Aiming at the problems of insufficient samples， semantic ambiguity and blurred entity boundaries in Chinese named entity recognition for UAV communication， this paper proposes a multifeature fusion Chinese named entity recognition model RCPAC （roberta-wwm-ext-large-CNN-PositionAwareAttention-BiLSTM-CRF）. Firstly， the UAV communication domain dictionary is dynamically integrated into the robertawwmextlarge model to obtain global features fused with domainspecific lexical information. A SqueezeandExcitation Networks （SENet） attention mechanism is introduced into the Convolutional Neural Network （CNN） layer to extraction key feature. Within the bidirectional positionaware attention mechanism， two regiondivided forward and backward attention modules are designed to capture entity boundary features. Then， dynamic weightbased fusion is conducted on the global features with lexical information， multiscale local features extracted by the CNN layer， as well as position information and entity features obtained from the bidirectional positionaware attention mechanism. Next， the Bidirectional Long ShortTerm Memory （BiLSTM） layer is used to model longrange sequence dependencies and filter redundant features. Finally， the Conditional Random Fields （CRF） decoding layer outputs the optimal label sequence. Comparative experiments and ablation experiments verify the effectiveness and superiority of the proposed model. On the selfconstructed smallscale UAV communication dataset， the F1score of the RCPAC model is improved by 1.01%-9.95% compared with other baseline models. In addition， the model achieves F1scores of 92.15% and 95.56% on the CCKS2021 Chinese address element parsing dataset and the MSRA dataset respectively， demonstrating its generalization ability and effectiveness.

In counter-unmanned aerial vehicle （UAV） swarm operations，the attack intent recognition（AIR），as the core preceding link in the decision-making chain，directly determines the effectiveness of counter-UAV resource allocation and countermeasure formulation.In response to the problems of existing intention recognition methods，such as difficult modeling，sensitivity to noise，and poor adaptability to non-attack scenarios，an attack intent recognition method based on long short-term memory （LSTM） and interacting multiple model （IMM） is proposed.The multi-classification problem is transformed into a binary classification task through scenario decoupling，and LSTM is used to extract the temporal motion features.The likelihood value of attack intent is calculated based on the additivity of mutually exclusive event probabilities，and the smooth update and normalization of intent probability are realized combined with the IMM mechanism.Multi-scenario simulations and Monte Carlo experiments show that the proposed method improves the recognition accuracy by more than 10% and shortens the convergence time by more than 2s compared with the traditional UKF-IMM method.It can effectively adapt to complex scenarios such as intent switching and non-attack maneuvers，and has stronger noise robustness，providing accurate intent support for air defense decision-making.

The damage expansion of the interface between the coating sleeve and the propellant in the initial operating section of a fully loaded solid rocket motor is analyzed using the LS-DYNA platform.The numerically simulated results show that the pressure difference between the inner and outer surfaces of the coating sleeve is determined only by the minimum flow cross-sectional area between the coating sleeve and the insulation layer without the consideration of fluid-structure interaction，which is inconsistent with the actual physical process.However，when the fluid-structure interaction is taken into account，the pressure difference between the inner and outer of the coating sleeve first increases and then decreases with the increase of the motor operating time.The pressure difference reaches its peak value of 0.604MPa at 0.4s.The additional stresses might exist at the interface between the coating sleeve and the propellant under the effect of this pressure difference，which could lead to the damage and expansion of the interface.During the design of motor，the gap between the coating sleeve and the insulation layer as well as the elastic modulus of coating sleeve should be comprehensively evaluated to ensure the stable interfacial mechanical properties and slower damage propagation，thereby enhancing the structural stability and reliability.

