With regard to the new development trends of artillery technology with the support of informationized and intelligent technologies, this paper summarizes the technological progress of artillery in aspects such as dynamic kill chains, long-range high-precision strikes, anti-jamming reliable strikes, high-density firepower strikes, and data-driven intelligent firepower. Furthermore, in combination with typical battlefield applications and new directions of equipment development, it analyzes the positive effects of these advancements on enhancing artillery performance. The research indicates that modern artillery has achieved severalfold increases in performance metrics concerning closed-loop strike time, range, accuracy, and rate of fire. It has also laid the foundation for artillery to adapt to the requirements of intelligent battlefield confrontation in terms of anti-jamming capabilities in complex battlefields and self-sensing/self-controlling functions of artillery systems. On this basis, this paper further explores the future development trends of artillery technology and proposes cutting-edge technological directions such as new-quality launching methods, extended-range technologies, and group firepower capabilities for artillery systems. It also preliminarily analyzes the significant implications of technologies such as high-initial-velocity launching through hydro-oxygen combustion, cannon-launched scramjet propulsion, and intelligent firepower coordination for the new-quality development of artillery, providing a preliminary vision for the future evolution of artillery systems.

Accurate determination of extreme boundary conditions in the forward design of a fuse is a robust guarantee for its safety. The safety margins hold paramount significance in the fuze forward design, where environmental excitations, detonation control components, and other safety factors directly impact fuze reliability. To address these challenges, this study proposes a conceptual framework for the fuze safety margins and establishes an analytical evaluation method for the electromagnetic environment safety margins in the fuze systems. The method specifically addresses two key scenarios: (1) voltage, current, and energy safety margin calculations under known electromagnetic environments, and (2) transient/pulse electromagnetic conditions. Furthermore, the paper systematically elaborates on experimental principles for safety margin verification in fuze detonation control circuits, principles for test point selection, and methods for component parameter selection.These contributions provide references and support for the theoretical analysis and engineering application of fuse safety margins.

Lightning poses a serious threat to the intelligent unmanned tanks in the wild. To clarify the protection requirements for the lightning indirect effects, firstly, the electrostatic simulation model of unmanned tanks with the cloud electrode was established, and the lightning attachment points were analyzed. Secondly, the field distribution and the conduction coupling laws of internal typical single wires and coaxial lines in an unmanned tank subjected to a 200 kA lightning current A-wave impact were investigated through field-circuit co-simulation. The influence of different attachment points, ground conductivity and termination resistances was discussed. Finally, an experimental model of the unmanned tank was constructed proportionally, and the small current equivalent injection test was conducted in accordance with GJB8848. Then the measured coupling currents were linearly extrapolated to verify the validity of the numerical simulation results. The results show that the tank metal shell is more effective in shielding the pulse electric field, but less effective in shielding the low frequency magnetic pulse; the coupling voltage at high impedance port is similar to the differential signal of wave A component and its peak is about 2 kV; the coupling current at low impedance port accords with the wave A component and its peak is about 2 kA.

At present, the refined damage assessment of reinforced concrete structures subjected to penetration explosion mainly relies on numerical simulations and experimental data analysis, both of which suffer from the drawbacks of time-consuming processes and high economic costs. To address these issues, a novel digital-intelligent prediction model for penetration-explosion damage is proposed to simulate the entire dynamics of the penetration-explosion process. Based on the graph neural network and a multi-task learning paradigm, this model can achieve rapid and accurate predictions of building damage morphology and node failures under various penetration positions and explosion yields, with a single penetration-explosion prediction taking only 0.2 seconds. Specifically, during the penetration phase, the structural response field prediction error is less than 2.9%, and the node failure prediction accuracy exceeds 99.5%; during the explosion phase, the structural response field prediction error is lower than 6.5%, and the failure prediction accuracy surpasses 98.8%. The evaluation of 100 penetration-explosion damage scenarios takes only 30 s.The proposed digital-intelligent prediction model demonstrates excellent generalization performance and highly efficient, accurate prediction capability, thus providing a new tool for building structure damage assessment and safety protection design.

Currently, human society is entering an era of Intelligent Internet of Everything (IoE), characterized by the integration of humans, machines, and objects. Information and communication networks, as the crucial information infrastructure supporting the "Intelligent IoE," are undergoing unprecedented transformations. The development of these networks has a profound impact on people's lifestyles, economic growth, and social progress. The deep integration of artificial intelligence and information communication technologies makes intelligent information networks an inevitable trend in the evolution of network intelligence. In response to the diverse application demands of complex scenarios in the intelligent era, networks should have the ability to proactively adapt to changes in both internal and external conditions. Key features such as autonomous learning, optimization, management, and evolution enable networks to enhance the timeliness, effectiveness, and personalization of services, thus meeting the demands of the intelligent age. 

