neural network is proposed. Dynamic distribution parameter adjustment, an adaptive nonlinear control factor, and an improved position update strategy are introduced to optimize the initial weights and thresholds of the BP network, thereby enhancing convergence performance and global optimization capability. A decoupling model is established using static calibration data of the six-axis force sensor and is compared with the least squares method, standard BP, GA-BP, PSO-BP, and GWO-BP approaches. Experimental results indicate that the proposed IGWO-BP method effectively reduces inter-dimensional coupling errors. The maximum relative errors of type I and type II components are 0.267% and 0.127%, respectively, demonstrating higher decoupling accuracy than the comparative methods. To evaluate practical performance, the decoupling model is implemented on an STM32F103C8T6 microcontroller, and a real-time data acquisition system consisting of an AD7606 module and a sensor calibration platform is developed. Real-time tests at a sampling frequency of 500 Hz show that the maximum relative error between the decoupled six-axis force/torque outputs and the reference loads does not exceed ±1.6%. No evident delay or data loss is observed during operation, confirming the feasibility and effectiveness of the proposed method in terms of real-time decoupling accuracy and system stability.

Abstract:Aiming at the problems that ground penetrating radar signals are easily interfered by noise in complex underground environments, traditional denoising methods have modal aliasing and strong parameter dependence, a denoising algorithm combining the grey wolf algorithm (GWO), improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and improved multi-parameter controlled wavelet threshold is proposed. Firstly, the GWO algorithm is used to optimize the noise amplitude (Nstd) and the number of iterations (NE) in ICEEMDAN to achieve adaptive adjustment of parameters and suppress modal aliasing. Then, an improved multi-parameter controlled wavelet threshold function is designed to selectively filter and reconstruct the decomposed modal components to retain the effective signal characteristics. In the simulation experiment, for the weak target scene with a burial depth of 0.1 meters, under the condition of an input signal-to-noise ratio of 10 dB, the output signal-to-noise ratio of this method is improved to 19.308 2 dB, which is an average improvement of 93.26% over the traditional ICEEMDAN with custom parameters combined with soft and hard threshold denoising algorithm. The waveform correlation coefficient is 0.994 15 and the root mean square error is 0.108 3. In the actual measurement experiment, the processed image is clearer and the hyperbolic characteristics of the target are more obvious. This method can effectively improve the signal quality of ground penetrating radar, and has the characteristics of strong adaptability, anti-aliasing and high engineering applicability, providing reliable technical support for the identification of underground targets.

Abstract:Ash deposition on the heating surfaces of coal-fired power station boilers causes a significant decrease in energy efficiency. If it accumulates continuously, it may cause serious safety accidents. Therefore, accurate prediction of the energy efficiency of the boiler heating surface is an important way to solve this problem. To achieve the improvement of energy efficiency prediction accuracy, this paper constructs an energy efficiency prediction model for coal-fired power station boilers based on a neural basis expansion analysis (NBEATS) model. This model is optimized by the honey badger algorithm (HBA) that incorporates a two-step decomposition strategy. This method uses the cleanliness factor as an indicator of energy efficiency in the boiler heating surface, and applies wavelet threshold theory to denoise the original cleanliness factor data. The two-step decomposition strategy of robust empirical mode decomposition (REMD) combined with variational mode decomposition (VMD) is used to obtain multiple mode functions of the denoised data. For each decomposed mode function, the HBA algorithm is applied to tune the hyperparameters of the NBEATS model for independent prediction, and the predicted values of all mode functions are superimposed to generate the boiler energy efficiency prediction results. Experiments on the boiler dataset of a 300 MW coal-fired power station in Guizhou Province show that the proposed model outperforms the ablation model in multiple performance metrics, the root mean square error performance is improved by 5.43%~93.31%; the coefficient of determination increases to above 0.98; the minimum mean absolute error achieves 0.000 2.