Ground penetrating radar （GPR），as a non-destructive detection technology，is widely used in geological exploration and engineering inspection.The antenna，as a core component，has a direct impact on the detection accuracy and depth of the system.Modern GPR systems impose the stringent requirements on the antennas in terms of miniaturization，ultra-wideband （UWB） and high gain.The traditional antipodal Vivaldi antenna （AVA） often relies on increasing the physical dimensions to extend the low-frequency bandwidth，which limits the portability of the equipment to some extent.Although the existing resistive loading techniques can improve impedance matching，they often come at the cost of significantly reduced gain，leading to insufficient signal-to-noise ratio for detection.To address the aforementioned contradiction，this paper designs a resistively loaded antipodal Vivaldi antenna for GPR.The design introduces structural chamfering at the antenna ends to optimize the current distribution，and incorporates a stepped resistive network to achieve a smooth impedance transition.This approach aims to effectively suppress the end reflections and enhance the radiation efficiency without increasing the antenna aperture.Simulated and measured results show that the antenna achieves a low-frequency cutoff of 1.37GHz，a stable gain over 1.37-26.2GHz band，and a voltage standing wave ratio （VSWR） of less than 2.5 within the frequency band of 1.37-26.2GHz.This study provides key technical support for the realization of high-resolution and lightweight GPR detection systems.

To address the difficulty of balancing the operational resources and spatiotemporal constraints in the cooperative interception against high-speed maneuvering targets，this paper proposes a cooperative interception timing optimization method based on an improved SA-ACO algorithm.Firstly，a bilevel programming model for launch scheduling is constructed：the outer layer determines the minimum number of interceptors based on a preset interception probability threshold，while the inner layer cooperatively optimizes the optimal launch time windows for the interceptor missiles.Secondly，to address the difficulty in the continuous-domain optimization of the inner layer，an adaptive simulated annealing-ant colony optimization （SA-ACO） hybrid algorithm is designed as the solver.The shortcomings of traditional algorithms，which are prone to getting stuck in local optima and have slow convergence are overcome by integrating the dynamic hybrid regulation and experience-inspired strategies.Simulations show that the proposed method quickly outputs the optimal combination of minimum interceptor quantity and launch timing while ensuring a high interception probability.It achieves the significant improvements in convergence speed，optimization accuracy and stability，providing an efficient and robust decision-making scheme for multi-missile cooperative scheduling.

In the process of constructing the ship damage matrix under the conditions of high-dimensional missile terminal parameters，the traditional uniform sampling method has insufficient accuracy due to the limited computing resources.To address this issue，this paper proposes a non-uniform high-precision damage matrix construction method based on the sensitivity analysis of ship damage characteristics.This method first uses the empirical formulas to quickly calculate the probability of damage at discrete explosion points in a ship and obtain the sensitivity distribution of the damage probability space.An then，the sensitive areas within the ship are projected onto the missile terminal parameter space by combining the firing line method and the pixel method to determine the parameter sensitive intervals.On this basis，a probability-weighted non-uniform sampling strategy is adopted to adaptively allocate the sampling resources to the high-sensitive areas while also considering global exploration.Simulation experiments show that，compared with the traditional uniform sampling method，the average prediction deviation mean，median and quartile range of damage matrix constructed by the proposed method are reduced by approximately 50.9%，57.8% and 52.9%，repectively，under the same number of sampling points.The proposed method can effectively improve the accuracy of damage matrix construction under the constraint of limited computing resources.

Infrared small target detection has attracted significant attention due to its strategic value in the key fields such as space-based early warning and maritime rescue.However，the extremely small pixel size，low signal-to-noise ratio，and complex background characteristics make it become a highly challenging visual task.Although the existing deep learning methods significantly outperform the traditional models，the intersection-over-union loss functions commonly used in the existing deep learning methods lack sensitivity to the absolute scale and spatial position variations of predicted targets，resulting in difficulties in achieving pixel-level precise localization and becoming a bottleneck for further performance improvement.To address these issues，this paper proposes an infrared small target detection method based on multi-scale spatial loss and morphological features.Firstly，an absolute-spatial ration （AR） loss function is designed，which enhances the perception of target scaletarget scale by introducing an adaptive dynamic weight based on area differences，and the radial-angular penalty terms in a polar coordinate system are constructed to refine the localization constraints of the center point.Secondly，a lightweight multi-scale prediction head structure is constructed in the U-Net decoder，thereby applying AR loss synchronously to the prediction outputs at different resolution levels to achieve a coarse-to-fine hierarchical supervision.Finally，a dual-stage morphological enhancement strategy for training and testing is constructed，embedding the structural priors via pooling-based differentiable morphological operators during training and correcting the connectivity of the predicted results through opening and closing operations during testing.On the IRSTD1k dataset，the proposed method achieves an intersection-over-union （IoU） of 67.59%，a detection rate （P_d） of 93.02%，and a false alarm rate （F_a） of 9.034×10^-6.Compared to the existing mainstream method DNANet，it improves IoU and Pd by 1.88% and 1.18%，respectively，and reduces the false alarm rate by 48.7%.This achieves a better balance between computational efficiency and detection accuracy.