Mobile edge computing (MEC) can effectively address the issues of increasing business demands and greater processing difficulties for mobile users. However, users' service demands have differentiated characteristics in time and space. At the same time, network characteristics such as deep uncertainty in the network environment also restrict the efficiency of network edge task processing and service caching. In response to these challenges, this paper studies the users' service offloading and service caching issues in a multi-user and multi-edge server MEC system. To address the differentiated service requirements of users in both time and space, with the optimization goal of minimizing the average delay and energy consumption cost for all users under the base station, a joint optimization of task offloading and service caching based on the differentiated needs in time and space is established.Taking into consideration the characteristics of the network environment with deep uncertainty, a multi-agent deep reinforcement learning algorithm based on graph neural networks and long short-term memory networks is further proposed for autonomous learning and decision-making of user task offloading, resource allocation, and service caching strategies. Finally, simulations have verified that the proposed algorithm has good convergence and energy efficiency.

Radar signal sorting is a key technology in the electronic warfare system and an essential part of battlefield situational awareness. The development of new system radar technology brings a serious challenge to radar signal sorting in the current complex battlefield electromagnetic environment. For the problem that the traditional K-means clustering algorithm is sensitive to the initial cluster number K and the initial points when performing signal sorting on radar full pulse data, an optimized K-means-based radar signal sorting algorithm is proposed. The combination of water wave center diffusion(WWCD) optimization algorithm and the Canopy algorithm realizes the optimal selection of the distance threshold of the Canopy algorithm, and optimizes the selection of the cluster number K for K-means clustering, effectively reducing the sensitivity of the K-means algorithm to the selection of the initial number of clusters. Three kinds of UCI data and three kinds of frequency-hopping radar pulse data are used to verify the sorting performance of the proposed method, and the clustering effects are also compared with common clustering algorithms such as DBSCAN, OPTICS, and Canopy-K-means. The results show that the proposed method is insensitive to the setting of initial parameters and has a high clustering and sorting accuracy.

Object detection technology has great potential for progress in the field of UAV remote sensing images. In order to design object detection algorithms suitable for remote sensing images, it is necessary to accurately grasp the advantages and disadvantages of the current algorithms and explore new solutions. Firstly, this paper conducts a comprehensive and systematic analysis of object detection detection methods based on traditional techniques and deep learning, and compares the various algorithms. Secondly, it summarizes the commonly used remote sensing image datasets and evaluation metrics for object detection detection algorithms. Thirdly, to address the key challenges in remote sensing image object detection, such as small target size, occlusion, and complex background environments, the paper deeply analyzes the difficulties faced by the existing methods in handling these issues. Finally, considering the background environments and characteristics of remote sensing images, the paper discusses  the potential improvement directions for the current algorithms from five aspects, providing a reference for the development of object detection technology in unmanned aerial vehicle (UAV) remote sensing images.

When the conventional feature matching color transfer method is applied to the fusion of ultra-wide field-of-view infrared and low-light images, the problems such as high mismatch rate and poor fusion effect may occur. Thus, A natural color fusion method based on superpixel features and color clustering joint constraints is proposed. First of all, the superpixel segmentation is carried out on low-light image and space of color reference image, and  low-level, mid-level, and high-level superpixel feature sets are constructed sequentially. Secondly, the superpixel color clustering of space of color reference image is processed by the improved Fuzzy ART network. Thirdly, the initial superpixel matching set is established based on the similarity of superpixel features. Additionally, with the adaptive color transfer, an initial colorized low-light image is generated and then fused with an ultra-wide field-of-view infrared image to obtain an initial color-fused image. Finally, based on the similarity between infrared-only information and infrared information with color, color transfer is applied to the infrared-only information, ultimately obtaining a naturally colored fused image. The experiment results indicate that compared with traditional methods, the proposed method improves fusion metrics and color metrics by over 19.5% and 9.4%, respectively. The fused images obtained not only retain the significant feature information of both infrared and low-light images well, but also exhibit natural and continuous colors. 