Abstract:To address the problems of scarce fault samples, high risk in reproducing faults on real equipment, and high cost, this paper proposes a marine elevator modelling method and a fault injection method based on multi-physics causal modelling for a fault sample generation method. Different from the traditional measurement-bias fault injection method that only superimposes deviations at the level of measurable signals, the proposed method establishes a fault injection model based on the mechanical-electrical-hydraulic coupling characteristics of marine elevators, and by dynamically modifying key component-level physical parameters, realises the evolution of the elevator from component-level performance changes to system-level functional failure. Based on the marine elevator fault-injection method proposed in this paper, the method’s effectiveness in fault propagation across multiple components is verified through fault-physical-quantity characterisation behaviour analysis and comparative experiments on performance with measurement-bias injection. Meanwhile, the effectiveness of the proposed method as a verification benchmark for fault diagnosis is demonstrated through the consistency of the effects of multiple typical fault diagnosis algorithms. The proposed method can generate fault samples with strong physical causality, providing reliable fault data support for the training of intelligent operation and maintenance and fault diagnosis algorithms for marine elevators.

Abstract:In nanoscale semiconductor manufacturing processes, the operating voltage of chips has been reduced to as low as 0.8 V, while the power consumption of some chips exceeds 700 W and the peak current exceeds 2 000 A. These trends impose stringent requirements on the power distribution network (PDN) that supports chip operation, including wide frequency bandwidth, ultra-low impedance, and low power loss. The intrinsic parameters of decoupling capacitors within the PDN are critical to its optimal design. To obtain accurate capacitor parameter models for PDN power integrity simulation and analysis, this paper proposes a high-precision broadband method for extracting the intrinsic parameters of decoupling capacitors based on a vector network analyzer (VNA). Vector fitting and error correction techniques are introduced to address the zero-crossing issue in test trajectories caused by trace noise, and the intrinsic parameters of the decoupling capacitors are extracted. An experimental system is constructed, with fixtures designed and multi-stage calibration and fixture de-embedding implemented. The test frequency range is 500 Hz to 3 GHz, achieving a capacitance error of ≤5%, an equivalent series resistance (ESR) error of ≤3%, and an equivalent series inductance (ESL) error of ≤6%. The measured intrinsic parameters of the decoupling capacitors are consistent with the reference values, and the impedance curves derived from substituting these parameters into the equivalent model align with the measured results. This method enables efficient and reliable extraction of the intrinsic parameters of decoupling capacitors, providing robust data for the optimal design of PDN and holding significant engineering value for the development of power supply systems supporting computing networks and digital systems.

Abstract:The gears of wind turbines operate under multiple noise and high-speed conditions for a long time, and the vibration signals exhibit nonlinear characteristics, which makes it difficult for single scale filtering methods in morphological filtering to accurately and effectively extract fault features, a method for gear fault feature extraction using the adaptive multi-scale composite differentiation product morphological filter (AMCDPMF) is proposed. Firstly, three morphological differential operators are introduced to construct the muti-scale composite differentiation product morphological filter (MCDPMF). Secondly, an adaptive selection of the optimal structural element (SE) scale for multi-scale filtering signals is achieved using a composite fault feature index. Subsequently, by integrating the MCDPMF with the optimal SE scale, the AMCDPMF filter is formulated to enable adaptive filtering, thereby effectively extracting fault features from simulated and experimental signals of gear dynamic models. Finally, through comparative analysis with single-scale filters and three classical multiscale filters, The energy ratio (ER) and improvement in signal-to-noise ratio (ISNR) of AMCDPMF have increased by 5.15% and 3.5 dB, respectively. The ability to extract gear fault features is better, effectively improving the reliability of wind turbine operation and promoting the efficient utilization of renewable energy.

