The target state data in the process of air combat confrontation presents the characteristics of time sequence and multi-dimensionality.In order to further improve the accuracy of target intention recognition, an LSTM target recognition method based on improved attention mechanism is proposed, and the target intention recognition that may occur in air combat is treated as a multi-classification problem.Firstly, the feature sequence is generated by the real-time state data of the target.Then, the attention mechanism is used to improve the feature learning ability of the target, enhance the state feature representation of the main target in the air combat process, and obtain the feature vector with weight allocation.Finally, the LSTM network is used to train the target feature vector, and the target intention is recognized through the softmax layer.The simulation results show that the proposed method effectively enhances the feature learning to the target by using the attention mechanism, and further improves the recognition accuracy of the LSTM network, which is scientific and effective.

This paper mainly studies the active defense guidance method in three-body confrontation scenario. Firstly, an active defense guidance method is derived by constructing a three-body confrontation model with strict feedback form, combined with Backstepping theory and differential game idea.Secondly, an evaluation neural network is established based on the Adaptive Dynamic Programming(ADP) algorithm to solve the guidance method online, and Lyapunov stability theory is used to prove the stability of the closed-loop system and the convergence of the evaluation network weights.Finally, simulation results verify the effectiveness of the proposed active defense guidance method.

An extended state observer based on sliding-mode control method is proposed to solve the stability problem of quadrotor UAVs with disturbances and uncertainties.First, according to the transformation of body coordinate system and ground coordinate system, the dynamic model of UAV is constructed by using backstepping control method.Then, an extended state observer is designed to restore the state of the system and estimate all disturbances and uncertainties of the system.It can achieve rapid error convergence and high enough estimation accuracy.At the same time, the sliding-mode controller is designed.Combined with Lyapunov stability theory and linear matrix inequality method, the stability condition of the observer is proposed.Finally, the validity is verified by numerical simulation.

Aiming at the problems existing in current image dehazing methods, such as dark color of output image, loss of scene information and incomplete dehazing, an end-to-end image dehazing method based on attention mechanism is proposed.Firstly, the channel attention mechanism is embedded into the inception network, and the shallow features are extracted from the fused network.Then, the deep image information is learned by multi-scale convolution and residual dense connection blocks, the depth and shallow feature fusion is realized by skip connection.Finally, it is returned to the pixel scale coefficient matrix through a single convolution layer, and a fog-free image is generated according to the improved atmospheric scattering model.Based on the Mean Square Error (MSE), the fidelity loss function is designed as a constraint in the network model.The experimental results on RESIDE dataset show that the Peak Signal-to-Noise Ratio(PSNR), Structure Similarity (SSIM), Learning Perceptual Image Patch Similarity (LPIPS) and CIEDE2000 of the proposed method reach 32.545, 0.970, 0.026 and 2.711 respectively.The method shows good effect, the output image is dehazed thoroughly and the color fidelity is high;It effectively avoids the loss of detail information in the existing methods.

In public security and other fields, high-resolution thermal infrared image can provide more scene details, and has a wide range of application requirements.However, high equipment cost limits the acquisition of high-resolution infrared images.In this paper, a multistage skip Deep Residual Convolutional Neural Network (DR-CNN) is designed to reconstruct high resolution infrared images by software super-resolution method.Multistage skip dual-channel attention residual blocks are used to increase the convolution depth to solve the problem of lack of correlation between convolution layers.In order to reduce the complexity of training and speed up the operation, Concat module is used to realize the local feature information fusion and the deconvolution layer is used to upsample the feature images directly from low resolution images to high resolution images.Compared with SRCNN, FSRCNN and ADSR, the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) are used as evaluation indexes.Experimental results show that the DR-CNN algorithm proposed is superior to other algorithms, and the generated high-resolution images are rich in detail and clear.

