The effect of discharge pulse interval and number on the discharge and ignition characteristics of nanosecond pulse surface dielectric barrier discharge （nSDBD） in NH₃/air mixture was studied. The experiment is conducted in a constant volume combustion chamber. Discharge characteristic tests show that nSDBD can generate multiple independently distributed discharge filaments， expanding the range of plasma. When using 30 pulse numbers with a pulse interval of 50 μs， at least 12 discharge filaments are excited， with filament lengths exceeding 14 mm. The ignition characteristic test shows that nSDBD has achieved spatial multi-point ignition. When the pulse interval is 50 μs， 180 discharge pulses generate 6 initial flame kernels surrounding the top electrode. In the NH₃/air mixture， there is a phenomenon of initial flame kernels dissipating， resulting in a discrepancy between the number of initial flame kernels and the number of stable combustion flame kernels. As the pulse interval or pulse number increases， the area of the initial flame kernels increases， and the number of stable flame kernels gradually increases from 1 to 6. This effectively shortens the ignition delay of NH₃/air mixture， with a reduction of up to 30 ms， approximately 55%. The adjustment of pulse interval and pulse number can effectively control the combustion phase of NH₃/air mixture.

A combination of genetic algorithm and recursive least squares method was used to estimate the stability parameters of a semi-trailer， solving the problem that parameters such as tire cornering stiffness， vehicle roll stiffness and body roll damping are difficult to directly measure through sensors. In terms of adaptability to working conditions， the shortcomings of traditional offline identification methods are effectively made up for via this method. Compared with general commercial vehicles， semi-trailers have a more complex structure and diversified operating conditions. Therefore， in the process of ensuring vehicle safety and quality， greater attention must be paid to the stability control of semi-trailers. The prerequisite for achieving this goal is to establish a high-precision and high-confidence theoretical model of semi-trailer dynamics. On this basis，the theoretical model can be used as a following target， and the difference between the output state of the actual vehicle or commercial software vehicle model and the output state of the theoretical model is used as the control variable for adjustment. Joint simulation of Trucksim and Simulink is used in this paper to compare the output overlap between the Trucksim software model and the theoretical model under specific input and working conditions. Results show that the theoretical model established based on the parameters such as roll stiffness indentifled by the method proposed in this paper is superior to traditional offline identification method in terms of operating condition adaptability and accuracy， the estimation error is reduced by about 6%. This result lays the foundation for subsequent semi-traller stability control research based on this theoretical model.

Based on a 2.2 L ammonia-fueled internal combustion engine equipped with an active pre-chamber， hydrogen was used as the pilot fuel， and the effects of different hydrogen injection parameters on combustion performance were analyzed via numerical simulation. The results indicate that compared with injection pulse width， hydrogen injection pressure and injection end timing have a more significant impact on combustion performance， particularly during the CA10-CA50 phase of the combustion duration. The research findings can provide a reference for the design of injection strategies for ammonia-hydrogen internal combustion engines.

To improve the magnitude and distribution uniformity of current density at various locations of the fuel cell membrane electrode， a polynomial function is used to approximate the distribution law of catalyst loading， and a weighted objective function is constructed for the average current density and current density uniformity of the fuel cell. By using numerical simulation to solve the mapping relationship between polynomial coefficients and weighted objective function， the distribution function can be optimized. The results showed that compared with fuel cells with uniformly distributed catalyst loading， fuel cells with optimized non-uniform catalyst loading distribution showed a significant improvement in current density uniformity， although the average current density slightly decreased. Considering both the magnitude and uniformity of fuel cell current density，non-uniform catalyst loading surface surpasses the uniform one. Furthermore， the first-order distribution function indicates greater advantages.

To address the issue of a single output port in traditional motors， a novel multi-port disc-type permanent magnet motor （MDPMM） was proposed. First， the structure and working principle of this motor is introduced. Second， the impact of the positioning force of the ring-arc stator on thrust was analyzed and calculated. Due to the numerous electromagnetic structural parameters affecting the positioning force， a generalized regression neural network （GRNN） was designd to establish a rapid calculation model for the MDPMM. By constructing a finite element model of the ring-arc stator region， a parameter sample library was obtained as input for the GRNN. The superiority of GRNN is verified by comparing it with support vector machines （SVM）. With the optimization objective of "no reduction of thrust density and minimum fluctuation of thrust" ， the particle swarm optimization（PSO）algorithm is used to optimize the structural parameters of the ring-arc stator region. Finally， the effectiveness of the hybrid GRNN-PSO algorithm is validated through comparative simulation analysis before and after optimization.