Abstract:Structured light 3D measurement technology is widely used in fields such as industrial inspection and medical imaging, but its measurement accuracy is often reduced due to the nonlinearity of the projector. In traditional phase-shifting methods, especially when combined with binary fringe projection techniques, a large number of fringe patterns are typically required to ensure measurement precision, which leads to low data acquisition efficiency. Focusing on how to improve measurement accuracy while reducing the number of projected patterns, this paper proposes a sixteen-step phase-shifting encoding method based on equally spaced binary fringes. The method generates binary fringe patterns with equally spaced distribution in advance, maps these patterns one-to-one to the corresponding sinusoidal intensity values within the same period, and then reconstructs the sinusoidal fringe through a weighted summation strategy. This innovative encoding approach enables each binary fringe to not only contribute information independently but also to be efficiently reused in the sixteen-step phase-shifting algorithm, thus significantly reducing the required number of fringe patterns and effectively improving 3D measurement efficiency. In the experimental section, the proposed method was quantitatively compared with the traditional sixteen-step phase-shifting method, the sixteen-step binary defocusing method, and the sixteen-step binary encoding method. The results show that, in terms of unwrapped phase accuracy, the proposed method achieved significant improvements, with the accuracy rates increasing by 71.42%, 87.17%, and 85.29% respectively compared to the other three methods, demonstrating its remarkable advantages in both enhancing measurement precision and reducing the number of projected patterns.

Abstract:To address issues such as communication topology symbolization, path directionality and external interference that cause a decline in speed coordination control accuracy in multi-motor systems such as unmanned aerial vehicle formations and industrial robotic arms, this paper proposes a speed coordination control strategy for multiple permanent magnet synchronous motors based on bipartite consensus. First, each single motor control system is considered an intelligent body, and the multi-permanent magnet synchronous motor speed cooperative control problem is transformed into a bipartite consensus problem containing a virtual navigator. The distributed control protocol is designed based on the principles of dividing by competitive relationships, directly participating in weight shares, and summation adjustment by in-degree weights. Second, a sliding mode speed loop controller based on an improved variable power reaching law is designed for the permanent magnet synchronous motors. Simultaneously, an ultra-local model-based design of model-free sliding-mode controllers for the current loops of the leader motor and each of the follower motors is used to improve the control accuracy and anti-interference capability of the multi-motor system. Finally, the proposed method is validated by constructing a simulation platform for five permanent magnet synchronous motors. Under no-load condition, it can realize high-precision speed tracking within a finite time, and the steady state tracking error is lower than 0.006%; under load disturbance, it can effectively inhibit both homogeneous and heterogeneous load disturbance conditions. When connectivity decreases due to communication topology switching, the overall convergence time and steady-state tracking error remain stable despite a slight increase in collaborative error. This method provides an effective solution for the high-precision and robust speed cooperative control of multiple permanent magnet synchronous motors under complex communication topologies.

Abstract:Pore pressure is a key parameter in oil and gas exploration and development. Accurate prediction is crucial for ensuring drilling safety and improving operational efficiency. Traditional models and shallow machine learning methods often lack effective integration of lithological spatial information, which limits their generalization performance. This paper proposes a deep graph attention network method (CLG) that incorporates lithology information for pore pressure prediction. The model takes multiple inputs including well depth, equivalent circulating density (ECD), mechanical drilling parameters (rate of penetration and weight on bit), natural gamma ray, porosity, and lithology encoding. Most data are obtained from the logging while drilling (LWD) system, and are processed by boxplot cleaning and normalization to improve quality. CNN extracts local features, LSTM captures temporal dependencies, and the graph attention mechanism integrates lithological structural information. Application in a Bohai exploration area shows that the CLG model achieves excellent performance across multiple wells (R2=0.997 9, MSE=0.003 5, RMSE=0.059 6, MAE=0.048 1), improving accuracy by about 15% and 7.4% compared to the Eaton method and various shallow models, respectively, with enhanced generalization ability. This method effectively advances pore pressure modeling with lithology awareness, providing reliable support for oil and gas drilling under complex geological conditions.