Aiming at the problem that stereo matching algorithm based on Census transform has poor matching accuracy and is susceptible to noise interference, a method combining adaptive window with improved Census transform is proposed.The size of the transform window is selected adaptively, the noise tolerance parameter is introduced into the Census transform to calculate the cost, and the improved Census cost and RGB color difference cost are weighted and fused.Then, the four-path cost aggregation strategy is adopted to complete the initial cost aggregation.Finally, the WTA algorithm is used to calculate the initial disparity, and the final disparity is obtained through post-processing optimization.The experimental results show that the proposed algorithm has low mismatching rate and better anti-noise performance.

In order to solve the problems of local minimum and unreachable target in traditional artificial potential field method, an improved artificial potential field method is proposed.Firstly, the collision range of obstacles is introduced and the influence range of obstacles is defined by using angles to eliminate the influence of obstacles at a certain angle and distance in the forward direction of the robot.Secondly, the repulsion function is improved based on the distance between the target point and the robot to solve the target unreachable problem.Finally, the circumscribed hexagon is used to create a virtual target point outside the collision range of the obstacle to avoid the robot from falling into a local minimum.Experiment results in basic and complex environments show that the improved algorithm can overcome the defects of the traditional algorithm to some extent, and is an effective path planning algorithm.

In order to improve task execution efficiency of muti-agent system and reduce costs, the sliding mode controller is designed based on dynamical model of Akerman unmanned vehicle, which realizes the tracking of the expected trajectory by the follower vehicle.The hyperbolic tangent function is adopted to replace the symbolic function in exponential reaching law for ensuring the continuity of the switch function.The convergence of the system is proved based on the new reaching law.A low-cost unmanned vehicle experiment platform is built based on ROS and Raspberry Pi 4B.ROS software is designed, including navigation, mapping, TF transformation and other functions of unmanned vehicle platform.The networking operation is completed for multi-unmanned vehicle system, so that the sensor data of unmanned vehicles can be shared and the relative position information between unmanned vehicles can be obtained through TF transformation.Finally, an experiment is performed on the unmanned vehicle experiment platform according to the sliding mode controller.The formation control performance curve is obtained and analyzed quantitatively and qualitatively.

Adaptive Optics(AO) imaging system is affected by residual atmospheric turbulence, closed-loop tracking error and photoelectric detection noise, and imaging results are blurred to varying degrees, which is not conducive to the later image screening and post-processing, thus, it is necessary to evaluate the image quality. Traditional image quality assessment methods are not reliable for no-reference AO image quality assessment, and even the assessment results deviate from the actual situation. Aming at the above problems, according to the imaging process of AO system, an AO degradation image dataset with image quality labels is generated. On this basis, a neural network model for assessing the AO image quality is trained by using a deep neural network with ResNet as the backbone, and the best Spearmans Rank Order Correlation Coefficient (SROCC) on the dataset is 0.994. The experimental results show that this method comprehensively considers various degradation factors in the process of AO imaging, a no-reference AO image quality assessment model is obtained by training deep neural network, and the assessment accuracy is better than that of other traditional image quality assessement algorithms.

Aiming at the problem that the variation of ellipse coefficient is not considered in the widely used magnetic target locating method based on magnetic gradient tensor invariant, which leads to the decrease of location accuracy, a new method of solving unit distance vector and distance vector modulus is proposed.By analyzing the position relationship between magnetic target and measurement system, a new unit distance vector is constructed, and a new distance vector modulus with first-order error and second-order error is constructed by Maclaurin formula and magnetic gradient tensor invariant respectively, to avoid the reduction of positioning accuracy caused by the change of elliptic coefficient.Simulation results show that the proposed method has better positioning accuracy than the previous method, and the distance vector modulus with first-order error and second-order error has almost the same effect on positioning accuracy.The method has high research and application value for the detection of submarine and unexploded ordance.