To address the problem that traditional signal separation algorithms cannot efficiently and accurately analyze specific faults, a signal extraction method combining Variational Mode Decomposition (VMD), Laplacian Energy (LE) and Variational Mode Extraction (VME) was proposed, and multi-class Relevance Vector Machine (mRVM) together with Dempster-Shafer (DS) evidence theory was adopted for intelligent fault diagnosis. This method is dedicated to the small-sample data scenario. First, the VMD-LE-VME method is used to extract effective fault information from fault signals and obtain multi-domain features. Second, the multi-domain features are input into the mRVM for fault identification. Finally, the classification results are fused by means of DS evidence theory to derive the final diagnosis results. Experimental results verify the effectiveness and superiority of the proposed method in handling small-sample data.

In order to solve the problem of feature corrosion and submergence caused by direct fusion of deep and shallow features and realize accurate lane detection in complex environment， a lane detection method based on adaptive fusion of double-branch features was proposed. Firstly， a dual-branch feature extraction network was designed in the method to enhance the feature extraction capability for lane lines in complex environments and reduce the loss of spatial detail information. Secondly， a feature adaptive fusion module was constructed， in which channel attention and self-attention were utilized to guide feature selection and fusion. The fusion process was adaptively adjusted to optimize the channel and spatial semantic information of feature maps. In addition， the improved parallel hybrid pyramid pooling module is more in line with the characteristics of long and narrow roads and captures remote context in multiple directions. Finally， the proposed method was tested on TuSimple， CULane and Curvelanes data sets， and the F₁ reaches 96.93%， 76.48% and 83.21% respectively. The experimental results show that the proposed method can effectively deal with lane line detection tasks in complex scenes such as occlusion and shadow， and its performance is significantly improved compared with the mainstream segmentation lane line detection methods.

The vehicle speed detection in the single Global Navigation Satellite System （GNSS） environment is subject to significant noise interference， and the sparsity of data sampling will further exacerbate the instability of speed calculation. To address this issue， the vehicle speed estimation under the single GNSS scenario is formulated as a speed expectation optimization problem based on spatiotemporal correlation， and a self-supervised bidirectional Long Short-Term Memory （LSTM） algorithm is proposed for its solution. Firstly， a sparse data spatiotemporal feature extraction LSTM is constructed by this algorithm， where time gate and spatial gate functions are introduced to analyze the changes in speed spatiotemporal correlations in sparse and unequal-interval vehicle detection data， and embedded vectors for speed spatiotemporal features are extracted. Secondly， the trend of vehicle speed changes is analyzed from both forward and backward directions by the noise bidirectional suppression LSTM network for vehicle speed estimation， enabling the accurate achievement of noise elimination and speed estimation. Finally， experimental verification of the vehicle speed estimation performance of the proposed algorithm was conducted using GPS signals as an example. The results show that an average noise reduction of 26.73 dB PSNR was achieved by the proposed algorithm in sparse speed data with sampling intervals ranging from 1 minute to 10 minutes， which is 28.93% higher than that achieved by the LWR， EnKF， and Noise2Void algorithms on average. Additionally， the speed estimation accuracy of the proposed algorithm is 2.02% higher on average.

An adaptive control algorithm based on the in-phase queue dissipation was proposed to address the issues of low control efficiency and high model complexity in existing signal control algorithms. Firstly， a prediction model was established to optimize signal timing parameters， and calculate the vehicle's stopping delay based on traffic shockwave theory. Secondly， the acceleration and deceleration delays of vehicles were calculated based on the speed variation curve. Finally， state transition matrix for constructing dynamic optimization model was proposed to minimize delays， thereby realizing full-stage optimization of intersection signal control. Compared with fixed-time control and actuated control methods， the proposed method respectively reduced the average delay by 55% and 27.8% per vehicle in ultra-high-traffic demand scenario and reduced the average number of stops by 58.5% and 10.1%. In high-traffic demand scenario， the average delay decreased by 36.1% and 14.6% per vehicle， and the average number of stops decreased by 23.4% and 8.7%， respectively. In medium-traffic scenario， the average delay decreased by 22.8% and 10.5% per vehicle， and the average number of stops decreased by 3.3% and 2.6%. The results indicate that the algorithm proposed in this study can effectively improve the traffic operational efficiency at isolated intersections under different traffic demand conditions.