Abstract:Advance detection of geological structures is crucial for ensuring efficient and safe coal mine production. To address the limitations of traditional geophysical techniques in meeting the precision requirements for advance detection in deep roadways, a new method based on spectral acoustic was designed. First, the resonant characteristics of roof strata during roadway excavation were theoretically analyzed, and the basic principle of advanced detection of geological structures by inversion of stress changes of coal and rock mass with elastic wave attributes was expounded. Subsequently, a continuous monitoring and real-time early-warning system platform was established, with system hardware including geophones and monitoring substations developed, and interactive control software created. Finally, field validation was conducted at the Ji15-15080 machine roadway in pingmei eighth coal mine, a representative deep mining site. Results indicate that: High-value zones of the geological structure monitoring index coincide with exposed geological structures; The prediction coefficient shows a rapid increase exceeding the critical value within a certain distance ahead of structure exposure; Relative stress coefficients exhibit elevated stresses unilaterally, bilaterally, or entirely at faults, consistent with geological structure evolution. According to the linear distance in the driving direction, the advance warning distance is 7.2~37.7 m, achieving a reliability of 92.85% for advance detection. The research results provide an efficient and reliable new technology for advanced detection of geological structure in deep roadway.

Abstract:This paper constructs a wavelet-gated temporal convolutional network model. First, the input electromyographic signals are subjected to multi-level discrete wavelet decomposition through a wavelet convolution module, and the components of each level are respectively subjected to one-dimensional convolution. Then, the detailed coefficients and approximate coefficients after convolution are reconstructed via discrete inverse wavelet transform. This process of multi-level decomposition, convolution, and step-by-step reconstruction enables the model to adaptively focus on key time-frequency features. The reconstructed signals are then input into a temporal convolutional network integrated with a gating unit. The proposed network structure achieves an accuracy of 81.85% for 52-class gesture classification on the Ninapro DB1 dataset, which is 4.9% higher than that of traditional temporal convolutional networks. Compared with recent mainstream deep models in this field (such as MSHilbNet, GengNet, etc.), this method achieves a relative accuracy improvement of 4.0%~7.8% while maintaining a smaller number of model parameters.

Abstract:This paper addresses key issues such as “insufficient quantification of ball stability in metal sphere systems、lack of systematic parameter constraints in ball search processes、unclear mechanisms of range-bin crossing and deviation amplification patterns.” Proposes an absolute far-field calibration scheme for weather radar that combines an equipped UAV, a laser-range-finding camera and a standard metal sphere. First, a stability metric constrained by a maximum horizontal displacement (S≥ 97%) is defined to determine the sphere’s optimal calibration position. Next, an “cross-scan” procedure is devised to locate the sphere rapidly and precisely at the antenna main-beam center and the midpoint of the chosen range bin. Finally, using the S-band dual-polarization reference radar (Z9740) at the Changsha Weather-Radar Calibration Center as a test case, results show that when the sphere is centered in the range bin, the mean deviations of reflectivity (Z) and differential reflectivity (ZDR) from their theoretical values are only -0.2 and 0.16 dB, the retrieved beamwidth, antenna gain and pulse width differ by just 0.01°、0.14 dB and -0.11 μs. If the sphere straddles adjacent range bins, the Z and ZDR errors enlarge to -4.03 and 0.45 dB, and the beamwidth、gain and pulse-width errors increase to 0.06°、-0.17 dB and 0.56 μs. Comparative tests reveal that “pulse-window overlap” is the root cause of energy splitting and bias growth, and improvement measures such as dynamic position control and pulse windowing are recommended. The approach can tighten absolute calibration accuracy of Z and ZDR to within 0.5 and 0.2 dB, providing an absolute validation method for nationwide radar-network consistency assessments.

Abstract:To enhance the detection accuracy of underwater platforms for targets, this paper designs and validates a sonar system mounted on the flank side. The system omprises a hytrophone array and an electronic subsystem, operating in passive sonar mode. System employs phase-shifted beamforming technology to process received signals in the spatial domain for target azimuth estimation, and utilizes matched filter technology to process echo signals for target range estimation. The electronic subsystem incorporates low-noise signal conditioning and synchronous acquisition modules to achieve low-noise synchronous capture of 48-channel acoustic signals, and features a high-speed, large-capacity storage system to ensure real-time data recording. Experimental results demonstrate that the electronic system self-noise spectral density is as low as 2.52 nV/√Hz @40 kHz, with inter-channel amplitude consistency error below 0.4 dB and phase consistency error under 2°. The system exhibits an overall amplitude consistency error of better than 3 dB and a phase consistency error of below 6°. It is capable of achieving a resolution of 2.10° and 1.35 meters, with a localization error of 3.96° and 2.12%.

