As the society emphasizes on the life safety of operators and the standard of machinery performance requirements for construction， engineering vehicles are developing in the direction of autonomy， efficiency and reliability. In order to realize the automatic transfer and operation of unmanned engineering vehicles， this paper systematically summarizes the relevant technologies at home and abroad， and analyzes the research progress of key technologies of unmanned engineering vehicles in detail in terms of environment perception， motion planning， engineering operation and condition monitoring， etc. It points out that the technologies of unstructured environment identification， path planning and trajectory tracking of vehicles with variable body structure and automated operation still need to be broken through， and proposes the adoption of mechanism/structure optimization design， advanced communication means， machine learning and digital twin， etc.， which is conducive to promoting the development of key technologies of unmanned engineering vehicles.

According to the algorithm processing method, image dehazing technology can be divided into two categories. One is image enhancement algorithm, which involves histogram equalization, wavelet transform and Retinex algorithm based on color constancy theory. The other is the image restoration and defogging algorithm, which mainly includes the traditional multi-image restoration based on the characteristics of optical polarization and the single image restoration based on a priori assumption, and the emerging image restoration algorithm based on deep learning. In order to better study the image dehazing algorithm in the future, the main development process of image dehazing techniques was reviewed in this paper, the existing problems for the study on image dehazing algorithm was analyzed, and the development trend was tried to explore.

To solve the cross-section deformation defects in the process of rotary draw bending the profile with “日”-shaped into the automobile protective beam, based on the self-developed bending device, the finite element analysis software ABAQUS and the orthogonal test simulation method are used. The comprehensive influences of the process parameters on the cross-section deformation, including bending radius, number of cores and core mold gap are investigated. The results show that the position of the maximum deformation section is affected by the number of cores, each deformation of the section decreases with the increase of the bending radius and the number of cores, first decreases and then increases with the die gaps. When the bending radius is 600 mm, the number of cores is 2, and the gap is 0.75 mm, the effect of reducing cross-section deformation is better, and the validity of the simulation results is verified by experiments.

In this paper, an equivalent modeling method of the cohesive zone weld model (CZWM) is proposed for the simulation of the most dominant tensile-shear loads in automotive body welds. Based on the measurement results of the initial damage and fracture energy from single spot weld joints through tensile-shear and multi-angle tension, an equivalent model of the joint is established. The model is in good agreement with the experimental results of the single spot weld joint in tensile-shear process. At the same time, the effectiveness of the CZWM is verified by three kinds of double weld joints. The results show that the maximum deviation of CZWM in tensile-shear failure load (TSFL) and failure displacement (FD) is 9.3% and 7.1% respectively. When the distribution of welds is parallel to the direction of the external load, the bearing capacity is the highest, the vertical is the lowest, and the diagonal is in the middle. This model provides an option for the numerical simulation of the mechanical behavior in automotive body FSSW welds.

Taking the truck radial tire 295/80R22.5 as the research object, the characteristics of vibration noise were analyzed using modal acoustic transfer vector technique and the acoustic boundary element theory. By means of modal acoustic contribution analysis method, the contribution of each modal shape to tire vibration noise was studied. Based on the this, the carcass chord contour design theory was used to redesign the radial tire. Then, the mechanism of reducing tire vibration noise was analyzed through acoustic contribution analysis. Compared to original designed tire, the redesigned tire reduces 7.48% sidewall mass and 1.98 dB tire vibration noise respectively. This is due to that the adoption of the modified carcass string contour design theory significantly reduces the vibration acceleration response of the tread and sidewall, especially at the peak of 440 and 488 Hz. Furthermore, the redesigned tire reduces noise pressure of positive acoustic contribution (tire tread and sidewall, respectively) and increases the numbers of the negative panels.

This paper first analyzes the mechanism of retraction of boom cylinder, and then puts forward a method to solve the problem of boom cylinder retraction by optimizing the force multiplication coefficient of lifting mechanism based on ADAMS. The kinematic ADAMS simulation model of the working mechanism and the AMEsim simulation model of the working hydraulic system are built. Experiments are carried out to verify the correctness of the simulation model. After constructing the objective function and the constraint function, the new hinge point coordinates of the working mechanism are obtained through the optimization of ADAMS. Simulation and optimization results show that, under the premise of satisfying other working performances of the loader, the lifting mechanism force multiplication coefficient in the unloading condition is increased by 30.8%, and the retraction of the boom cylinder is reduced by 89.3%. This research provides a theoretical basis for the design of the future working mechanism.

In order to optimize the outlet dynamic characteristics of the axial piston pump, a method based on support vector regression machine (SVR) is proposed to optimize the triangular groove of the valve plate. First, the outlet characteristics of the axial piston pump are modeled. The accuracy and feasibility of the simulation model are verified through experiments. The error between the experiment and simulation results is 0.31%, which proves that the theoretical model is in good agreement with the test. Second, the sample data under different conditions of triangular groove structures are obtained through simulation. Based on the SVR model, the corresponding relationship between the outlet flow pulsation and the depth and width angles was found out, and the optimal solutions of the depth and width angles are obtained as 11.2° and 51.7° respectively. Finally, under the same working condition, the optimized calculation results of the triangular groove structure are compared and analyzed. The results show that the flow pulsation after the optimization of the pump is 1.87% lower than that before the optimization, so as to shorten the development cycle and cost of new product.