To address the issue of inaccuracy of vehicle lane change decision in the mixed traffic flow scenario of HV and AV， a vehicle lane change decision model was proposed. The model is based on game theory， multiple game functions were constructed for the continuous cooperative lane change situation of adjacent vehicles， the influence of communication uncertainty was eliminated by Harsanyi transform， and the driving style of vehicles was distinguished by K-means++ clustering， the game returns were further adjusted by risk factors. The lane change model was verified by using the SUMO simulation platform. The experimental results showed that under the fixed AV permeability， the average passing number of vehicles is effectively increased and the average passing time is reduced by applying the combined game theory and driving style lane change model， at the same time， no accidents occurred in the test， which proves the stability and safety of the lane change model. Under different AV permeability， the average passing time of vehicles decreases significantly with the increase of permeability， which indicates that the AV can effectively utilize the lane.

To improve traffic safety in highway tunnels， it is suggested to use line shaped visual guidance facilities to enhance local brightness and contrast of the tunnel， and form a linear guidance system that outlines the tunnel outline and road alignment. Firstly， set up a tunnel simulation scenario， conduct driving simulation experiments and questionnaire surveys， and analyze the impact of various linear induction facilities/systems on drivers' vision， psychology， and driving tasks； Then， explore the effects of the size， direction， continuity， and combination of various induction facilities on the optimization of visual reference frames and spatial rights-of-way perception. The results show that linear guiding system can effectively enhance the salience of the tunnel environment， reduce the difficulty of obtaining visual information， and ease the driver's nervousness. Longitudinal linear guiding facilities are conducive to clarifying the road boundary and the direction of travel， and enhancing the driver's perception of location. Vertical and horizontal linear guiding facilities help to clarify the outline of the tunnel and enhance driver's speed control. Through a reasonable combination of different linear guidance facilities， to complement the shortcomings， to achieve the overall improvement of driving safety indicators.

Aiming at the problem that drivers' depression and tension are easily induced by narrow space， insufficient lighting and limited sightline in tunnels， which further leads to the phenomenon of "slowing down whenever passing through a tunnel" and thus impairs traffic efficiency， a coordinated control method for tunnel vehicle speed and lighting is proposed.First， drivers' visual adaptation characteristics to light intensity in tunnel environments are analyzed， calculation models for the optimal lighting intensity and range under different driving speeds are established， and the control scheme of lighting variation in tunnels is optimized to reduce the interference of the black hole and white hole effects on drivers. Then， combined with indicators such as actual vehicle speed and traffic density in tunnels as well as the levels of emergencies， a method for determining the safe operation speed of tunnels is designed， and the tunnel lighting illuminance is adjusted to meet the requirements of drivers' safe driving. Finally， simulation verification is carried out by using the actual operation data of tunnels in Shandong Province. The results show that the average operation speed and safety efficiency of tunnels are improved by this method through the coordinated control of tunnel vehicle speed and lighting. At present， this method has been put into practical application in some tunnels of the Jiwei Expressway in Shandong Province.

In order to improve the utilization efficiency of existing parking resources and reduce the traffic congestion， exhaust emissions and other problems caused by travelers' invalid parking space search behavior， a multi-objective nonlinear integer programming model for optimizing the allocation of parking resources was proposed. Firstly， based on the comprehensive consideration of user walking distance， parking fees， the balance of utilization of various parking lots in the area， and the additional traffic pressure attached to parking behaviors， the personal cost and social cost functions of parking allocation were established. Then， with the optimization objective of minimizing the comprehensive cost of the system， a multi-parking intelligent parking allocation model was constructed， and at the same time， considering the complexity of model solving， an improved Kepler optimization algorithm integrating multiple strategies was designed for model solving. Finally， in order to test the effectiveness of the model， numerical experiments under different parking supply and demand situations were designed， and the proposed model and algorithm were compared and analyzed with the classical allocation model and the traditional solution algorithms. The results show that the proposed model reduces the individual cost by 4.4% on average， and the balance of utilization of each parking lot is significantly improved. Meanwhile， the proposed model has obvious advantages in reducing the additional traffic pressure of the road network caused by parking groups， with a maximum reduction of 33.6% in the impedance growth rate of the road segment. Compared with traditional genetic algorithm and simulated annealing algorithm， the proposed improved Kepler optimization algorithm has faster convergence speed and a better ability to search for optimal solutions.