Abstract:Accurate lane line information is crucial for path planning and decision-making modules in autonomous driving systems. To achieve high-precision lane detection, this paper leverages the inherent characteristics of lane lines—being slender and structurally continuous—as prior knowledge, proposing a lane-prior guided multi-scale lane detection network. The network employs ResNet-18 as the backbone feature extraction structure. Gated fusion modules are introduced at different stages to effectively suppress redundant information propagation and enhance cross-layer information interaction efficiency. Addressing the slender, continuous, and spatially extensive nature of lane lines, a large kernel attention mechanism is incorporated into the feature pyramid network (FPN), significantly expanding the receptive field and strengthening the network’s perception capability for lane lines across different scales. In the detection head, lane priors are combined with global features, effectively resolving the challenge of accurately identifying lane lines when relying solely on local pixels. Furthermore, an auxiliary segmentation branch is designed, utilizing a cross-entropy loss function to optimize the parameter update process and enhance the model’s expressive power for fine-grained features. Experiments conducted on the public TuSimple and CULane datasets demonstrate the effectiveness of the proposed method: it achieves 96.96% accuracy on the TuSimple dataset and an F1-score of 80.1% on the CULane dataset. The proposed method enables high-precision lane detection in complex scenarios and exhibits strong competitiveness among existing lane detection approaches.

Abstract:The impedance matching and radiation performance of the RFID tag antenna will be affected in the metal environment. Meanwhile, in order to reduce the thickness of the tag, a miniaturized low-profile double-arm spiral anti-metal UHF RFID tag is proposed. The tag antenna is composed of an upper radiating patch and a lower ground patch. The structure of the upper radiating patch is that two dipole arms spirally extend outward from the center, presenting a central symmetry, and the chip is located at the symmetrical center of the two spiral arms. A metal short-circuit arm is introduced from the dipole arm of the upper radiation patch to connect the lower layer to the ground. By changing the width of the dipole arm and the position of the short-circuit arm in the radiation patch, the resonant frequency of the tag antenna can be effectively adjusted and the size of the antenna can be reduced. The size of the tag antenna proposed in this paper is 27 mm×0.6 mm. The tag is attached to the upper surface of the metal cylindrical water valve. Under the condition of an equivalent omnidirectional radiation power of 3.28 W, the measured reading distance around one circle in the horizontal direction is 8.5~9.3 m. The tag antenna designed in this paper features low profile, metal resistance, omnidirectionality, long reading distance and easy production. It is suitable for cylindrical metal application scenarios such as metal water valves and metal bottle caps for alcoholic beverages, and is managed in aspects such as production transportation, identification and monitoring using RFID technology.

Abstract:To improve the accuracy of automatic recognition of depression, a fusion deep residual network is proposed as the encoder, and an improved long short-term memory network and self attention mechanism are introduced to enhance the network model’s perception ability of data temporality, resulting in a well performing PLSTM AttnNet model. Firstly, preprocessing operations such as filtering and denoising are performed on the EEG signal to extract six typical time-domain data and differential entropy and power spectral density covering five frequency bands: δ, θ, α, β, and γ. Input the preprocessed data into the ResNet encoder to extract spatial dimension information, use the PLSTM module and self attention module to process the output features of the ResNet encoder to capture the temporal dimension information of the data, and improve the recognition accuracy of depression. A large number of comparative experiments were conducted on the publicly available EEG dataset, and the quantitative results showed that the proposed PLSTM AttnNet model achieved a classification accuracy of 97.3%. Compared with the single ResNet model and ResNet PLSTM model, the accuracy was improved by 10.1% and 3.9% respectively, and it was significantly better than the comparative model in evaluation indicators such as F1 score. The results indicate that the innovative fusion of ResNet and PLSTM Attention effectively enhances the collaborative representation ability of EEG signal temporal spatial features, solves the problem of one-sided feature capture in traditional models, and provides a reliable solution for efficient automatic recognition of depression, which has important clinical application value.