For reducing the impact of random vibration generated by Tunnel Boring Machine (TBM) on the displacement and stress of hydraulic straight pipe in its hydraulic system, a design model of hydraulic straight pipe under random vibration environment was established, and the correctness of the model was verified by experiments. The influences of pipeline structural parameters on the stress response and displacement response of hydraulic straight pipeline were analyzed. It is found that the stress response and displacement response of hydraulic straight pipe change with pipe diameter and pipe length in the opposite direction. The multi-objective optimization algorithm based on genetic algorithm was adopted to optimize the structural parameters of the fixed-supported hydraulic straight pipeline at both ends under the random vibration environment. The stress response and displacement response of the pipeline before and after the design were compared and analyzed. The results show that the mean square value of the maximum displacement of the pipeline after the design was reduced by about 16.21% and the mean square value of the maximum stress was reduced by 21.04%. The research results could provide a theoretical basis for the design and selection of vibration resistance of hydraulic pipeline in random vibration environment.

The active suspension based on the reference sky-hook model has problems such as nonlinear force, parameters uncertainty, and actuator non-ideality. To weaken the influence of these problems on the control effect, a linear disturbance observer (LDO) was proposed for estimating the uncertainties caused by the actuator imperfections, the nonlinear and uncertain parameters in the active suspension. Combined with sliding mode control (SMC) a control system is established. The influences of the sky-hook damping coefficient and uncertain parameters on the control effect are analyzed. After that, the method for selecting the sky-hook damping coefficient is proposed and the effectiveness of the observer is proved. Finally, the performance indexes of active suspension system with LDO-SMC based on reference skyhook model (LDO-SMCRSM) and SMC based on reference sky-hook model (SMCRSM) are calculated and compared in a bump and class-C road running condition. Simulation results shows that a better suspension performance can be obtained by using the proposed LDO-SMCRSM.

To work out a reasonable and scientific passenger-flow organization scheme with an accurate assembling passenger prediction for transportation terminals, a combination forecasting model based on the KNN regression algorithm was proposed. Grounded on the analysis of the assembling laws of passengers at transportation terminals, the KNN regression algorithm was applied to forecast the number of assembling passengers based on the principles of the numerical similarity and the trend similarity. With the comprehensive consideration of the respective characteristics, the combination forecasting was realized by introducing a time-varying weight coefficient. As a result, the proposed model could solve the shortcomings of the previous KNN regression prediction model, such as large amount of historical data and long running time. The experimental results suggest that the average prediction accuracy of the proposed method can be guaranteed over 95% during non-holidays and 90% during the Spring Festival travel rush, which is superior to moving average method, Kalman filter model and gray prediction method respectively.

Traffic sign detection is the key step of a traffic sign recognition system. To solve the problems of only detection specific categories of traffic signs and over?fitting due to the lack of training samples in deep learning based methods， we propose a deep learning based traffic sign detection method ? pseudo sample regularization Faster R?CNN. In our method， we first design a pseudo sample regularization scheme using the traffic signs and unlabeled background samples from the training set. Then， the region proposal network （RPN） in the deep learning framework is used to generate region proposals. At last， the RPN network and Fast R?CNN detection network are working alternatively and jointly using the alternative training， shared CNN and co?training strategies to obtain the Faster R?CNN traffic sign detection model. We demonstrate the effectiveness of the proposed method through comprehensive experimental analysis. The results show that our method shows boosted traffic sign detection performance and the over?fitting problem can be suppressed。

In order to analyze the preferences heterogeneity of travelers and the influence of autonomous vehicles on travel behavior, this study incorporates the latent psychological variables into the latent-class conditional Logit model based on the theoretical framework of the extended technology acceptance model. A hybrid choice model is established to conduct empirical research. The results show that compared with the traditional conditional Logit model, the latent-class conditional Logit model has higher fitness. Respondents can be divided into three subgroups: the shared autonomous vehicles preference subgroup, car preference subgroup, and private autonomous vehicles preference subgroup. The three subgroups account for 46.5%, 13.0% and 40.5%, respectively. Travelers from the car preference subgroup evaluate the walking and waiting time and travel costs positively, while travelers from the private autonomous vehicle preference subgroup evaluate parking costs negatively. The two latent variables of perceived ease of use and perceived trust have a significant influence on the classification of latent class.