On the basis of standard Ethernet, Time Sensitive Network (TSN) can ensure the real-time and reliability of data transmission by adding a series of key technologies such as high-precision time synchronization, resource reservation and path control, and traffic shaping and service scheduling.However, when heuristic algorithm realizes gating scheduling, the excessive cycle ratio of data leads to the problems of increased cache capacity and uncontrolled scheduling when data with the same priority are transmitted in the same direction at the same time.Therefore, the heuristic algorithm of gating scheduling is optimized and improved, and the problems of insufficient cache capacity and delay jitter are solved by adopting technologies such as data cycle sensing and output cache length sensing cycle.The simulation results show that, compared with the traditional heuristic algorithm, the algorithm can effectively reduce the storage space of the gate control list by 90% when the data cycle ratio exceeds 60 times, and the average delay jitter is reduced by about 50% under the same priority data transmission.

Aiming at the problems of slow convergence and easy to fall into local optimum in three-dimensional path planning of Unmanned Aerial Vehicles (UAVs), this paper propeses an ant colony algorithm with improved artificial potential field, constructs a gravitational potential field, and initializes pheromones in the preliminary searchable area with the resultant force of improved artificial potential field as a coefficient.A random pheromone volatilization factor updating mechanism is proposed, which improves the heuristic function and pheromone updating rules of ant colony algorithm, introduces gravitational potential energy to simulate UAV flying at high altitude, and applies it to pheromone updating.Finally, two groups of comparative experiments are set up to compare the four algorithms.The comparison results show that the improved algorithm can effectively solve the problems existing in the ant colony algorithm, improve the efficiency and ability of the algorithm to search the path, and converge to the optimal value in different environments, which proves the adaptability and effectiveness of the algorithm.

Hyperspectral image has been considered as one of the greatest research directions in the remote sensing science due to its advantages of high spectral resolution as well as allowing for the synchronous acquisition of both images and spectra of objects.Most conventional hyperspectral image classification methods, however, are based on “shallow” handcrafted features, and highly relies on expert knowledge, which are difficult to meet the current technical requirements.In recent years, with the wide application of convolutional neural networks in the field of artificial intelligence, hyperspectral image classification methods based on convolutional neural networks have achieved breakthroughs in classification accuracy and speed.Firstly, hyperspectral image classification methods are introduced, and the limitations of traditional classification methods are analyzed.Secondly, according to the different extraction methods of hyperspectral image features by convolutional neural networks, the algorithm is divided into three types:Spectral features, spatial features and spatial-spectral features, and the merits and demerits are analyzed.Finally, some suggestions are put forward for the insufficient sample training, practical application and classification results of hyperspectral image classification.

Simultaneous Localization and Mapping(SLAM) technology is the main technical support of autonomous mobile robot and has become a research hotspot today.The development history and main sensors of SLAM technology are introduced, then the SLAM technology based on vision, lidar and multi-sensor fusion is sorted out, and the common SLAM algorithms are summarized.The advantages and disadvantages of each implementation scheme are compared and analyzed.Finally, the technical problems and development trends of SLAM are discussed.

Focusing on the decision-making problem of aircraft battle damage repair, an intelligent decision-making support system is designed.Differential evolution algorithm is used to solve the multiple linear regression model of time estimation, which improves the estimation accuracy compared with traditional methods.On this basis, a repair sequence decision-making model with the goal of mininium time is constructed, and the improved adaptive real number coding genetic algorithm is used to solve it, which effectively improves the convergence efficiency.Based on the above model, an auxiliary decision-making system is developed to realize functions such as data management interaction and graphical user interface, which provides support for intelligent decision-making on aircraft battle damage repair.

A design method of attitude controller for quadrotor UAV based on cascaded linear Active Disturbance Rejection Controller (ADRC) is proposed, and the position controller is designed by using Proportional-Integration-Velocity (PIV)control to realize the trajectory tracking control of quadrotor UAV.For the attitude control system, a Linear Extended State Observer (LESO) is designed for pitch, roll and yaw channels to observe and compensate for the internal uncertainties and external disturbances in real time.For the position control system, the PIV control algorithm is used to design the controller and achieve stable tracking control of translation variable.The stability of the cascaded ADRC is analyzed and explained.Finally, trajectory tracking experiments on Quanser flight control experimental platform verify the effectiveness of the proposed method.