The technological innovation and transformation in the new-generation internet domain are pivotal to national security, economic stability, and the enhancement of global competitiveness, serving as a crucial support for China in great power competition. However, the internet system and mechanism constitute a "complex giant system" with two major technological approaches to its innovation research: the evolutionary development route and the revolutionary innovation route. Both have their respective advantages and disadvantages and have played a vital role in addressing significant issues in today's internet. In the future, the further development of the new-generation internet should draw on the strengths of various approaches by integrating diverse new architectures, technologies, and capabilities. This integration will transform the internet from a traditional passive information carrier into a new type of proactive intelligent service facility, achieving leapfrog capabilities in intelligence, security, and diversified service provision. It will provide more intelligent and flexible support for emerging businesses and applications in the future, fostering a new landscape of deep integration and collaborative development between internet technology and various vertical industries.

The pendulum waves in blocky rock masses exhibit nonlinear characteristics, making it difficult to interpret their spectral features and the mechanism of quasi-resonance phenomena. To address this issue, a nonlinear vibration model of pendulum wave that considers cubic terms is established. The approximate analytical solution of multi degree of freedom strong nonlinear vibration is obtained by using the improved L-P method and the mode superposition method. This, to a certain extent, explains the nonlinear phenomenon mechanism of pendulum wave in blocky rock mass observed in the experiments. The results show that the vibration frequency can be expressed as high order polynomials of the input energy and changes non-monotonously as the energy increases. The ratios between the resonance frequencies are related to the system structure and are approximately equal to integer multiples of 2. By establishing a hierarchical structure model for pendulum waves, it is concluded that quasi-resonance serves as an indicator of transitions between pendulum waves across different-sized blocks.

Depth information derived from binocular disparity is a critical cue enabling observers to accurately identify camouflaged targets. Given that stereoscopic vision induced by binocular disparity is effective only within a specific range of distances, its applications are primarily focused on enhancing the probability and accuracy of identifying camouflaged targets. The feasibility of applying stereoscopic vision induced by binocular disparity to camouflage design remains to be explored. This study employs random dot stereograms(RDS) to induce stereoscopic vision in observers and collects their electroencephalogram(EEG) signals. The analysis covers the behavioral data of the observers, EEG fatigue indicators, and changes in dynamic functional networks throughout this process. Behavioral data demonstrate that small disparity RDS can induce stereoscopic vision in participants. Significant differences in fatigue indicators were detected at four electrodes, though the trends varied. Behavioral data also indicate that observers did not experience subjective feelings of fatigue. Dynamic functional network analysis reveals that binocular disparity stimuli induced stronger subnetworks in the α and β frequency bands of the brain functional network. This study confirms that small disparity RDS can effectively induce stereoscopic vision in observers,leading them to pay more attention to erroneous target contours without causing noticeable fatigue. This finding provides a physiological basis for employing binocular disparity in camouflage design, enhancing the stereoscopic effect of flat fake targets and camouflage patterns.

Aiming at the issues of underestimating target reflectivity and difficulties in accurately extracting targets structural features in sparse imaging of synthetic aperture radar (SAR), a sparse SAR imaging algorithm based on non-convex and relative total variation (RTV) regularization is proposed. This algorithm utilizes non-convex penalties to suppress bias effects and leverages RTV to adaptively preserve image structures. Subsequently, under the distributed optimization framework of the alternating direction method of multipliers (ADMM), it achieves coordinated optimization enhancement of multiple regularization terms. Additionally, to further enhance imaging efficiency and reduce memory usage, a measurement matrix constructed using match filter (MF) operators is employed for approximate observations, and quantitative evaluations of the reconstructed SAR image quality are conducted. Both simulation and real-data processing results demonstrate that this method can effectively suppress noise and clutter, significantly improving target reconstruction accuracy and radiometric resolution without compromising spatial resolution.

High-mobility scenarios are known to pose a significant challenge to wireless communications systems due to the resulting doubly-dispersive wireless channel. The orthogonal time frequency space (OTFS) modulation is a two-dimensional modulation method that maps the transmitted signal to the delay-Doppler domain. By converting the doubly-dispersive channel in the time-frequency domain into a flat channel in the delay-Doppler domain, the OTFS modulation can effectively overcome the effects of frequency selective fading and time selective fading. To address the channel parameter estimation problem in multi-user massive multi-input multi-output (MIMO) OTFS systems, firstly, through in-depth analysis of the multi-antenna channel structure characteristics, the channel between users and base stations is modeled as a sparse structure model. Afterwards, the massive MIMO channel is divided into multiple groups, and a pilot pattern suitable for multi-user massive MIMO-OTFS systems is designed. A sparse Bayesian learning channel estimation algorithm based on group block structure and common sparse threshold is proposed. Finally, with the estimated channel state information, a method for estimating channel parameters such as fractional Doppler and angle of arrival is designed to further perceive the users' states. The simulation results show that the proposed channel parameter estimation algorithm outperform the traditional methods in estimation accuracy and system spectral efficiency.