To address the issues of insufficent safety performance and low learning inefficient in traditional autonomous driving systems， an autonomous driving safety decision-making model capable of continuous learning and understanding linguistic information was proposed. Referring to the reasoning decision-making and experience accumulation processes in human driving， this model leverages a large language model （LLM） as the decision-making agent， integrating chain-of-thought reasoning， a two-stage attention mechanism， and cognitive memory storage and retrieval into the contextual safety learning of the driving process. Meanwhile， a kinematic module is employed to convert LLM decisions into executable driving commands， enabling the continuous learning of safe driving experiences. Experimental results demonstrate that the proposed decision-making model significantly improves safety and efficiency compared to rule， reinforcement learning， and knowledge-based approaches， and possesses the capability of continuous learning and adapting driving behaviors based on human instructions， providing a reference for human-like autonomous driving.

The calculation methods for the direct shear carrying capacity of Ultra-High-Performance Concrete （UHPC） flat wet joints and keyed wet joints are investigated. By considering the construction and force-bearing characteristics of UHPC wet joints， a direct shear strength model for UHPC wet joints is proposed， which incorporates key factors such as interfacial cohesion， frictional resistance， aggregate interlock strength of keys， and fiber reinforcement contribution. Based on the results of direct shear tests on joint specimens and finite element simulation， linear regression analysis is conducted to determine the interfacial bond stress coefficient and friction coefficient. Through the derivation of the aggregate interlock strength of keys and the discussion on the fiber contribution， a formula for calculating the shear carrying capacity of UHPC wet joints is presented. Nine existing formulas for calculating the shear carrying capacity of UHPC wet joints are summarized， and the proposed direct shear carrying capacity calculation model is validated by comparing the predicted carrying capacity from the nine existing formulas with experimental values. The results indicate that with the increase in lateral confinement stress， the failure load of flat joints exhibits a nonlinear relationship； the ratio of the proposed formula to the experimental values has an average of 1.04， which is closer to 1 compared to the nine existing formulas， demonstrating superior prediction accuracy of the proposed formula. The verification of different experimental approaches on the shear carrying capacity calculation method for UHPC wet joints shows that the predicted values from the proposed direct shear carrying capacity calculation model align well with experimental values， thus proving the universality of the calculation method. This method can provide a reference for predicting the direct shear carrying capacity of UHPC keyed wet joint interfaces.

To deepen the research on the mechanisms of foundation and inundation， the effect of inundation on the bearing characteristics of long and short pile foundations in loess areas was explored through modeling tests. Sixteen combined piles （8 long and 8 short） were adopted in the test， and the wet subsidence deformation was comprehensively analyzed， pile bearing capacity and deformation characteristics under the conditions of vertical load and water immersion. The results show that the Q-s curve of the long and short pile foundation under vertical load and water immersion shows a typical slow-varying curve， and the settlement of the pile end gradually increases with the increase of water immersion and load. Water immersion intensifies the deformation of the wetted soil layer around the pile， and the wetted deformation is slowed down then sharply and then slowed down again， and the increase of the wetted depth leads to the development of pile lateral friction and the downward shift of the neutral point. After 10 days of water immersion， the maximum value of negative friction resistance of long pile is 61.27 kPa， and that of short pile is 53.85 kPa. When immersed in water， in terms of the speed of negative friction resistance reduction and the range of change of the depth of the neutral point， both are the most significant side piles， and the center piles and corner piles are close to each other； the maximum negative friction resistance occurs in the corner piles， and the center piles are second， and the side piles are the last. The maximum negative friction resistance occurs in the corner pile， followed by the center pile and the last one in the side pile. After the soil body is saturated， the neutral depth ratio of the long pile （side pile） is 0.55， the neutral depth ratio of the corner pile and the center pile is 0.64， and that of the short pile is in the range of 0.62-0.64， which is close to the recommended value of the pile foundation specification. Finally， the finite element simulation experiment simulated the water immersion test through the hydraulic equivalence principle， and the results were consistent with indoor tests， indicating the reliability of the experiment.