In order to study the influences of Vertical Vibration Test Method (VTM) and Marshall compaction on the mechanical properties and void distribution of asphalt mixture, AC-25 asphalt mixture samples were prepared by VTM and the Marshall method respectively. The influences of molding method, vibration time and compaction time on the volume parameters and mechanical properties of AC-25 asphalt mixture were studied through VV, VMA, VFA, Marshall stability, compressive strength, split strength and shear strength. The correlation between the volume and performance indexes obtained by two indoor molding methods and the actual pavement was analyzed. Finally, the influences of molding mode, vibration time and compaction time on the grading of molding samples were analyzed by CT scanning technology. The study shows that the AC-25 samples formed by VTM have better mechanical properties and higher correlation with the real road surface. Moreover, the pore structure and mechanical properties of core samples show exponential distribution. The constitutive relationship between the pore structure and mechanical properties of AC-25 asphalt mixture can be described in the form of exponential function.

To study the performance evolution of asphalt mixture under the impact of multiple pore water pressure cycles, three kinds of SBS modified asphalt mixture specimens with different asphalt content (3.5%, 4.5% and 5%) were prepared. Different cycles of pore water pressure (500 to 7000 cycles) were applied on specimens by Moisture Induced Sensitivity Tester (MIST), and then the change of internal structure of asphalt mixture is characterized by the physical indicators (volume, bulk relative density) of the moisture conditioned specimens. Indirect Tensile Strength (ITS) and Tensile Strength Ratio (TSR) under different times of pore water pressure measured by Indirect Tensile Test (IDT) are used to characterize the evolution of mechanical properties of asphalt mixture. The results show that the volume of the specimens increases first and then decreases, the Bulk Relative Density (BRD) slightly decreases with the pore water pressure cycles. ITS and TSR results show that the mechanical properties decrease first and then improve. It is found that 500 pore water pressure cycles are the critical times for specimens to change from unsaturated state to saturated state. When cycles are less than 500, the specimens are unsaturated and the pore water pressure produced by MIST is large, so the mechanical properties are continuously reduced. When cycles are higher than 500, the specimens are saturated, MIST cycle has little contribution to pore water pressure and the effect of temperature becomes the main impact, resulting in the improvement of mechanical properties. Thus, for the small dense asphalt mixture specimen with porosity of about 4%, it is recommended to use 500 MIST cycles as moisture condition to evaluate the water damage resistance of asphalt mixture.

In order to investigate the relationship between gas permeability and meso-structure of carbonated concrete, carbonation tests and gas permeability tests of ordinary concrete (OC) and polypropylene basalt fiber concrete (BP) with 0, 7, 14, 28, 56 and 80 days carbonation were performed respectively. The parameters of concrete meso-structure were obtained by RapidAir 457 concrete pore structure analyzer (Denmark). Based on the influence levels of different pore size ranges and calculation methods on the air content and fractal dimension of concrete, the pore size of concrete is divided into 0-160 μm, 160-300 μm, 300-4000 μm. The air content and fractal dimension were calculated respectively. The influences of carbonation age and materials on the macroscopic and microscopic properties were analyzed. The internal relationship among the gas permeability, fractal dimension and meso-structure of concrete was studied. Then, the sensitive pore size range was given. The results show that the gas diffusion coefficient and air content of BP are less than that of OC in general, which shows that hybrid can improve the carbonation resistance of concrete by improving the balance of three-dimensional distribution of fiber from different levels. With carbonization age increasing, for OC, the correlation between pore mass fractal dimension in small pore size range and gas diffusion coefficient is higher than other ranges. For BP, the correlation between solid mass fractal dimension in medium pore size range and gas diffusion coefficient is higher than other ranges. The solid mass fractal dimension is employed to characterize the concrete internal structure more accurately than pore mass fractal dimension. Based on the comprehensive analysis of air permeability, fractal dimension and meso-structure, it is identified that the sensitive pore size range of OC is the small pore size rang, and the sensitive pore size range of BP is the medium pore size range.

To reasonably analyze the failure probability of the long-span bridge girder, considering the correlation among the failure modes of the multiple control monitoring points, a new data fusion method about the failure probability analysis for the long-span bridge girder is presented. With the extreme strain information, the optimal R-Vine copula model considering the correlation among the failure modes of the multiple control monitoring points is built with the corresponding performance functions, bivariate pair-copula model and the optimal R-vine. Further, with the first order second moment (FOSM) method, the failure probability of the long-span bridge girder considering the correlation among the failure modes is analyzed , the feasibility of which is compared with the other analysis method using the monitoring data of the existing bridge. The results show that the optimal R-Vine copula information fusion method for the failure probability analysis of long-span bridge girder considering the correlation among failure modes is more reasonable.

In order to study the restoring force model of Scarp Tire Rubber Pads (STPs) under the thermal oxidation aging, cyclic loading tests with different aging time were performed and the hysteretic dissipated energy characteristics of STPs were analyzed. Based on the double spring model, and combined with the results of STP pseudo-static test, this paper presented a restoring force model for STPs and established a two-spring restoring force model for STPs with different aging time and design compressive stress. By comparison of MATLAB simulation of hysteretic performance and experimental hysteretic performance, it is verified that the model could well reflect the energy dissipation performance of STPs. The results show that, with the increase of aging time, the horizontal equivalent stiffness, unit cycle energy dissipation area and equivalent damping ratio of STPs all increase first and then decrease, reaching the maximum in actual aging of 50 a, which can meet the isolation energy consumption of rural buildings within 50 a. The establish restoring force model under the condition of hot oxygen aging can better simulate the hysteretic characteristics of STPs under different aging time and different design compressive stress, which can provide theoretical reference for the application of STPs in isolated buildings.

