With the development of Prognostics and Health Management (PHM) technologies and the increasing intelligence level and informatization level of the device,the knowledge in PHM domain is increasing.Knowledge graph technology attracts attention from scholars in this domain owing to its powerful knowledge organization,management and representation capabilities and its support for related data/knowledge-driven approaches.Oriented to PHM domain,the concepts,key technologies,applications,challenges and prospects of knowledge graphs are reviewed.Firstly,the definition and components of domain knowledge graphs are introduced.Secondly,the methods of domain knowledge graph construction are discussed,and the common construction modes and technologies in the domain are briefly summarized.Then,in view of the characteristics of knowledge in the domain,the applications of knowledge graphs in the domain are introduced in detail.Finally,the challenges and directions of knowledge graph research for the domain are analyzed.The overview aims to help researchers deepen their understanding of knowledge graph and its application in PHM domain,so as to promote further development and innovation of knowledge graphs applied in this domain.

To solve the problem of time-varying quadrotor formation tracking control under the condition of directional topology switching,a consensus control protocol based on special orthogonal group SO(3) under performance constraints is constructed by using the leader-follower strategy and local neighbors'state information.In this protocol,the information of the leader UAV can only be given to the neighbor UAVs,and the set of neighbor UAVs will change.The rotation matrix in the special orthogonal group SO(3) space is used to characterize the attitude of the quadrotor,and an attitude controller is designed based on this to accelerate the convergence of formation control.The experimental results show that when the dwelling time of switching the directional topology is greater than the fixed threshold and the feasibility conditions of time-varying formation tracking control are met,the multi-UAV system can realize the time-varying formation control by using the designed control law,which verifies the effectiveness of the proposed algorithm.

For a fixed-wing UAV system with limited input under disturbance,a reinforcement learning controller based on Actor-Critic is proposed.Firstly,the longitudinal model of fixed-wing UAV under disturbance is established,and the performance function with integral term is designed to evaluate the flight state and system performance of UAV.Then,an Actor-Critic structure based on reinforcement learning is adopted.The Actor is used to solve the control law of minimization strategy performance function,and the Critic is used to approximate the nonlinear performance function.Finally,Disturbance network is used to apply disturbance to the UAV to test the UAV's capability of operating under disturbance,and the momentum gradient descent algorithm is used to improve the learning speed and stability of the neural network to enhance the control performance of the UAV controller.The simulation results show that,compared with the traditional control methods,the proposed reinforcement learning controller can realize the control faster and more stably under the condition of limited input.

A Distributed Sequential Ellipsoidal Intersection (DSEI) fusion algorithm based on the conditional Kalman filtering method is proposed for the state estimation of a multisource asynchronous measurement system with measurement missing phenomena.The system measurement missing phenomena are modeled and described by a set of random variables obeying Bernoulli distribution.The synchronized state space module synchronizes the asynchronous measurement system at the measurement update moment using the state iteration method.The local filtering module is implemented by using the conditional Kalman filtering algorithm to ensure the integrity of the filtering results.In fusion estimation modulea DSEI fusion algorithm is designed to fuse the filtering results provided by each sensor without considering the correlation information among multiple sensors.The simulation example verifies the superior performance of the proposed algorithm with an accuracy improved by 35% and an iteration time less than 3 ms.

An improved YOLOv7 target detection algorithm is proposed to solve such problems as sharp changing of target scales,low detection accuracy and high missing rate of small targets in UAV acquisition scenarios.Firstly,a minimal target detection layer is added on the basis of the original YOLOv7 to adapt to targets at different scales and reduce the missed detection rate of small targets.Secondly,the non-parametric attention mechanism is introduced into the feature fusion network,and an MP-SimAM module is constructed based on the attention mechanism to fuse more important feature information.Finally,a new box regression loss function,named SCIoU Loss,is proposedto further improve the model's convergence speed and detection accuracy.The experimental results show that the model performs well on VisDrone 2019 dataset.The mAP50 of the proposed algorithm model reaches 44.0% on the test set,which is 2.6 percentage points higher than that of the benchmark model YOLOv7.The detection effect of small targets is significantly improved.

Particle Swarm Optimization (PSO) algorithm is widely used in path planning due to its simple principle and easy implementation.To address the problems of the traditional PSO algorithm,such as poor search accuracy and easy to fall into local optimal solutions,an improved PSO algorithm is proposed for 3D path planning of UAVs in urban environments.To improve the search efficiency and accuracy of particles,a chaotic sequence is used to initialize the particle swarm,making the initial particle distribution more uniform.Adaptive segmented inertia weights and adaptive exponential learning factors are introduced to balance the global and local search ability of particles.An acceleration factor is added to the velocity update formula to enhance the ability of particles to leave the poor regions.Adaptive adjustment coefficients are introduced to optimize the particle position update formula.Four test functions are selected for testing and simulation experiments are conducted.The quality of the results obtained by the improved algorithm is better than the quality of those obtained by the Genetic Algorithm (GA) and the traditional PSO algorithm.

