Since the concept of resilience was introduced into the field of transportation， resilient transportation has received extensive attention from researchers in the transportation field. Road transportation infrastructures with good resilience can effectively deal with various natural and man-made disasters， and further meet the needs of efficient and safe transportation in the future. In order to clarify the current research status of road transportation infrastructure resilience around the world， the research results of road transportation infrastructure resilience were summarized from the definition， measurement methods， and resilience improvement technologies of road transportation infrastructure， and the future road transportation infrastructure resilience research was discussed especially the development direction and research focuses of resilience improvement technology. The analysis shows that the research on the resilience of road transportation infrastructure mainly focuses on the overall traffic network organization and planning level of the road network. The research on the resilience of the infrastructure structure is relatively scattered， and there is a lack of unified and comprehensive definitions and metrics. In addition， in the research on the catastrophic failure mechanism of structural resilience， there is a lack of a comprehensive research understanding of all elements and the coupling between structural systems， and it is difficult to reveal the chain process and catastrophe characteristics of catastrophic evolution. Therefore， the research on the resilience of future road transportation infrastructure should further reveal the theories and methods of catastrophe analysis of different types of facility structures， and establish a comprehensive and unified definition and measurement standard for the structural resilience of road transportation infrastructures. At the same time， from the perspectives of disaster monitoring， structural safety and resilience improvement， flexible operation and post-disaster recovery， etc.， more effective technologies for improving the resilience of road transportation infrastructure should be formed and furtherly promoted in the future.

In response to the problem of vibration caused by the large length/diameter ratio of the boring bar during deep hole boring， which affects the processing quality and efficiency， three vibration control methods， passive control， semi-active control， and active control， were summarized. The specific structures， vibration reduction mechanisms， characteristics， shortcomings， and development trends of the three methods have been sorted out. Comprehensive analysis shows that the structure， materials， and control methods of vibration damping boring bars are currently the focus of research. With the continuous development of structural design， material science， vibration reduction mechanism， control theory， big data， artificial intelligence and other technologies， the research on vibration damping boring bars is gradually becoming diversified， integrated， and intelligent. Meanwhile， intelligence is a new development direction for vibration damping boring bars.

As a new generation of automotive materials， smart materials provide new research directions and design ideas for the realization of automotive lightweight and intelligent design goals. In order to further promote the research process of smart materials for automobiles， this paper reviews piezoelectric materials， magnetorheological materials and shape memory alloys respectively， describes the special properties of various materials， and systematically summarizes domestic and foreign research achievements in important research fields such as automotive energy recovery， structural vibration suppression， sensors， actuators， and safety protection. Finally， the challenges of commercialization of smart materials for automobiles are analyzed， and the directions of future research are pointed out.

In order to promote the application of road safety intelligent perception technology， the research progress about intelligent monitoring of pavement icing condition based on optical fiber sensing at home and abroad was overviewed. Firstly， the principle of fiber-optic pavement icing detection was revealed. Subsequently，based on the analysis of different optical fiber performance indexes， the distribution mode of common probes and various weak signal detection methods， the fiber-optic pavement icing detection system was constructed. Then， common icing detection data pre-processing and its thickness analysis methods were explored， the main environmental influencing factors and measures to enhance detection effectiveness were outlined， and burying technology of fiber-optic road icing detection sensor was analyzed. Finally， current research status of icing pavement monitoring and early warning were summarized， the existing problems were discussed， and the development direction of intelligent sensing of pavement icing condition was attempted to outlook.

In view of the adhesion performance between asphalt-aggregate， the research results on the adhesion and raveling mechanism， evaluation methods， influencing factors and improvement measures in domestic and overseas are summarized. Formation and failure of adhesion are complex processes involving physics， chemistry， thermodynamics， and micromechanics. The characteristics of the interface between asphalt and aggregate are influenced by the properties of asphalt and aggregate， the void and asphalt film thickness of the mixture， and the external environment. The different test systems are dedicated to the development of methods that that can not only simulate the process of damage occurrence in the field but also provide an assessment method through which the suitability of mixtures would be estimated in designing steps and would be guaranteed during the pavements service life. Combined with the existing research contents，the future research directions of asphalt-aggregate adhesion performance and asphalt mixture moisture sensitivity are prospected.

In order to reasonably evaluate the impact of various factors on the charging load， this paper introduced the concept of user portrait， and generated a typical user portrait that could describe the charging behavior of users through the construction and extraction of the characteristics of vehicle charging behavior data. At the same time， it is found that the shape of load curve can be adjusted by adjusting the proportion of different types of users. Through practical examples， the effects of user behavior characteristics and attribute characteristics on key grid indicators such as charging load form， peak valley time and load rate are comprehensively analyzed， so as to reasonably guide users to charge in order， provide basis for power grid planning and capacity expansion considering electric vehicle charging load.

In order to reduce the amount of historical data of text data and improve the efficiency of text data compression and storage， a text data compression and storage algorithm based on time series model is proposed. The wavelet threshold denoising method is used to estimate and eliminate the error and noise of text data； from the perspective of text data features， the features are described in detail， and the combination and inheritance relationship between feature types are set to build a time series model. Convert the preprocessed text data into binary coded bytes with similar structure using the time series model， perform XOR operation to compress the redundant part in the result， and store the compressed data in the corresponding database， and finally complete the text Data compression storage. The simulation results show that the proposed algorithm can effectively improve the compression performance and obtain more satisfactory compression and storage results of text data.