In this paper, 22 groups of center pull-out tests of reactive powder concrete were carried out. The effects of concrete compressive strength, steel fiber volume content, steel bar diameter and bond length between steel bar and reactive powder concrete (RPC) on bond performance between deformed steel bar and RC RPC are analyzed. The occurrence conditions of three failure modes, such as steel bar pull-out failure, concrete splitting failure and simultaneous occurrence of steel bar pull-out and concrete splitting failure, are summarized. Combined with the test data, the calculation formula of the characteristic value of each bond anchorage under RC curing mode is obtained, and the average bond stress-slip constitutive model of RC RPC reinforcement is established., This model is verified by the test results, and the effect is good. Through the pull-out test of the center of the strain gauge attached to the steel bar, the distribution law of bond stress is analyzed, and the bond position function is obtained by fitting. The test results show that the bond strength between deformed steel bars and RPC increases with the increase of compressive strength of concrete. With the increase of steel fiber content, τ_u and τ_r tend to increase. With the increase of rebar diameter, τ₀ decreases first and then increases, while τ_u and τ_r decrease. With the increase of bond length of steel bars, τ₀ and τ_r increase, while τ_u decreases. With the increase of load, the peak value of bond stress moves from the loading end to the free end. The longer the bond length, the more uneven the bond stress distribution. The formula obtained by combining the average bond stress-slip constitutive model with the paste position function can fully reflect the bond stress-slip constitutive model of RC RPC and deformed steel bars.

To clarify the speed behavior on typical scenes and speed variation when driving scenes switching, three road sections were selected in Chongqing, field driving experiments with more than 70 subjects were carries out. The vehicle operation data under the natural state were collected through on-board instruments, and the speed change characteristics when driving scenes alternate was analyzed, as well as the restraint of roadway condition on the speed behavior for various driving scenes. The results show that the driver's selection behavior on the main line of river-spanning bridge is highly discrete, that is, the road environment has little restraint on the driver's behavior, and the probability of rear end collision is higher on the bridge. When driving from the main line of the bridge into a ramp with small radius of the interchange, the driver will slow down on the bridge in advance, and the sharp ramps have stronger binding on the driving behavior than the bridge. The driver will continue the deceleration behavior to a certain distance in the tunnel, and the dispersion of speed amplitude will be weakened at the same time, indicating that the change of environment at tunnel entrance will significantly affect the driver's speed choice behavior. When driving into a ramp with a small radius, the driver's deceleration behavior will continue into the range of circular curve, and higher speed will lead to greater deceleration. For a ramp with a relative longer circular, only a few drivers will keep a constant speed near the middle of the circular curve. When the circular curve of the ramp is shorter, no driver would drive at a constant speed within the circular curve. The driving speed is constant in the range of multi-layer helical ramps. Speed amplitude of 70 percent drivers is distributed in a narrow range. Therefore, adjusting the design parameters in a small range can take care of most drivers' behavior habits.

In order to improve the measurement accuracy of three-dimensional (3D) texture of asphalt pavement, the traditional binocular reconstruction algorithm was improved in threefold. First, the laser line constraint was introduced. Second, the improved binocular reconstruction test system and the 3D texture measurement precision evaluation device were produced. Finally, the regional segmentation matching algorithm was established. The results show that the introduction of laser line constraints can improve the accuracy of both overall measurement and single point measurement. Moreover, the measurement accuracy of the improved algorithm can be improved with the increase in the number of laser line constraints. Using six laser constraints, the improved algorithm can satisfy the precise measurement of pavement 3D texture. Additionally, the improved algorithm has good anti-interference ability to light, and can keep stable in the illumination range of 50~ 350 LUX.

Cloud operators charge bandwidth fees for data from user jobs that are transferred between data centers, and this accounts for a large percentage of user spending. It is of practical significance for users to reduce the communication overhead of the submitted job significantly without affecting the completion of the job through reasonable assignment of tasks and resources when determining the leased resources. This paper analyzes and models the traffic cost problem across data centers when workflow tasks are distributed in geographically distributed data centers, and proposes a heuristic algorithm for solving the problem. Several sets of experiments are carried out on the proposed algorithm, and the results are compared with that of the resource allocation algorithms commonly used in the current practice.. The experimental results verify that the heuristic algorithm proposed in this paper can effectively reduce the user communication traffic cost by at least 9.75% in the workflow operation across the geo-distributed data center, which serves to reduce the communication fee between users across geo-distributed data centers.