To address the long-term credit assignment and multi-agent coordinative planning challenges in dynamic multi-aircraft mission planning，this paper proposes a novel reinforcement learning framework based on Value Mix Network （VM-Net）.VM-Net consists of three key components，i.e.，agent-level prediction module （VM-P），value mixing module （VM-M） and geometry-aware dense reward function，working in concert under the centralized training and decentralized execution （CTDE） paradigm.The VM-P integrates an improved gated recurrent unit （I-GRU） with a self-attention mechanism，enabling the agents to leverage the historical trajectories for accurate state evaluation and policy learning.The VM-M aggregates individual value functions into a global Q-function to achieve the explicit cooperation among agents under the CTDE framework.The geometry-aware dense reward function based on relative distance and angle effectively mitigates the sparse reward problem in task planning and accelerates policy convergence.Extensive experiments in both symmetric （up to 12v12） and asymmetric （6v24） scenarios demonstrate that VM-Net outperforms SAC，MAPPO and QMIX.It is still able to maintain real-time inference （<20ms per step） and high win rates （>80%） even in complex environments where baselines degrade significantly.The emergence of sophisticated cooperative behaviors during training—such as coordinated flanking and dynamic target allocation—further validates VM-Net’s practical utility and its potential for deployment in real-world adversarial environments.

The overpressure waveforms are disturbed，missing，or distorted in the field measurements of explosive shock waves，resulting in the failure to extract the feature values of shockwave.Therefore，a deep learning signal restoration and evaluation method that integrates local morphological features with global temporal dependencies is proposed.This method takes the incomplete sequences and binary missing masks as inputs，employs one-dimensional convolution to capture the microsecond-level abrupt changes in the shock wave front，and uses a cascaded bidirectional long short-term memory network with a self-attention mechanism to reconstruct the long-range attenuation trends.To overcome the underestimation of peak caused by traditional regression，the quantile regression and a dynamic peak-weighted loss function are jointly used for optimization，and the temporal generative adversarial networks are employed for sample augmentation to improve the generalization capability.The shockwave data measured during the damage process is processed.The results indicate that the proposed method effectively suppresses peak clipping，significantly reducing the absolute errors of peak overpressure，positive phase duration and specific impulse； at a 90% confidence level，the coverage rate of the predictive interval reaches over 0.94，and the interval width adaptively increases according to the degree of local information loss.The study demonstrates that the proposed method not only faithfully reconstructs the strong transient physical parameters，but also provides the quantitative metrics for restoration quality，offering highly reliable data support for the assessment of weapon damage effectiveness and the interpretation of engineering risk.

The power amplifier （PA） is one of the indispensable components in radio frequency （RF） communication systems. However， its nonlinear distortion characteristics and memory effect can easily affect the quality of the transmitted signal. In addition， with the popularization and application of orthogonal frequency division multiplexing and high-order modulation technology， the high peak-to-average ratio characteristics of signals have higher requirements for the linearity of power amplifiers. Considering that the digital predistortion technology or PAPR reduction technology cannot be only used to maximize the energy efficiency of transmitter， This paper focuses on studying the existing PAPR compression and digital predistortion joint scheme. Then a peak cancellation-crest factor reduction-digital predistortion （PC-CFR-DPD） joint scheme is proposed based on it. The PC-CFR-DPD scheme applies the peak cancellation （PC） technology to the joint structures， and reduces the signal PAPR by generating time-domain cancellation pulses and signal peak point cancellation that are the same as the signal spectrum， greatly reducing the nonlinear distortion caused by direct peak clipping on the signal. The linearization performance of the system is improved without increasing its implementation complexity. The research results has certain reference value for the research and application of digital predistortion technology under high peak-to-average ratio signal conditions.