Abstract:To address the problems that fuse trimming in chip manufacturing still relies on sequential search of trimming codes, leading to frequent read-write operations, long testing time, and high hardware resource consumption, this paper designs a single-shot trimming-code prediction model based on a GBDT multi-classification framework. Overfitting-resistant mechanisms are introduced at the structural level: during residual learning, part of the residual samples is randomly discarded to weaken excessive memorization of abnormal and noisy samples, while Gaussian perturbations are injected into the leaf-node outputs to enhance robustness and generalization under fluctuations in data distribution. In addition, a fuzzy-region discrimination strategy is incorporated to construct a mathematical description of trimming-code fuzzy regions, enabling secondary discrimination and correction of samples located near adjacent class boundaries, thereby effectively reducing misclassification between neighboring categories and achieving refined identification of uncertain trimming codes. Experimental results show that, in the fuse-trimming indexing task, the proposed model achieves an overall prediction accuracy above 97.8% and an R2 value greater than 0.99, while significantly shortening test time and reducing read-write operations and hardware resource overhead.

Abstract:In response to the high-precision requirements for ground calibration of micro-thrusters in major aerospace missions such as space gravitational wave detection, and the severe limitation of thermal noise caused by environmental temperature fluctuations on the performance of measurement devices, this paper conducts research on thermal noise suppression technology. Based on the torsion pendulum micro-thrust measurement principle, a comprehensive analysis of error sources such as the thermoelastic effect of torsion wire stiffness, structural thermal expansion, and sensor drift is performed, revealing the mechanism of environmental thermal noise on measurement accuracy. On this basis, a precision environmental temperature control system based on field programmable gate array (FPGA) control is developed. This system adopts a hardware architecture of an external liquid circulation constant temperature enclosure, combined with an adaptive smith fuzzy PID control algorithm, to construct a highly stable, interference-resistant quasi-adiabatic working environment for the measurement device. Experimental results demonstrate the superior performance of the developed system, successfully suppressing the peak-to-peak temperature fluctuation in the core measurement area to within 20 mK; compared to the state without temperature control, the temperature gradient at both ends of the torsion wire is reduced to 0.5% of the original value, and the theoretical relative measurement error caused by temperature fluctuations is reduced from 0.66% to 0.003 7%. The research results indicate that this temperature control system effectively mitigates the interference of environmental thermal fluctuations on micro-thrust measurement, enhances the measurement accuracy and thermal stability of the device, and provides critical technical support for the high-precision calibration of micro-thrusters.

Abstract:With the rapid development of artificial intelligence models, the reasonable development and utilization of these models to serve people’s daily lives have become increasingly important. Given that traditional recommendation systems are unable to perceive users’ real-time emotions and therefore struggle to provide emotion-aligned book recommendations, this study proposes an intelligent book recommendation agent that integrates facial expression recognition technology with artificial intelligence models. In order to make the expression recognition results more accurate, this paper innovatively uses MTCNN-MobileNet_V2 fusion architecture for training. The training results show that the average accuracy of this method is about 23.6% higher than that of the traditional HOG based recognition method. The agent is based on the Raspberry Pi 4B hardware platform. Firstly, the user’s face image is collected in real time through Camera V2, and the face region is detected and aligned using MTCNN algorithm. Then, the expression classification is performed using MobileNet_V2 lightweight convolutional neural network; Then, the system integrates the large language model Qwen-3-32B to perform semantic reasoning and generate personalized book recommendations; Finally, MQTT and Home Assistant are used to realize the real-time display of recommendation results. The experimental results show that the average accuracy of this system is 94.72%, and the accuracy of emotion recognition is 97.7% under standard light (≈800 Lux); 83.1% under low illumination (≈350 lux). In the five person scene, the recognition accuracy is about 92%, the response delay is about 5 seconds, and the system response delay is moderate, which can realize real-time recommendation. The agent significantly improves the recommendation accuracy and user experience, and verifies the effectiveness of this method, which provides an effective basis for personalized recommendation based on emotional input in the future.