In the process of constructing the knowledge graph, the existing method ignores the processing of semi-structured data, which leads to the inaccuracy and time-consuming in construction of the knowledge graph. Therefore, an automatic knowledge graph construction algorithm based on massive text data is proposed. A triplet extractor is used to extract massive text data sources, and to extract semi-structured data, while eliminating redundant data. According to the data processing results, the appropriate data objects are selected using the data collection function as the text data source constructed by the knowledge map. The data source is subjected to standardized processing such as text format conversion, word segmentation and feature extraction. The underlying semantics of the data are analyzed and an XTM visualization map is drawn to form a preliminary knowledge map. The triples of users, ratings and items are composed by mining the existing knowledge in this knowledge map, applying potential vectors to information recommendation, and the graph evolution algorithm is used to predict the ratings, users and items, constructing latent vector models Domain recommendation to realize the automatic evolution of the knowledge graph. Experimental results show that the algorithm has higher construction accuracy and less time consumption, which shows that the algorithm is reliable and practical.

With the development of cloud storage technology, more and more users store data to the cloud server in order to reduce the cost, but at the same time they lose the right to control the data, which leads to the fact that the integrity and security of cloud storage data cannot be ensured. To solve these problems, this paper proposes a security data integrity audit scheme based on fast tag query. Firstly, chaotic system is used to encrypt cloud storage data to ensure data confidentiality. Then, the Hash Value of the data block is used as the label to classify data blocks. In this paper, each leaf node of Merkle Hash Tree (MHT) corresponds to a label, and each label corresponds to multiple data blocks, which increases the speed of data query index. This scheme not only meets the security requirements of the cloud data storage integrity audit scheme, but also reduces the communication expenses of the audit process. The experimental results suggests that the scheme is safe and effective.

In the current video monitoring dynamic target detection process, the pre-estimation of the target position is ignored, which leads to a long detection time and a large detection error. To solve this problem, a dynamic target detection method for video surveillance based on motion vector space coding is proposed. Motion vector space coding method is applied for background modeling, CamShift target tracking algorithm based on kalman filter is used to detect the target, and estimate the search scope and the location of the target at the next appearing moment. CamShift is combined with estimation results to search the true location, correct the target search scope, acceleration and speed, thus, completing the video monitoring dynamic target detection. The experimental results show that the proposed method has high efficiency in detecting dynamic targets and high accuracy in detecting results.

When encryption or compression algorithms are used to transmit data over the network, the payload data is generally random. Using existing traffic detection methods, it is difficult to effectively distinguish encrypted traffic from compressed traffic. To solve this problem, based on the differences between randomness of encrypted data and compressed data, this paper proposes the ECF randomness feature set. Without relying on the information of the network protocols, the packet headers, and the compression identifiers, the current mainstream machine learning algorithms are used to achieve accurate identification of encrypted or compressed data. Experiment results show that this method has higher accuracy compared with current methods and it also has good performance with generalization and migration.

In this paper, attention mechanism is combined with deep reinforcement learning, and label information is used to study how to learn effective Korean language text structured representation independently. Two structured representation models are proposed, which are called Information Distilled Attention (IDA) and Hierarchically Structured Attention (HSA). IDA selects the important words related to the task, and HSA finds the phrase structure in the sentence. The structural discovery in both presentation models is a sequential decision problem that can be implemented using Policy Gradient (PG) in reinforcement learning. The experimental results show that the proposed IDA can recognize the important words of Korean, and HSA can extract the sentence structure well, and have good performance in the task of text classification. At the same time, the results of the two models have a good auxiliary effect on corpus tagging.

Due to the absorption and scattering of light by suspended particles in water, the original underwater image has a low definition, fuzzy details, and color distortion. To solve these problems, an underwater image recovery method based on multi-scale attention fusion and Convolutional Neural Network (CNN) is proposed. First,the Space Channel (SC) module is constructed by using the attention mechanism. Then, by adding the SC module into the multi-scale feature extraction, the information in the image can be extracted effectively, and the image sharpness and color correction can be realized. Finally, the Laplacian operator is used to construct the multiple loss functions to further enhance the detail features of the image, so that the recovered image quality can be significantly improved. The method in this paper is compared with other methods qualitatively and quantitatively on the two test sets. The experimental results show that this method is superior to other methods in image sharpness, detail enhancement, and color correction.

An online learning prediction model for fuel injection based on a large amount of experimental data is designed by utilizing back propagation neural network (BPNN), and a real-time torque tracking controller consisting of the BPNN predictive feedforward and a PID feedback is proposed. A simplified and discretized real-time model is adopted in the BPNN prediction model. The threshold value of the BPNN prediction model with parameter adaptability can be learned and updated online. Several experimental results show that the BPNN prediction model with online variable threshold proposed in this paper has higher prediction accuracy comparing with the model with fixed threshold, and that the torque tracking controller proposed in this paper can satisfy the real-time control requirement and has smaller error rate under the transient condition comparing with the PID controller.