In the UAV-assisted railway moving edge computing (MEC) system, an effective method for computing task unloading and power control is proposed with service delay and task failure execution cost as the performance index, while ensuring the stability of energy consumption of UAVs and the requirement of computing delay constraints. With the Lyapunov optimization theory, the optimization problem based on long-term performance index is transformed into the sub-problems of multiple time slots. A dynamic computational unloading algorithm based on domain adaptive learning is proposed to make unloading decisions for computing tasks in each time slot. Meanwhile, in each time slot, the proposed algorithm also optimizes the transmission power of the service data download when the unloading decision determines that the computing task is executed on the local device. The experimental results show that the proposed algorithm can effectively reduce the service delay and improve the efficiency of task completion.

To address the problems of visual burst and feature redundancy in the classical bilinear pooling model in the field of fine-grained image recognition, this paper proposes a graph bilinear pooling model. This model integrates graph networks into the bilinear pooling framework，leveraging the aggregation capabilities of graph networks to encode differential image features into higher-order features，thereby alleviating the phenomenon of visual burst during the encoding process. The results of the experiments conducted on the three public datasets of CUB, Cars and Aircrafts show that the proposed model achieves accuracies of 87.8%, 93.5% and 89.6%, respectively. Compared with decomposed bilinear pooling, this model's parameter count is only 25% of the baseline model, while the recognition accuracy is improved by 2.4%, 1.7%, and 1.3%, respectively, which fully verifies the effectiveness of the model and can provide a method reference for fine-grained recognition of military targets.

Aiming at the problem of high complexity of the existing algorithms for intelligent reflecting surface (IRS)-assisted cell-free massive multiple-input multiple-output (mMIMO) systems, this paper proposes a low-complexity algorithm based on IRS-assisted cell-free mMIMO system combined with beamforming. To begin with, under the constraints of restoring the signal received, IRS phase shift and transmit power, and aiming at obtaining the optimal transmit beamforming vector, a joint optimization problem of access point (AP) active beamforming and IRS passive beamforming is constructed. Then, the original problem is converted into two tractable sub-problems by the alternating optimization method. Finally, the sub-problems are solved by the Lagrange multiplier method and the artificial bee colony algorithm through the idea of solving non-homogeneous linear equations. The simulation results show that compared with traditional algorithms, the proposed algorithm has lower complexity and lower bit error rate, and can use lower resource consumption such as transmission power and spectrum efficiency to obtain performance comparable to other solutions.

In order to solve the small sample problem of insufficient feature quantity in the laser atmospheric transmission field test dataset, a laser atmospheric transmission model was constructed by the multi-layer phase screen algorithm, and more than 4 000 sets of simulation data were obtained. The simulation data, together with 332 sets of near-infrared laser atmospheric transmission field test data, form the machine learning training data set. With the atmospheric parameters, transmission distance, and other parameters as input items, and four evaluation factors for the target spot light field as output items, the random forest algorithm was used to construct a prediction model of four light field evaluation factors to achieve assessment of laser atmospheric transmission efficiency. The effectiveness of the model was verified using 202 sets of real field data. The research results show that the R2 coefficient of the model constructed by the random forest algorithm is above 0.89, which can better fit the multiple regression relationship between input and output; the average relative error between the predicted value and the experimental value is less than 10%, which indicates that the prediction accuracy of the model is high. The research results can provide certain scientific basis for the application of machine learning algorithms in laser transmission effectiveness evaluation, and provide certain technological support for the evaluation of laser atmospheric transmission effectiveness.

The direct current power supply system with pulse constant power load (PCPL-DCPS) has multiple different operating states as the pulse parameters change. However, the current method for large-signal stability analysis of power systems in a single mode is not applicable to the stability analysis of pulse loads with medium and high frequencies. To address this, a unified large-signal stability analysis method that combines switch theory and mixed potential function method is proposed in the paper. In the method, the fast and slow switch stability domains of switch systems are divided by the dwell time of fast and slow switching, and different principles for large signal stability analysis in different domains to meet different dwell time conditions are proposed. Taking PCPL-DCPS as a simulation experiment, the fast slow switch stability domain boundary division and analysis method are verified to be more reasonable. Furthermore, by correlating the stability of switch system with pulse frequency, a more unified and intuitive characterization formula is obtained. The method described in the paper is suitable for switched power systems throughout the entire time period, and can also be applied to the stability research of alternating current power supply systems with pulse loads, further guiding the selection of pulse loads parameters, such as peak power and pulse frequency.