Abstract:As the core component of wind power equipment, the health status of wind turbine blades is crucial for the normal operation of the entire system. In response to the real-time monitoring requirements of blade condition parameters during operation, a distributed real-time monitoring system based on 5G communication has been designed and implemented. The system employs a two-layer distributed architecture, consisting of a multi-node data acquisition unit cluster and upper computer software. A two-level synchronous trigger strategy has been proposed to achieve high-precision synchronous sampling across all data acquisition channels, with the trigger delay being only 0.13% of the minimum sampling period. A dynamic adjustment model of trigger rate and data transmission rate was developed by fusing the compressed sensing technology, and the parameters were optimized accordingly. This approach enables the dynamic matching of data collection and transmission in the variable environment of 5G networks, thereby improving the reliability of system data transmission. Experimental results indicate that the fluctuation of background noise is less than 0.5%, the nonlinear error is below 1.1%, the channel consistency error is under 1%, and the reconstruction accuracy of compressed sensing reaches 98%. This design not only provides an effective solution for the real-time monitoring of wind turbine blades but also serves as a technical reference for the development of other online real-time monitoring systems.

Abstract:For the systematic evaluation of the anti-jamming performance of 5th generation advanced (5G-A) systems in complex electromagnetic environments, this study addresses the shortcomings of existing jamming test methods in simulating realistic non-orthogonal interference. A dedicated 5G-A test platform based on an orthogonal mismatch interference mechanism is designed and implemented. The platform innovatively proposes an interference generation method that destroys subcarrier orthogonality by introducing a normalized frequency offset, establishing an analytical relationship between the attacker’s frequency offset and the efficiency of interference energy diffusion. In terms of hardware implementation, the platform employs an FPGA combined with high-speed ADC/DAC modules and a configurable RF front-end, achieving precise generation and high-precision control of interference signals in terms of frequency, power, waveform, and timing. To comprehensively verify the platform’s effectiveness and evaluate the anti-jamming capability of key 5G-A channels, two test scenarios—RF cable direct connection and static/dynamic over-the-air measurement—were employed. The performance of the synchronization signal block (SSB), physical downlink control channel (PDCCH), and physical uplink control channel (PUCCH) under orthogonal mismatch interference of varying intensities was analyzed and examined. Experimental results indicate that, compared to traditional narrowband jamming, orthogonal mismatch interference can increase the terminal bit error rate by a factor of two under the same transmit power, while enabling accurate testing of the antijamming performance of the aforementioned key signals and channels. This work holds positive significance for promoting the reliable deployment of 5G-A networks.

Abstract:In long-range reconnaissance missions, multimodal image fusion aims to generate fused images that preserve salient features from each modality while maintaining photorealism. To address the challenges of ambiguous feature representation and structural-textural distortion in distant targets, this paper proposes BalanceFusion, a novel fusion architecture integrating diffusion-based reconstruction with wavelet-domain frequency-aware perception, specifically designed to enhance detail representation and fidelity in infrared and visible image fusion under long-range conditions. The proposed method first employs an efficient image restoration network to strengthen target feature representation. Subsequently, a wavelet-conditioned fusion module is introduced to incorporate frequency-domain awareness, emphasizing thermal radiation contours of targets while mitigating the adverse effects of super-resolution methods that overly prioritize foreground targets at the expense of background structural integrity. Finally, modality-specific high-frequency features exhibiting significant differences are hierarchically aggregated across multiple scales to ensure the fused image retains both sufficient discriminative details and authentic shared structures. Experimental results on real-flight datasets demonstrate that the proposed algorithm outperforms state-of-the-art methods, achieving a 33% improvement in the structure-preserving perceptual quality metric Qabf and 21% gain in spatial frequency (SF).