To solve the problem of low target detection accuracy caused by closely arranged targetscomplex background information and numerous small targets in remote sensing imagesa remote sensing image target detection algorithm based on the improved ConvNeXt is proposed by using YOLOv5s.Firstlyan improved ConvNeXt Block is introduced at the bottom of the feature extraction network to widen the perceptual field and enrich the semantic information through the interaction between large kernel convolution and selfattention.Secondlya set of bottomup pyramidal structures is added to the part of multiscale feature fusion to amplify the role of shallow feature maps and compensate for the position information of small targets in remote sensing imageswhich is lost due to deep convolution.Finallythe SIoU loss function is introduced to redefine the penalty index and accelerate the convergence of the overall network.The proposed detection algorithm is ablated on the RSOD dataset with a mean Average Precision（mAP） of 92.27%and the experimental results show that the proposed algorithm can realize accurate detection of remote sensing image targets.

To solve the problems of high computational complexity and low detection efficiency of Transformer-based detection model,a lightweight dynamic Transformer object detection algorithm is proposed.Firstly,the dynamic gate strategy is introduced to filter important regions in the self-attention module,and a local-to-global dynamic sparse self-attention mechanism is designed,which enhances the multi-scale generalization capability of the model while reducing the computational load.Secondly,dynamic layer-skipping mechanism is introduced at the structural level of the model.Then,the model is able to adaptively adjust the parameters and structure according to the input during inference to achieve a better tradeoff between detection efficiency and accuracy.The experimental results demonstrate that the improved detection model effectively reduces the computational redundancy,which is more efficient and has a broader practical application space compared with the existing benchmark models.

To solve the trajectory planning problem of dual-arm Free Floating Space Robot (FFSR) in space grasping task,a novel FFSR trajectory planning algorithm is designed,which can realize error convergence within predefined arbitrary time.Cuckoo Search（CS） algorithm is used for parameter optimization to realize fast error convergence of the end actuator and obtain a smooth trajectory.Firstly,the FFSR kinematics model based on attitude errors is derived.Then,the cumulative danger field collision avoidance algorithm is applied to the predefined arbitrary time trajectory planning to obtain a quick collision avoidance trajectory with high tracking accuracy.Finally,the CS algorithm is used to optimize the predefined arbitrary time parameters,and the manipulators'trajectory with low joint angular velocity is obtained.The simulation results show that the fast error convergence of the end actuator can be achieved within a predefined time,and a smoother trajectory can be obtained.

The grayscale changes caused by uneven illumination and sudden changes in illumination affect the detection effect of image features.Therefore,an adaptive FAST-9-12 corner detection algorithm based on grayscale mean value is designed.Firstly,a small-area double-detection template is designed based on the extensibility of feature points,which reduces the number of comparisons between pixels and central points,and improves the region positive detection rate and detection speed.Secondly,based on the local grayscale mean value of the image,the threshold is adaptively adjusted in the detection template of each pixel to avoid the impact of grayscale changes on the detection effect.Finally,the corner radius suppression principle is designed according to the idea of flexible non-maximum suppression so as to screen more robust corners.The experimental results on the dataset of Inria remote sensing images show that the corner detection speed of FAST-9-12 is about 22% higher than that of FAST-12-16 and FAST-9-16 templates,and since the extraction method of adaptive threshold is not easily affected by the illumination,the detection accuracy is improved by 4.16 and 3.11 percentage points respectively.FAST-9-12 realizes rapid and accurate detection of image features.

The development status of optical reconnaissance mission payload technology of large UAVs at home and abroad is introduced.To meet the requirements of ground resolution capability performance index for various types of optical reconnaissance mission payload of large UAVs,an equivalent ground resolution performance index conversion method is proposed.The design idea and layout method of ground resolution test targets are defined,and the intuitive and objective evaluation methods of ground resolution are given,so as to realize the quantitative testing and evaluation of ground resolution capability of large UAVs.The engineering flight test shows that the evaluation method is reasonable,scientific and accurate,and can be used to verify the ground resolution performance index of various types of optical reconnaissance mission payload of large UAVs.

A new carrier landing control method is proposed,that is,integrated direct side force controlled carrier landing.The direct side force is applied to carrier landing research for the first time.The longitudinal control concept of the U.S.Navy MAGIC CARPET carrier landing control is extended to the lateral control.The lateral direct force surface and the conventional lateral surface are comprehensively designed to realize the control decoupling of the three lateral channels.The lateral trajectory incremental control is designed on the direct side force channel,so that the lateral stick displacement is directly proportional to the lateral line-up drift rate.HUD symbology is improved,so that the pilot can observe the situation of carrier landing more directly.A carrier landing flight simulation environment is designed to simulate pilot control in real time.The simulation results show that the pilot can easily complete the lateral line-up correction with the “one stick control” strategy,just like MAGIC CARPET longitudinal control.At the same time,it can well track the lateral deck motion and resist crosswind disturbances,and the heading angle and the roll angle are basically unchanged in the whole process.Compared with conventional carrier landing,this method greatly reduces the pilot's handling burden and improves the landing performance.