To explore the cracking characteristics of asphalt concrete， a finite element software ABAQUS was employed to simulate the semi-circular bend （SCB） test of asphalt concrete using a bilinear cohesive zone model （CZM）. The simulated load-load line displacement curve was compared with the experimental result to validate the applicability of the CZM in cracking analysis. Based on this， the internal stress distribution， system energy balance， fracture behavior parameters during the cracking process of the SCB specimen， and the effects of cohesive zone model parameters on the load-load line displacement of the specimen were analyzed. The results indicate that the bilinear CZM has good applicability in cracking analysis of asphalt concrete. During the loading process， the SCB specimen undergoes elastic， damage， and fracture stages. The point of instability of crack propagation does not correspond to the peak load， leading to an overestimation of fracture toughness when calculated using the peak load； a portion of the fracture work converts into elastic strain energy in the uncracked region of the specimen， resulting in an overestimation of fracture energy when calculated using fracture work. The peak load that the SCB specimen can withstand during the cracking process mainly depends on the tensile strength of asphalt concrete rather than fracture energy.

To address the melting and softening issues of phosphogypsum （PG）-filled asphalt pavement after exposure to water， a hydrophobicity treatment of PG was introduced to reduce solubility. An in-depth investigation was conducted into the moisture stability characteristics and solubility reduction mechanisms of asphalt pavement incorporating modified PG. The results showed that the immersion residual stability and freeze-thaw splitting strength ratio did not meet the moisture stability requirements for pavements. Additionally， after 600 hours of immersion， the number of PG filler particles decreased by 52.60%. Following hydrophobic modification， the polar component of PG decreased by 72.46%， the interface adhesion strength improved by 11.15%， and the thickness of the interface transition zone increased by 546%. The mechanism is that silane coupling agent can undergo dehydration condensation reaction on the surface of phosphogypsum and graft to form a silanol group coating. The hydrophilic property of the phosphogypsum surface is altered by the coating， which effectively prevents the intrusion of water and the displacement of the asphalt film， thereby improving the water stability performance. These results provide a theoretical basis for expanding the application of PG in road construction.

In order to study the shear resistance of the interface between ultra-high performance concrete （UHPC） and normal concrete （NC）， the effects of different matrix concrete strength， interface reinforcement planting rate and reinforcement buried depth on the shear strength of the interface were investigated through Z-type direct shear test， and the calculation formula of shear bearing capacity of uhpc-nc reinforcement planting interface was established the results show that when the strength grade of the matrix concrete is C30 and C40， the interface planting rate is recommended to be 1.22%， and when the strength grade of the matrix concrete is C50， the interface planting rate is recommended to be 1.59%. In order to ensure the shear resistance of the interface， the buried depth of the reinforcement on the UHPC side should be greater than 8D. Finally， the formula for calculating the shear bearing capacity of UHPC-NC planting reinforcement interface was proposed.

Network security issues are becoming increasingly prominent， and IoT network security urgently needs further investigations. Traditional IoT intrusion detection methods have weak feature representation capability for sequence data， and most of the methods based on machine learning and deep learning rely on complex feature preprocessing techniques and have weak global modeling capability for high-dimensional sequence data. To address the above problems， we propose a FATIDS-based IoT intrusion detection method， which achieves end-to-end feature selection and feature extraction through the self-attention mechanism， dynamically adjusts the attention to sequence features， and improves the global modeling capability for high-dimensional sequence features. To solve the imbalance problem faced by IoT intrusion detection， the Focal Loss is utilized to dynamically scale the model gradient， adaptively reduce the weight of simple samples， and focus on classes that are difficult to classify. Finally， the performance of the proposed method is validated on the ToN_IoT and DS2OS standard datasets， and the experimental results show that the proposed method achieves superior detection performance compared to other remarkable methods， and the impact of important hyperparameters on the performance of the proposed method is also validated.