The existing methods need a lot of labeled data as support in the detection, but the lack of data in the case of cell adhesion and occlusion is not conducive to the improvement of cell detection accuracy. In order to solve this problem, a cell image generation detection method based on condition generation antagonism network is proposed. Pix2pix network model is used to control the generation of cell image with adhesion occlusion, the loss function is extracted, pix2pix is used to realize image to image conversion, and regular term error control is used to generate network error. On this basis, the detection network is constructed, including the generation network structure, discrimination network structure and detection network structure. The target detection is carried out at the output of the generation network, so that the image generation and cell detection are completed in the same network. Experiments show that compared with the existing model, the designed method has a significant improvement in detection accuracy, reaching 90.2%, which can meet the needs of medical cell detection.

In order to make full use of the information in the data distribution of the Satisfiability Problem (SAT) and thereby improve the performance of the algorithm, this paper proposes a stochastic local search heuristic method based on deep reinforcement learning. The selection of the variables in stochastic local search algorithm is regarded as a reinforcement learning task. The strategy agent learns is used as a heuristic for selecting variables, so that we can obtain a more efficient variable selection heuristic in an end-to-end way. Experiments show that the method in this paper is effective. Compared with the classical stochastic local search algorithm ProbSAT, our method also has certain advantages in performance, and can solve the problems in fewer decision steps.

The task of object detection plays an important role in the safe driving of driverless vehicles. Currently, the object detection technology of environment percept is mostly one-class object detection or all the objects in an image are listed as the target to be detected. Numerous studies have not yet focused on object division and detection of the objects in front of the vehicle. To solve the above problems, in this paper, the objects to be detected in front of vehicles are divided into two categories. One is the dynamic targets with high risk and displacement at any time, including four-wheel vehicle, two-wheel vehicle and people. The other one is the static targets with less danger and no displacement, including traffic lights and traffic signs. For the dynamic multiple objects in front of the vehicle, an improved algorithm of object detection based on YOLOv3 is proposed, which can be transplanted to the embedded system. To overcome the shortcoming of the original YOLOv3 algorithm, that it is difficult to get real-time detection in the embedded terminal, the original backbone network Darknet-53 was replaced with MobileNetV2 to extract features, adding Group Normalization operation in the training process and using Adam as optimizer. The extracted BDD100K dataset is used for training. The model is tested with BDD100k partial dataset not involved in training and Team_test dataset produced by our research group. The results show that compared with original YOLOv3, the missing rate (MR) of the algorithm in this paper can be kept within 5%, and based on the increase of 0.020 in mAP, comparing with the basic model of YOLOv3, the parameters of YOLOv3-MobileNetV2 model are reduced by about 89%, the Inference Time is reduced by about 70% under the CPU.

In view of the disadvantages of the existing simultaneous localization and mapping (SLAM) algorithm, which has low real-time performance and the positioning accuracy is greatly reduced in dynamic environment, a DG-SLAM algorithm based on deep learning and GPU parallel computing was proposed. The deep learning-based object detection algorithm was introduced to detect dynamic objects in the driving environment, and the feature points of dynamic objects were removed before the matching of image frames, so as to eliminate the impact of mismatching dynamic objects on the positioning accuracy of SLAM system. In the tracking of local maps of ORB-SLAM2, the discriminant method of 3D interior points was used to distinguish the inner points and outer points, and the GPU parallel computing model was established to efficiently search the local map points. The Saturated kernel function was used to minimize the reprojection error of the two norm terms to ensure the parallel calculation of the reprojection error when the local map was optimizing. The algorithm was verified on the KITII dataset, DG-SLAM has high tracking accuracy, and the average calculation efficiency was more than 3.4 times faster than that of ORB-SLAM2 system under the same conditions, more than 85 frames per second, which could realize efficient and high precision SLAM in dynamic scene.

A new fuzzy c-harmonic means clustering algorithm, which is based on cluster volumes constraint, is proposed in this paper to solve the problem of imperfect clustering performance of traditional algorithm for imbalanced data set. Firstly, a quantity is defined by the membership matrix to measure the volume of each cluster, which is then used to construct a new objective function by combining with that of traditional algorithm. Secondly, new membership matrix and cluster center formulas are obtained by minimizing this new objective function. The proposed algorithm was tested on the UCI data sets, simulated imbalanced data sets and actual machine vibration detection imbalanced data sets. Experimental results show that, compared with several peer algorithms, the proposed algorithm achieved good clustering performance for imbalanced data sets while maintaining the global optimal performance of the traditional one.

The artificially extracted features are computationally intensive and subjective, fail to fully reflect the nature of the signal, and take too long to generate time-frequency images. To overcome these problems, We propose an improved residual neural network (ResNet) ResNet32 as a framework to extract and identify radar time-domain signal features. We build a time-domain signal dataset of 9 types of intra-pulse signals and input them into the ResNet32 framework for training and classification. The algorithm saves a lot of time to generate time-frequency images, and the experimental verification algorithm has a better recognition rate at low signal to noise ratio (SNR). In the experimental conditions of mixed SNR, the recognition rate of the 9 modulation types with SNR=-14 dB and SNR=-8 dB achieved more than 90%.

