Detection and classification of target at early stage are critical in modern defense systems. Previous studies have primarily focused on binary classification between targets and non-targets, or single-task learning which is structurally limited in detecting and classifying targets simultaneously. This paper proposes a multi-task learning framework based on LSTM that jointly learns prediction and classification tasks, enabling both detection and classification within a single model. The model captures both the temporal patterns of target sequences and decision boundaries between target classes. To improve detection performance, we introduce a two-dimensional score vector that integrates prediction error and k-NN(k-nearest neighbor) distance, followed by Mahalanobis distance calculation. Experimental results show that the proposed anomaly score outperforms conventional methods on target detection. The model achieves high accuracy and macro F1-scores using full-length sequences and maintains reliable performance even with shorter input segments. These results confirm its potential for early stage classification during tracking.

In the aviation industry, early detection of external defects is essential for flight safety and maintenance efficiency. Traditional Walk Around Inspection(WAI), which depends on human vision, is vulnerable to errors caused by lighting, weather, and inspector variability. To address these limitations, this study evaluates the applicability of deep learning-based object detection models for automating defect detection. Specifically, YOLOv5, YOLOv8, and YOLOv9 were trained and tested under identical conditions using two datasets: a public aircraft defect dataset from Roboflow and a custom-built dataset comprising 30 riveted aluminum panels that simulate real-world fuselage defects. Model performance was assessed using Precision, Recall, and mAP@0.5. Among the three, YOLOv9 achieved the highest accuracy across all metrics, followed by YOLOv8 and YOLOv5. These results demonstrate the effectiveness of YOLO-based models for detecting aircraft surface anomalies and support their potential for integration into automated inspection workflows in real maintenance environments.

This paper introduces a Blur Components Extraction Model(BCEM) and presents a synthetic image deblurring dataset specialized for maritime environments, Maritime Blur Dataset(MBD). The proposed BCEM extracts blur kernels from unaligned pairs of sharp and blurred images captured with a single camera, without requiring additional hardware or motion sensors. Using the extracted blur kernels, MBD is constructed by convolving them with high-resolution sharp images of maritime scenes that include ships, buoys, and ocean waves-elements rarely considered in terrestrial benchmark datasets. The proposed MBD is used to train deep learning-based image deblurring models, and their performance is evaluated through both qualitative and quantitative comparisons. By efficiently isolating motion blur components such as engine-induced vibrations, the proposed approach allows for the construction of high-quality and realistic deblurring datasets.

Multi-Function Radar(MFR) systems must concurrently perform search and tracking within strict time and energy limits, rendering resource allocation critical. This study applies Particle Swarm Optimization(PSO)-noted for its simplicity and global search capabilities-to optimize resource management in MFR operations. Three operational presets(SEARCH, BALANCED, and TRACK) were predefined to represent distinct operational modes. Monte Carlo simulations quantified each preset's performance in search coverage, tracking retention, and time-budget compliance. Results demonstrated clear differentiation between presets, indicating that operators can seamlessly shift from search-oriented to tracking-oriented operations by merely adjusting preset parameters without redesigning system architecture. These findings provide a practical framework for enhancing flexibility in both design and operational phases. Future work will incorporate multi-objective optimization and adaptive parameter scheduling to further improve responsiveness to dynamic battlefield environments.

This study proposes a keypoint-based structural anomaly analysis system using Electro-Optical(EO) sensors mounted on unmanned ground vehicles. Unlike conventional object detection approaches, which struggle to interpret fine-grained structural deviations, the proposed method extracts keypoints of components and evaluates their geometric relationships-such as relative height ratios and tilt angles-to assess structural integrity. A three-stage pipeline consisting of object detection, keypoint detection, and post-processing validation is implemented. Experiments under various rotated conditions (0°, 90°, 135°, and 180°) show that the proposed method achieves error rates below 0.5 %, significantly outperforming conventional bounding box-based methods, which show over 50 % error. The system also demonstrates strong robustness against occlusion, viewpoint variation, and partial visibility. These results highlight the system's potential for real-time autonomous diagnostics, predictive maintenance, and quantitative evidence-based monitoring in dynamic field environments.

This study proposes a UAV-mounted LiDAR system for high-precision ground surveillance. The proposed LiDAR system is designed to achieve an angular resolution of 0.01 degree, providing a beam spot diameter of 20 cm at an altitude of 500 m. A compact pulsed laser operating at 1550 nm and a high-performance SSPM were used for eye safety. 3D mapping of ground structures including buildings, trees, and monuments were conducted with the hexa-copter drone. The drone-mounted experiment was conducted at an altitude of 200 meters, and the shapes of complex structures were successfully measured.