The storage and computing capabilities of a single mobile edge computing (MEC) server are limited, often unable to meet the demands of heavy computation tasks. Therefore, choosing a collaborative approach among multiple MEC nodes to handle migrated tasks becomes an effective measure. To optimize resource allocation in multi-server collaboration, the game theory is employed, and a resource allocation problem based on the game theory for multi-server collaboration is further proposed. In response to this problem, servers are categorized as either migrators or collaborators based on their status. With the game theory, the profit models for buyers/migrators and sellers/collaborators are proposed respectively. By seeking flexible allocation of computing resources, an equilibrium between buyer server profits and seller server profits is achieved. A dynamic payment bidding mechanism for multi-server computing resources is studied, and an on-demand resource allocation algorithm oriented towards server collaboration is proposed. Simulation results show that the proposed algorithm not only meets the computational needs between different servers but also maximizes system performance.

This paper focuses on the optimal fixed-time leader-following consensus of multi-agent systems (MASs). Firstly, based on the goal of performance optimization, an event-triggered optimal control strategy is designed, which takes into account the fixed-time optimal consistency control objectives and limited system communication and computation resources. Secondly, an adaptive dynamic programming （ADP）online learning algorithm is proposed to approximately solve the solution of Hamilton-Jacobi-Bellman (HJB) equation to obtain the expression of the optimal value function，where the Critic neural network structure is only utilized. Thirdly, combined with the gradient descent method and experience replay approach，the weight vector is updated to approximate the cost function and its gradient at the triggering instants by employing the historical record and current data. Finally，the unmanned swarm systems composed of unmanned ground vehicles (UGVs) are utilized to verify the feasibility of the proposed method.

Armored vehicles are the main targets for reconnaissance and strikes on land battlefields. Due to constraints such as long reconnaissance distance and large detection range, armored vehicles in reconnaissance images often present characteristics of low-resolution and missing details, making their identification extremely difficult. Although existing large models based on deep learning have achieved higher accuracy, they lack consideration for model parameters and computational complexity, making it difficult to implement and apply them. In order to improve computational speed, a low-resolution armored vehicle recognition approach via joint super-resolution and knowledge distillation is proposed. By constructing a lightweight model that integrates super-resolution with classification and designing a learning method for lightweight models based on knowledge distillation, the model's parameter quantity and computational volume have been greatly simplified. The proposed lightweight model demonstrates performance similar to that of the larger model.The effectiveness of the proposed model and method were verified on the three public datasets and a self-built armored vehicle recognition dataset.

The current research on railguns is mainly focused on small-caliber railguns, lacking an understanding of the multi-physical field environment characterization of projectiles during the launching process of large-caliber railguns. Therefore, this paper develops a dynamic simulation model for the integrated projectile launching by large-caliber railgun, fully considering practical factors such as contact characteristics, material property changes and structural deformations among components. The spatial and temporal distribution characteristics of the electromagnetic, temperature, and structural fields of the launching components during the launch process are also discussed. Finally, based on the simulation results, the reliability of the projectile is evaluated. The research results indicate that the 130 mm railgun can accelerate a 26 kg projectile to 1 023 m/s, with a maximum overload of approximately 6 700g. In addition, at the moment when the projectile exits the muzzle, the induced current of the fuse and the stress on the TNT powder charge will sharply increase, posing possible safety hazards to the projectile. However, the temperature field has a relatively insignificant impact on the projectile.