In this paper, the error transfer model of three-dimensional ultrasonic positioning (TDUP) system in air medium is provided. The distribution of error sensitivity under different baseline length is obtained through experiments. The positioning accuracy of the TDUP system mainly depends on the distance measurement error and the error transmission brought by the process of calculation. This paper is focused on the study of the error transition. Firstly, the partial differential equations of the positioning coordinates with respect to the ultrasonic propagation distance are established, and the error transfer model, the expression of the error sensitivity ε, is derived by the total differential equation. According to the expression of ε, the relative length of the receiving array baseline and the direction of the distance error symbol (differential mode distance error and common mode distance error) directly affect the error sensitivity. The spatial distribution of error sensitivity is given by experiment. Next, under the condition of long baseline and short baseline, the influences of common mode and differential mode distance errors on the error sensitivity are discussed, and the spatial distribution of error sensitivity is given through experiments. The experimental results show that the positioning error in the long baseline ultrasound positioning system is the result of the combination of differential mode and common mode distance error. The positioning error in the short baseline ultrasonic positioning system mainly comes from the differential mode distance error, and the common mode ranging error could be ignored. In addition, the error sensitivity of short baseline system is much higher than that of long baseline system under the condition of difference mode. T The contour of sensitivity error is distributed layer by layer in the form of Russian dolls. The spatial distribution of error sensitivity further explains the uneven distribution of errors in ultrasonic localization. Finally, the paper proposes solutions to improve the positioning accuracy.

When the current algorithm is used for communication scheduling between cross-platform processes, there are problems of high packet loss rate and high average delay. To solve the above problems, a cross-platform inter-process communication scheduling algorithm based on weighted queue is proposed. By analyzing the two indexes of packet rate and end-to-end delay, it is concluded that in order to achieve cross-platform process communication scheduling, we need to pay attention to queue length and data retransmission hops, and initially reduce the packet rate and average delay in the communication process. According to the random number and the service probability, the queue manager, the packet length observer, the adaptive service probability calculator and the queue scheduler are used to realize the weighted fair scheduling among the cross-platform process communication, and the average packet length of each queue is updated to further reduce the probability of packet loss in the platform. Experimental results show that the proposed algorithm has a low packet loss rate and low average delay.

Although the existing method of rotation of each sensor can improve the matching accuracy of the drift angle,it results in the discontinuity of the image synthesized by the detector with optical butting on the focal plane. In order to solve the above problems, an electronic imaging system based on CMOS image sensor GMAX3265 and FPGA is designed. The system uses the windowing mode of CMOS image sensor to Real-time clipping the nonlinear part of the image synthesized by adjacent detectors, thus forming a continuous image strip. Firstly, the difference between the traditional drift angle compensation method and the slice rotation drift angle compensation method is compared and analyzed when the detector adopts optical butting. Secondly, the number of cutting rows is calculated and the cutting precision is analyzed. Finally, the electronic system is completed through the overall system design and driving timing design. The experimental results show that when the rotation angle difference between adjacent detectors is 0.1 °, the ratio of error pixels to total pixels is 0.1165%, which does not lose too much image and improves image continuity.

In the estimation of the motion state of non-cooperative targets such as aircraft and satellites, due to changes in the features detected by the visual system, the system estimation error will increase or even diverge. This paper proposes a strong tracking volume Kalman filter algorithm (STCKF). When using the CW equation to describe the relative motion of the satellite, considering the eccentric installation of the visual measurement system, a suboptimal fading factor is introduced in the covariance matrix of the STCKF prediction state error, and the gain is adjusted online to ensure that the residual sequences are orthogonal to each other, which guarantee the reliability and stability of the system tracking when the state changes suddenly. The simulation results show that, compared with the standard UKF and CKF algorithms, STCKF can adapt to the time-varying characteristics of the target features, and significantly improve the accuracy and stability of target tracking.

In recent years, the number of electric vehicles has been drastically increasing. This situation gives rise to an urgent demand for designing a real-time optimal power allocation inside electric-vehicle charging stations. The goal of this paper is to present one possibility that addresses the aforementioned demand. First, the paper formulates the power-allocation problem as a linearly constrained non-linear optimization where the objective function is constructed via logarithm utility functions. Second, the paper mathematically shows that the optimal solution to the formulated power-allocation problem is user-friendly. Third, the paper presents an efficient algorithm that is able to obtain the optimal solution in real time. Finally, the paper validates the performance of the proposed power allocation through numerical simulations.

In order to realize the self-adaptive adjustment of the motion speed of mobile robot in different terrain environments, a method of terrain clustering analysis based on robot vibration information and self-adaptive adjustment of motion speed based on clustering results is proposed. In this method, the acceleration information in the vertical direction and pitch angle information of the robot in the process of motion are used, and the terrain information is clustered based on the improved Gaussian mixture model to obtain the membership degree of the terrain relative to the typical terrain. In combination with the terrain undulation information, the adaptive control of the robot's motion speed is realized through the fuzzy control strategy. In order to verify the correctness and practicability of the proposed method, relevant experiments are carried out on the platform of pioneer 3-At four-wheel drive all terrain robot. The experimental results show that this method can make the robot cluster analysis on different terrain accurately and realize its speed adaptive adjustment under different terrain, which effectively enhances the terrain adaptability of the robot.