In this study, the effects of atmospheric scattering signals caused by laser designation on a laser seeker were investigated. A computational simulation model was developed to evaluate the output characteristics of the laser seeker when scattering signals are incident on it. We conducted experiments on the effects of atmospheric scattering on the seeker and compared them with the model, confirming its suitability for laser seeker performance analysis. In an operational environment where both target and scattering signals coexist, the differences between the two signals were analyzed to access the performance of laser-guided munitions.

It is necessary to reflect the sound propagation characteristics considering the ocean environment to analyze sonar detection performance effectively. However, to analyze the performance of a multiple sonar platform over time in the Modeling & Simulation(M&S) at the engagement level, repetitive sound propagation modelings considering the ocean environment are required. This leads to an increase in computational amount, reducing the time efficiency of sonar performance analysis. Therefore, we propose an operation method for sound propagation modeling based on the spatial variability of sound propagation according to the ocean environment in the East Sea. Instead of dividing the area into equal intervals, regions requiring modeling were divided into unequal intervals according to the spatial variability of the sound propagation. The efficiency of the proposed method was confirmed through comparison with the sonar performance analysis results when the area of interest was divided into high-resolution and low-resolution equal intervals in the same scenario.

This paper presents a novel method to estimate radiation pattern distortion by radome using GPGPU(General Purpose Graphics Processing Unit)-based IPO(Iterative Physical Optics). IPO has been widely used to analyze the electromagnetic characteristics of radome because it can accurately interpret the various effects caused by multiple reflections occurring on the inner wall of the radome. However, the conventional CPU-based IPO consumes huge amount of computation time due to its recursive electromagnetic operations in many meshes. To solve this problem, GPGPU is applied to accelerate IPO by performing the calculation in each mesh in parallel. Simulation results show the proposed method can analyze the radiation pattern distortion efficiently without loss of accuracy.

This study analyzes the effectiveness of ground-based uplink jamming against geostationary orbit(GEO) communication satellites and evaluates potential unintended interference on other satellite systems, including those in low Earth orbit(LEO). The analysis employs the Jamming-to-Signal Ratio(JSR), using a 10 dB threshold to assess communication degradation. Simulations were conducted for a representative GEO satellite, considering multiple ground-based jammer locations. JSR contours were analyzed as functions of transmission power and antenna gain. To examine collateral effects, the beam intersection and exposure time of a selected LEO satellite were evaluated. Results indicate that GEO jamming effectiveness is primarily influenced by transmission parameters rather than jammer placement, and that interference with LEO satellites remains minimal due to their high orbital velocity and short beam-crossing durations. This study provides insight into effective jamming design while minimizing unintentional disruption.

As autonomous technologies continue to advance, their integration into ground weapon systems has accelerated. Among these systems, self-propelled artillery has recently become the subject of research in autonomous operation. Given that self-propelled artillery is a specialized type of tracked vehicle, both path planning and path following algorithms must be carefully tailored to suit its unique mobility characteristics. In this paper, we propose a complete framework for autonomous driving of a self-propelled artillery platform. For path planning, we employ a simplified bicycle kinematic model combined with a cost function that accounts for the platform's specific physical constraints and dimensions. For path following, we design an LQR-based controller using a dynamic bicycle model to ensure accurate trajectory tracking. Additionally, we implement a low-level Model Predictive Control (MPC) module that outputs driving torques for smooth and responsive control. To validate the feasibility of the proposed approach, we conduct experiments in a high-fidelity simulation environment that models both the nonlinear dynamics of tracked vehicles and the effects of uneven terrain. The results demonstrate that our algorithm enables autonomous navigation of a self-propelled artillery system in complex terrain environments.

This study systematically examines the temporal evolution of technology convergence and applicant network structures in the field of South Korean unmanned maritime systems, based on patent data collected from 2011 to 2022. The analysis is conducted across three distinct periods using centrality and network metrics. Findings reveal a progressive deepening of technological convergence, with aircraft (B64U), detection (G01S), and ship structures (B63B) emerging as core technologies. Network visualization illustrates an initial phase of limited interconnectivity, transitioning towards enhanced cluster integration and centralization over time. Applicant analysis indicates that individual applicants dominated during the first period, while technology sharing and collaborative networks expanded from the second period onward, underscoring the growing significance of industry-academia cooperation. The convergence of maritime unmanned systems with unmanned aerial vehicles, autonomous control, and surveillance technologies highlights a shift towards complex, integrated systems. These findings suggest the need for strategic policy support that fosters multi-stakeholder collaboration, promotes open innovation, and encourages knowledge-sharing across sectors. The results offer a foundational basis for future roadmap planning, ecosystem development, and investment prioritization in the evolving maritime unmanned systems domain.