To solve the problems of traditional image rain removal methods such as image distortion and artifact generation,a generative adversarial network based single image rain removal method is proposed,which combines convolutional auto-encoding with patch penalty.Firstly,the method uses convolutional auto-encoding to form a generator network,and uses symmetric skip connections to improve the training efficiency and convergence performance of the generator network,and realizes the reconstruction of image detail information and two-dimensional signal spatial information.Secondly,PatchGAN,a Markov discriminator,is introduced to penalize on the level of image patch to remove artifacts in the generated image.Finally,a new refined loss function is proposed to participate in the training of the network model to further enhance the depth of the model's rain removal.The peak signal to noise ratio and structural similarity are taken as the evaluation criteria of the model.The experimental results show that the method has good performance in the rain removal processing of real rain images and synthetic rain images,which can elaborately restore the details of the image and ensure high visual quality.

3D terrain shading technology is one of the important technologies of airborne Synthetic Vision System (SVS),traditional texture mapping requires a large number of texture database storage resources,and the ray-tracing based global illumination model has high requirements for embedded software and hardware resources.To solve this problem,an airborne SVS terrain shading method based on the improved BlinnPhong illumination model is designed,the illumination model is improved,and the vertex color mapping and normal vector calculation algorithms are optimized.The simulation test results show that the improved method has significantly lower storage and memory usage than texture mapping,and the improved method is better than the method before the improvement in terms of CPU usage,GPU usage and display frame rate,and is suitable for the application scenarios of airborne SVS.

To ensure the performance of small moving target detection in dynamic backgrounds based on depth optical flow estimation,fewer times of down sampling are generally adopted to maintain a high resolution,which leads to large computational time consumption.Feature matching is a core processing link of depth optical flow estimation,which takes up a large proportion of the overall time consumption of optical flow estimation,and is very sensitive to the operation times of down sampling.Therefore,a fast optical flow estimation algorithm based on local feature matching is proposed.The target motion information is introduced,the spatial range of feature matching is narrowed,and the amount of data to be processed is reduced.A block-based local matching strategy is designed,and the batch processing mechanism is introduced to avoid the problem of large time consumption in data processing of the pointwise local matching strategythus to accelerate the algorithm.Based on this,CenterNet network is adopted to detect the optical flow anomaly areas corresponding to the moving target in the optical flow field obtained by optical flow estimation.Experimental verification is conducted from the perspectives of optical flow estimation time consumption and detection accuracy.The results show that as for small moving target detection,the block-based feature matching optical flow estimation has about 25% less time consumption than the global feature matching optical flow estimation,while the target detection performance is roughly equivalent.

As for the abnormal behaviors of tri-rotor UAV's actuator,a Two-Stage Kalman Filter (TSKF) based diagnosis algorithm is designed.Firstly,the dynamics of the tri-rotor UAV under its actuator's abnormal behaviors are analyzed and then approximated under the hovering state to obtain the system state space equation.Secondly,the abnormal behavior estimation sub-filter is designed based on the Kalman filter to obtain the optimal estimation of the abnormal behaviors of the UAV's actuator.The coupling equation is introduced to put the optimal estimation into the state estimation sub-filter to obtain the optimal estimation of state errors.The abnormal behaviors are isolated from the system states,so as to diagnose the actuator's abnormal behaviors.Finally,numerical simulations are implemented and the results show that the proposed algorithm can well diagnose the actuator's abnormal behaviors.

As for target recognition in Synthetic Aperture Radar (SAR) images,this paper employs the original images and the target reconstruction results based on attribute scattering centers for decision fusion.The Kernel Sparse Representation-based Classification (KSRC) is taken as the basic classifier to classify the original and reconstructed SAR images.The KSRC improves the classification adaptivity by introducing the kernel function,and target reconstruction can effectively reduce noises in the original SAR images.According to the energy relationship between the reconstruction results and the residual in the target reconstruction process,the noise level of the original SAR images is evaluated.Accordingly,the weights of the original images and the reconstructed images are determined.The weighted fusion method is used to process them and judge the target categories of the test samples.The proposed method is tested based on MSTAR dataset,and the experimental results prove its effectiveness.

To solve the problem of color distortion,such as low contrast and color imbalance in photos and videos taken in an underwater setting,a method of underwater image enhancement is proposed,which combines adaptive color equalization with depth measurement in the underwater scene.The method is composed of the following three steps.Firstly,regarding the issue of severe visual color bias,the preprocessing of adaptive color equalization is conducted on the channels with severe color bias to highlight the texture features.Secondly,according to the images'general texture distribution characteristics,feature extraction is conducted,and background light estimation,scene depth estimation and transmission coefficient mapping estimation are conducted on the images'channels.Finally,the scene brightness recovery model is used for feature fusion,and the histogram equalization algorithm is used to obtain the enhanced underwater images.In order to prove the effectiveness and advancement of the method,real underwater degraded images are selected as the processing object,and the enhanced images are compared in general by using the quality evaluation index of UIQM.The results show that the algorithm effectively enhances the sharpness and chromaticity of the underwater degraded images,and clearly restores the image details.