To ensure that the embedded area of the watermark is preserved with a high probability during cropping， we propose an effective and anti-cropping attacked watermarking method. Firstly， we use a saliency detection network to select the embedding region； then we use a Transformer network to embed the transformed watermark into this region； to ensure the robustness of the watermark， a noise layer is used to attack the embedded image before extracting the watermark； finally， we use a fully convolutional network to extract watermark from the salient region. The experimental results show that the proposed method has higher robustness than compared methods under common nine attack methods. Our method can accurately extract the embedded watermark under saliency detection-based cropping attack.

An image dehazing algorithm based on a pyramid structure with multi-scale detail enhancement and hierarchical noise suppression is proposed to address the issues of detail blur and noise amplification in existing algorithms. Firstly， a multi-scale detail enhancement algorithm is designed to weight and fuse multiple different exposure images generated by gamma correction， resulting in a fog image after detail enhancement， along with corresponding detail layer and fuzzy layer images， for enhancing the details of the restored image. Secondly， a non-local weighted average algorithm is constructed to optimize the initial transmittance estimated by prior dark direct attenuation， so as to reduce morphological artifacts， while the final transmittance is obtained using a small radius Weighted Guided Image Filter （WGIF）. Finally， through the proposed multi-scale hierarchical noise suppression and fog removal algorithm， the fog-free image is restored while noise amplification is suppressed. Experimental results demonstrate that the proposed algorithm can better suppress noise amplification， producing fog-free images with clear details， natural colors， and higher-quality sky region restoration. Furthermore， multiple objective evaluation metrics are significantly improved compared to those of current mainstream algorithms.

In order to capture the changes of statistical features of images caused by stegography more comprehensively and improve the detection accuracy of spatial steganalysis， the embedding characteristics of the steganographic algorithm are incorporated. The derivative and Gabor double filter is used to preprocess the image， and the filter extraction is enhanced to produce a variety of residual images， which effectively increases the diversity of steganographic features. The optimized CBAM module is embedded into the residual block to guide the network to effectively focus on the region with rich steganographic signals， thus strengthening the discriminant learning ability and training effect of the network. The proposed model is compared with the classical model on BOSSbase v1.01 and BOWS2， and the experimental results show that the detection accuracy of the proposed method is superior to the existing mainstream models of Ye-Net， SRNet and ZhuNet.

Aiming at the complexity of features in remote sensing images and the existence of an elongated and continuous distribution of roads that are easy to obscure， a Road Extraction Model for Remote Sensing Images Combining Attention and Context Fusion （ACFD-LinkNet） was proposed. The network is based on the D-LinkNet network. Firstly， a strip attention module was used in the codec part of the D-LinkNet network to enhance the feature extraction capability of roads at different scales， to better capture the global features of the roads， and to capture the long-distance information of the roads. Secondly， a Context Fusion Module （CFM） was proposed and added to the feature delivery part of the network codec to predict road connections between neighboring pixels， fusing road information between different layers of the context to solve the problem of obstacle obstruction interfering with road connections. Finally， the cross-entropy loss function and Dice loss function of the improved model were set up with multiple loss function hyperparameter weight assignments to solve the dataset positive and negative sample inhomogeneity， and the optimal segmentation accuracy was obtained by adjusting the weight ratios. Experiments on the DeepGlobe and CHN6-CUG datasets resulted in F₁ values of 86.76% and 92.12% for the composite metrics， respectively， which is an improvement of 3.96% and 1.13% compared to the D-LinkNet model， in addition to optimal performance compared to semantic segmentation methods such as Unet， Deeplabv3+， A²-FPN， etc.

A end-to-end entity relationship extraction model based on span and semantic features was proposed to address the problems of traditional entity relationship extraction methods relying on distance measurement or simple span recognition， which make it difficult to capture potential relationships between entities， and the high computational complexity and error propagation of the model. Firstly， the text vectors were randomly segmented into span sequences so that the model can learn a wider range of semantic feature information. Secondly， semantic relationships were judged to screen out subsets of candidate relationships， thus reducing information redundancy. Finally， the candidate relationships were transformed into relationship-span combinations containing important relationship semantics， and the Transformer decoder was used to achieve the joint extraction of entity relationships. The experimental results show that the F₁ value of this model is significantly improved in the NYT and WebNLG datasets compared to other baseline models， proving its effectiveness.