For autonomous mobile robot navigation in long-distance complex environments, RatSLAM, which simulates rat brain, is more efficient than traditional SLAM.However, it completely relies on visual information, resulting in its low accuracy and insufficient reliability.By introducing polarization skylight for orientation, which efficiently provides precise absolute heading angles, RatSLAM can be compensated.A polarization skylight bionic rat brain navigation system and an experimental platform are designed.The performance of the system is verified through outdoor road experiments, and is compared with that of the RatSLAM.The experimental results show that the mean position errors of the polarization skylight bionic rat brain navigation system and RatSLAM are respectively 3.65 m and 26.29 m, and the mean angle errors are respectively 0.36° and 2.39°.This system reduces cumulative error and achieves efficient, robust and lightweight autonomous navigation.

In order to solve the problem that it is difficult to determine the minimum sample size for the reliability life test of an accelerometer, based on the reliability statistical inference theory and the life distribution characteristics of mechanical and electrical products, a minimum sample size sampling method suitable for engineering application is proposed, which combines the original data modification, distribution model establishment with sampling method design.It is obtained through this method that the sample size of the accelerometer for reliability life test is 11, which is experimentally verified.The results show that the relative error between the prior parameter value calculated by the sampling method and the original value is less than 1.5%, which can meet the engineering application.

Aiming at the problems of low detection accuracy, low recall rate, and poor real-time performance of remote sensing images, a remote sensing aircraft detection algorithm based on GhostNet and CoT(Contextual Transformer) Multi-Branch Residual Network (MBRNet) is proposed. Learning from the YOLOv4 network model, MBRNet is adopted as new backbone network to reduce the problem of gradient disappearance and makes up for the lack of global feature calculation capabilities of CNN. In order to reduce the problem of small target loss, multi-directional feature extraction and fusion are introduced into the backbone and PANet. The idea is to realize full complementation of information between high and low feature layers and between feature layers of the same scale.The proposed algorithm has an accuracy of 97.64% and a recall rate of 89.11% on RSOD and LEVIR data sets in the circumstance of complex background, overexposure and dense targets.

Integrated Modular Avionics(IMA) system can effectively improve the efficiency of the system and reduce the allocation of resources, but the integration work increases the complexity of the system, and causes the failure of the system to spread to other systems during the integration process, which has a great impact on the system security.Traditional system safety analysis methods mainly rely on engineering experience and personal skills, and have the limitation of ignoring system failure state and misjudging the impact of failure.In order to solve this problem, a new integrated modular avionics system safety analysis method called model checking is proposed, which uses the traversal algorithm to search all the system states.The use of this analysis process realizes automation and reduces the need for engineering experience.The practicability of this method in the safety analysis of integrated modular avionics system is verified.The analysis process is given based on the system requirements and specifications, and the model checking tool NuSMV is adopted for safety analysis.This method can automatically identify the minimum combination of faults that cause top-level events, so as to achieve the purpose of traditional fault tree analysis.Finally, the practicability of the method is demonstrated by the case analysis.

In order to solve the problem that the navigation system filter diverges due to the step mutation of sub-sensor measurement information in the integrated navigation system of AUV during the recovery and docking process, on the basis of unscented Kalman filter, an adaptive fault-tolerant asynchronous fusion algorithm is proposed based on the distributed federated filter structure.According to the sampling rate of the sensor, the algorithm establishes a distributed integrated navigation multi-scale system model.When the auxiliary information changes suddenly, the adaptive extinction factor is used to optimize the UKF, and the gain matrix is dynamically adjusted to weaken the influence of fault information on filtering accuracy, which makes the multi-source navigation system more stable and fault-tolerant.Simulation experiment shows that the proposed adaptive fault-tolerant UKF asynchronous fusion algorithm can better suppress the filter divergence and improve the accuracy and reliability of the navigation system, and improve the stability and fault tolerance of the navigation system in the process of recovery and docking.