To evaluate the lightning detection performance of the FY-4A, this paper proposes a matching method based on the statistical characteristics of the satellite-ground lightning data. First, a larger time-space window is selected for satellite-ground data matching. Then, the time deviation and distance deviation of the matched data are statistically analyzed to determine the final suitable time-space window. By statistically analyzing the summer lightning data (from June to August) in 2020 from FY-4A and ground-based advanced direction and time-of-arrival detecting system (ADTD) in China, it is finally determined that the time interval is 3.4 s and the spatial latitude and longitude interval is 0.6° as the time-space window for satellite-ground data consistency analysis. Based on this time-space window, a comparative analysis of satellite-ground lightning data was carried out. The results show that the average time deviation of satellite-ground data is 1.3 s, and the average space deviation is 23.73 km. The detection efficiency of FY-4A lightning is 19%. It is found that the detection efficiency of FY-4A has obvious diurnal characteristics, with the detection efficiency during the day significantly lower than that at night. The average detection efficiency at night is 31.1%, while during the day it is 7%. The detection efficiency is only 2.3%, the lowest from 12:00 to 14:00 during the day. The diurnal variation of the light radiation intensity of the FY-4A lightning data positively correlates with the change in the solar height angle. The light radiation intensity is the highest from 12:00 to 14:00 during the day, reaching 1 391 μJ/m2/sr, which is an order of magnitude larger than that at night (124 μJ/m2/sr). This shows that only lightning with strong light radiation can be captured by the FY-4A lightning mapping imager (LMI) during the day, which further explains the reason for the low lightning detection efficiency during the day.

In the massive multiple-input multiple-output (MIMO) system, the base stations generally conduct adaptive coding and modulation according to the channel state information (CSI) feedback from the users to improve the spectrum efficiency. Aiming at the problem that the existing CSI feedback methods based on DL cannot be effectively applied to the users' equipment, a lightweight CSI feedback network based on convolutional neural network is proposed. Meanwhile, the depthwise separable convolution is used to reduce the model complexity of the feedback network. Considering the actual deployment of the users' equipment, the multi-task integrated feedback network is designed under different compression ratios and different scenarios. The simulation results are compared with the ConvCsiNet and ShuffleNet feedback networks based on DL in terms of normalized mean square error and network model complexity. The results show that the proposed feedback network can greatly reduce the model complexity of the network during the actual deployment while maintaining a high CSI reconstruction accuracy.

In the traditional deep learning methods, multiple complete models need to be trained and saved independently to achieve the tasks of various text style transfer(TST). These approaches have the following two problems. First, it requires a significant amount of storage space to store different models, which brings significant storage costs. Second, because each model is independent, it is difficult to modularize and reuse them, thereby limiting the reusability and scalability of the models. Inspired by the prefix tuning technology, a style-prefix-guided model for unsupervised style transfer is proposed, which adds a style prefix before the activation layer of the language model and captures the features of the style prefix through an attention mechanism to guide style transfer. By keeping the parameters of the pre-trained language model frozen and only optimizing and storing the parameters of the style prefix, the cost of separately storing multiple language models is avoided. In addition, since multiple tasks share the same language model, the reusability and scalability of the model are improved. The experimental results of the two tasks of common style transfer demonstrate the superiority of the proposed model over the existing baseline models.

In order to study the feasibility of Ti/Al explosive welded laminated composite plate in lightweight design, manufacture and application of equipment, a series of low-velocity impact mechanical property tests of the composite plate were carried out by drop weight testing. The test results show that the composite plate has good impact resistance, and the energy absorption process of laminates can be divided into three stages: elastic deformation energy absorption, plastic deformation energy absorption and damage energy absorption. In the case where the overall thickness is the same, the increase of titanium layer thickness can effectively improve the low-speed impact resistance of the plate. The numerical calculation model of the composite plate under low-velocity impact was established by ABAQUS. Combined with the test results, the accuracy of the finite element model was verified, and the displacement and energy changes of Ti/Al composite plate under different working conditions were analyzed. It can provide reference for the application of composite plates to equipment.

The adversarial examples generated by the existing methods generally suffer from a low attack success rate and are easy to perceive. To address these problems, this paper first designs an audio adversarial example generation framework based on Time-Frequency Partitioned Perturbation (TFPP). Leveraging the time-spectral characteristics of the audio signal, the framework divides the magnitude spectrum of the input audio signal into critical regions and non-critical regions, and generates the corresponding perturbations. Building upon this framework, this paper further proposes a Generative Adversarial Network (GAN)-based adversarial example generation method named TFPPGAN. TFPPGAN takes magnitude spectra as inputs and uses adversarial training to simultaneously optimize the adversarial perturbations in critical and non-critical regions by adaptively adjusting the partitioned perturbation constraint coefficients. Exhaustive comparison experiments are conducted on three typical audio classification datasets. The experimental results show that, compared with baseline methods, TFPPGAN can improve the attack success rate and signal-to-noise ratio by 4.7% and 5.5 dB respectively. The perceptual evaluation score of generated adversarial speech quality also improves by 0.15. Besides, this paper theoretically analyzes the feasibility of the combination of TFPP with other attack methods, and experimentally verify the effectiveness of this combination.