To address the problems of poor attack performance and low authenticity of adversarial examples caused by the perturbations generated by the adversarial example generation method based on generative adversarial networks （AdvGAN） deviating from key image regions and lacking controllability， a Diff-AdvGAN adversarial example generation method was proposed. Firstly， an Adaptively Spatial Feature Fusion （ASFF） module was employed to fuse featusare maps of the images at different scales. Then， the fused feature maps werere input into a generator to produce perturbations， and a Stochastic Differential Guide Module （SDGM） was used to enhance the controllability of the perturbations and generate adversarial examples. Finally， the adversarial examples are fed into a discriminator and a target model， the loss values were iteratively calculated and fed back to the generator to generate stronger perturbations with improved attack performance. Experimental results show that the Diff-AdvGAN method achieves attack success rates of over 99% on the MNIST dataset for the LeNet C， VGG11， and C&W models， and attack success rates of 96.17% and 95.82% for the ResNet18 and ResNet32 models on the CIFAR-10 dataset. Moreover， the perturbations generated by this method can accurately locate in the critical regions of the images， exhibiting high sparsity and small magnitudes， demonstrating significant advantages compared to comparison method.

The limitations of current image manipulation localization methods， such as the extraction of features at a single scale， the misdetection and omission of small tampered regions caused by background confusion， and the uncertainty existing in prediction results， are addressed. An image manipulation localization method based on edge uncertainty learning is proposed. The base features of the tampered image are extracted by means of a pyramid vision transformer. A coarse localization map is then generated through multi-level interactive coarse localization branches. To enhance the detection of small tampered regions， a small target-aware refinement branch is employed. Multi-scale feature fusion is achieved with the use of a dedicated module， which enables the full interaction and integration of features across different scales. Additionally， entropy-based perceptual loss is introduced to supervise boundary uncertainty， thus significantly reducing the uncertainty of the prediction results. The proposed method is evaluated on five widely-used public image tampering datasets in both in-domain and cross-domain experiments. It is demonstrated by the results that the method can effectively localize tampered regions and outperform existing approaches.

To inhibit the formation and development of energy holes in wireless sensor networks （WSNs） and to equalize the energy consumption and hole distribution， the effects of parameters such as node density， node residual energy， and data transmission within the survival zone of hole edges， as well as the distance of multi-energy holes， on the generation and development of energy holes are systematically analyzed. To objectively describe the characteristics of energy holes， a series of definitions such as energy hole， fusion energy hole， hole edge domain， and hole distance are proposed. Drawing on the theory of diffusion of multi-component liquid mixtures， an energy hole evolution model is constructed by calculating the diffusion angle and diffusion velocity of energy holes. On this basis， the WSN clustering method based on the diffusion coefficient of the energy hole， the WSN data transmission optimization method to suppress the energy hole， and the node dormant scheduling strategy in the edge domain of the energy hole are studied. Simulation results show that the proposed algorithm has significant improvement in terms of network lifetime， data transmission， energy equalization， and hole distribution compared with LEACH， UCDTS， and EHSRA algorithms.

To address the problem of inadequate infrastructure and inability to provide reliable communication in remote areas or disaster scenarios， a non orthogonal multiple access relay system with multiple drones assisting users is proposed. The system comprehensively considers constraints such as information causality， transmission power， and return delay， and jointly optimizes the offloading strategy， transmission power， resource allocation， and trajectory to achieve the goal of minimizing system energy consumption. Due to the non convexity and complexity of optimization problems， a two-stage online resource allocation scheme is designed for solving. In the first stage， to eliminate the dependence on uncertain information in complex scenarios， Lyapunov optimization theory is applied to decompose the optimization problem into three sub problems. In the second stage， an alternating iterative optimization algorithm was proposed： firstly， the user offloading decision was obtained based on the maximum weight matching algorithm； secondly， closed form solutions for transmission power and frequency allocation were obtained by utilizing auxiliary variable method， continuous convex approximation， and Lagrangian duality； finally， the trajectory planning problem is transformed into a convex problem and solved using the convex optimization tool. The simulation results show that compared to the benchmark algorithm， the proposed scheme significantly reduces system energy consumption while satisfying the long-term stability constraints of the system.