Aiming at the control problems when load-carrying quadruped walking vehicle weight changes, a control method based on centroidal dynamics and adaptive sliding mode control method was proposed. The control of the walking vehicle was divided into two parts: the control of the torso and the control of swinging legs. The centroidal dynamics and task space PD control method were applied to the torso motion control, and the virtual model control method was applied to the motion control of the swinging legs. Then, the adaptive sliding mode control algorithm was applied in the height direction of the vehicle to realize the adaptation to weight changes and the weight recognition. The tracking accuracy of the forward speed and lateral speed was improved by combining the centroidal dynamics. Adams and Simulink were used to simulate the trotting gait on flat ground and slope when the weight of walking vehicle changes, and comparisons were made with the virtual model control method. The results show that the centroidal dynamic and adaptive sliding mode control algorithm realizes the weight adaptability of the vehicle, and reduces tracking errors of the forward speed and lateral speed, which proves the effectiveness of the proposed control method.

In order to achieve the goal of maize inter-row weeding, this study designed the inter-row weeding device which can locate the position and avoid the injury of maize seedlings. This device includes the comb type weeding executing part designed with 65 manganese carbon spring steel wire, the maize seedling positioning part designed with travel switch and the driving part mainly based on servo motor. This inter-row weeding device was tested in June 2019 by the National Agricultural Machinery Quality Inspection Center. The test results show that the average weeding rate was 95.1%, and the average seedling injury rate was 1%. Variance test results show that the plant spacing and the stabbing depth have a significant impact on the weeding rate, and the plant spacing also has a significant impact on the seedling injury rate. The results of this study can provide useful reference for the design of inter-row weeding device in dry field.

In view of the path planning problem in the visual navigation of the residual film recovery operation after the autumn, a method of path extraction for residual film recovery operation is proposed. The color space and texture features of the three types of images of stubble, residual film and inter-row noises were extracted to constitute a training set and test set with ratio of 8:2. With the goal of detected root, a Random Forest (RF) model is built to classify the sample images. The feature dimensions are reduced by the feature importance and correlation, the optimal model parameters of the RF model are determined by grid search. Based on the upright characteristics of the stubble, the upper and lower vertices of each stubble are detected as the feature points. All the feature points are fitted by the least square method as the navigation line. The experimental results show that the 30 feature dimensions are reduced to 16 feature dimensions by feature selection, and the processing time of a single image is reduced from 0.24 seconds to 0.16 seconds. Under the optimal parameters of the model, the accuracy of the test set is 92.5%. The classification test was selected for 450 images of 10× 20 pixels, the accuracy rate is 91.8%, the accuracy rate (sensitivity rate) of stubble detection is 90.7%. 200 field operation images with ROI area size of 100×200 pixels were selected for the navigation line fitting experiment, which included different weather, different depression angles and abnormal driving images, 171 images were successfully fitted. The stubble detection method has high stability and accuracy, and can provide reference for root crop detection and path extraction of different crops.

In this paper， based on the laser melting additive manufacturing technology of NiTi alloy， two intelligent bionic protective structures with memory recovery characteristics are established by using the design principle of structure bionics. The bionic structure model is prepared by selective laser melting technology. The mechanical properties and deformation recovery properties of the bionic structure model are analyzed by numerical simulation and test. The results show that by comparing the experimental force-displacement curve and its deformation mode， the simulation calculation results more accurately simulate the deformation and load-bearing characteristics of the bionic protective structure model sample in the static pressure process； the compression of the bionic protective structure can reach 15%， its compression strength is 81.4 MPa. Its deformation recovery rate reaches 99.04% under heating conditions after unloading. This research has realized the verification of the preparation and the function of restoring deformation of bionic protective structure under pressure load by using additive manufacturing method， which provides a theoretical basis for the development of new intelligent bionic protection structure。

In order to explore the effect of radiation dose on the storage quality of packaged spiced beef, vacuum-packed and modified atmosphere packaged (75% CO₂+25% N₂) spiced beef were treated with different doses of radiation. The quality changes during storage were studied. The quality indexes include the hardness, elasticity, total number of colonies, pH, color, TBARS, TVB-N and sensory characteristics. The results show that the radiation dose is positively correlated with the bactericidal effect, the quality of the spiced beef is good when the irradiation dose is 4-6 kGy, and the radiation odor is obvious when the dose is greater than 6 kGy. The vacuum-packaged spiced beef has better storage quality after 6 kGy irradiation, the shelf-life can reach 20 days. The modified atmosphere packaged spiced beef has better storage quality after 4 kGy irradiation, and the shelf life can reach 18 days.