The development of effective torpedo countermeasure tactics is critical for submarine survivability, yet training autonomous agents for this task remains challenging. In this paper, we propose a novel framework that enhances reinforcement learning agents via curriculum learning based on quantified tactical difficulty. First, we quantify tactical difficulty from simulation data, then use grid search to determine an optimal, multi-stage learning schedule. Our model, trained with this curriculum, achieved a final countermeasure effectiveness of 97.94 % in the final performance evaluation, a 3.5 percentage point improvement over a baseline agent trained on a randomized order (94.44 %). Notably, this superior final performance was achieved even when the baseline agent exhibited higher rewards during the training phase. This result indicates that the proposed curriculum is a more effective method for achieving higher final performance than standard randomized training.

This study develops a data-driven framework to optimize battle command training using actual combat data from the Korea Combat Training Center(KCTC). Traditional models rely on theoretical assumptions, limiting their realism. By analyzing KCTC data, this research extracts key parameters to enhance the firepower domain of the Battle Command Training Program(BCTP) model. The proposed framework aligns simulations with real battlefield conditions, improves decision-making, and advances scientifically structured training. Statistical analysis ensures a realistic representation of combat scenarios, reinforcing adaptive learning. The findings contribute to a more effective battle command training system, enhancing operational readiness for future warfare.

In future battlefields, unmanned systems will play a crucial role in collecting visual intelligence, which enables fast and accurate decision-making by operators and commanders. Among various data types, visual information often offers the most intuitive and reliable insights. The level of analysis of such intelligence especially the granularity of object classification can significantly influence tactical and strategic decisions. Fine-Grained Visual Classification(FGVC), which enables model-level identification(e.g. distinguishing a "T-55 tank" from a generic "armored vehicle"), is essential for achieving information superiority. However, the defense industry faces significant challenges in acquiring high-quality training data due to issues of security, scarcity, and limited diversity. To address this, the present study introduces a fine-grained labeled image dataset containing 34 distinct tank models currently in operation by key countries around the Korean Peninsula. The dataset includes over 4,400 images captured under various conditions to support robust training of AI models. We validate its effectiveness by fine-tuning an Inception-v4 model and analyzing performance across multiple pre-processing methods. This work is expected to support the advancement of automatic target recognition, situational awareness, and AI-based combat simulations within MUM-T frameworks, contributing to smarter and more resilient defense systems.

The safety criteria of a helicopter-launched missile is defined in MIL-STD-1289D as rotor disk clearance; 5-degree half angle cone between rotor blade and the outermost surface of the missile. Unlike an unguided rocket, missile's performance is not affected by the initial variation due the helicopter downwash. The safety condition, however, can be quite different according to the missile's initial dynamics. CFD(Computational Fluid Dynamics) is a conventional approach to analyze this kind of safety problem, but it may take several days or even months to solve a single case. For this safety problem, Monte-Carlo simulation method should be applied to consider various error conditions, which means CFD-based analysis is not applicable. In this paper we acquired CFD-based downwash database and applied the data on the missile's aerodynamic database. This method can check the safety conditions of 1,000 cases in a few hours and was validated by the launch test results.

This study proposes an optimization framework for selecting the Minimum Operating Strip(MOS) on wartime-damaged runways using the Harmony Search(HS) algorithm. The MOS selection problem involves multiple constraints, such as minimum runway length, repair resources, allowable damage rate, and maximum repair time. Traditional greedy methods often provide feasible solutions but fail to consider complex constraints or diverse candidate regions. By applying HS, this research demonstrates that MOS candidates can be identified more effectively, balancing repair time reduction and operational feasibility. Simulation experiments confirmed that HS improved repair time by approximately 8 ~ 15 % on average, with higher benefits under complex conditions. The results suggest that HS not only enhances computational efficiency but also strengthens decision-making support for airbase recovery operations. This study highlights the potential of metaheuristic algorithms in military engineering applications.