Based on large eddy simulation， different sub-lattice turbulence models are used to simulate the flow field of a peach cavity hydrodynamic coupling under braking condition. The Q-criterion vortex recognition method is used to extract the three-dimensional multi-scale vortex structure inside the turbine. The unsteady multi-scale vortex spatiotemporal evolution law， energy transfer and loss mechanism inside the turbine is analysed based on the vortex dynamics theory. In order to verify and evaluate the accuracy and reliability of the simulation results， the velocity field and vorticity field are extracted based on flow field visualization experiment by particle image velocimetry. From the perspective of the qualitative identification of three-dimensional vortex structure spatiotemporal characteristics and the quantitative extraction of two-dimensional flow field parameters： WMLE S-Ω model simulation can provide rich information of three-dimensional multi-scale vortex of the near-wall region on the blade，the small-scale vortex can be captured accurately. WALE model simulation can accurately identify small-scale vortices in the corner area where the blade and the outer ring intersect. The simulation results of the two-dimensional flow field in the mainstream area by WMLES S-Ω model tend to be true， the secondary flow phenomenon in the corner area can be presented accurately. The numerical distribution of the two-dimensional flow field simulation in the near-wall region on the blade by WMLES model is consistent with experimental values. The research results can provide certain theoretical and technical guidance for the flow field simulation of hydrodynamic coupling.

In order to solve the problem of high cost of labeling named entity data and difficulty in obtaining large amounts of labeled data in the biomedical field，this article proposes a two-stage learning framework to realize BioNER under low resources. In the first stage， Word2Vec and BERT are used as the basic model to pre-train and fine-tune to obtain the word embedding representation in a specific field； In the second stage， the generated word embedding representations are input to the neural network composed of BiLSTM and CRF and then used for the training of the final task. This paper conducts experiments on the Yidu-S4k dataset， and even in the case of a small number of labels， the results show that the algorithm in this paper achieves an accuracy of 80.94% and has great performance.

There are many uncertainties in the contact of bolted joints. It is difficult to determine the reasonable interval of contact stiffness in existing models. Therefore， the range of contact stiffness of joint surface with uncertain micromorphology is obtained by using interval Estimation theory. Firstly， based on fractal theory to characterize the contour height of the micro-convex body of the combined surface， the structural function method and interval algorithm are used to solve the uncertainty interval of the fractal parameters of the combined surface. Then， the surface topography parameters are connected in parallel with the fractal parameters by moment spectroscopy. The Chebyshev envelope function is introduced to calculate the surface morphology parameters interval under the influence of uncertainty. Finally， the uncertainty interval of surface morphology parameters is introduced into the statistical model to establish the contact stiffness model of joint surface considering the uncertainty of surface morphology parameters. The range of joint surface contact stiffness is obtained， and the influence of surface morphology parameters on the joint surface contact stiffness is explored. The results indicate that the model can accurately predict the range of changes in the surface contact stiffness of bolted connections， which can provide guidance for the design of the joint surface.

At present， in the field of image data encryption， chaos theory encryption method is more prominent. The high randomness， unpredictability and extreme sensitivity of initial parameters of chaotic system provide favorable guarantee for image data security. Based on this kind of mapping， a large number of complex chaotic mapping models have been proposed by scholars at home and abroad. Compared with one-dimensional chaotic mapping， the security has been greatly improved， but the computational complexity has also increased. Therefore， this paper proposes a one-dimensional compound Logistic-Tent algorithm for chaotic mapping， which increases the mapping range of chaotic mapping， reduces the computational complexity and ensures the security of chaotic sequences. Combined with Fibonacci scrambling transform， this algorithm adopts scrambling - diffusion encryption process. Simulation experiments show that the scheme has high encryption rate， and can effectively defend against brute force cracking， entropy attack， difference attack and statistical image attack.

Firstly， a seismic resilience assessment framework was proposed for highway bridge networks by comprehensively integrating various influencing factors （network topologies， deterioration effects of bridges， regional seismic hazards， etc.） and their uncertainties. Then， the network topologies， seismic performance of regional bridges and traffic functionalities of highway segments were analyzed， and integrated to establish a time-dependent post-disaster simulation of highway bridge networks. Subsequently， an artificial neural network based methodology was proposed for rapid seismic fragility assessment of regional bridges. Finally， the life-cycle seismic resilience of the highway bridge network in San Francisco， USA was analyzed as a case study， which reveals the evolution law of seismic resilience the service life. The results demonstrate the developed data-driven fragility model can act as a reliable surrogate to traditional time-consuming fragility methods， with the goodness of fit of 0.73. Consequently， the proposed methodology for the life-cycle seismic resilience of highway bridge networks is effective and accurate， providing decision-making strategies for disaster prevention and mitigation.

On low-resource corpus， the mainstream translation quality estimation models have poor performance. Meanwhile， the sentence embedding strategy is naive. In view of reasons mentioned above， a Korean-Chinese translation quality estimation based on cross-lingual pretraining model is proposed. Firstly， a cross-lingual sentence embedding method is proposed by drawing on the idea of attention. The method can effectively fuse the cross-layer information and token positions of the pre-trained model. Second， a cross-lingual pretraining model is introduced to the task as a way to alleviate the few-shot caused by the low-resource of Korean. Finally， the regression is performed on the sentence embedding vectors， so that the Korean-Chinese translation quality estimation can be completed. Experimental results show that the method can effectively improve the performance of the Korean-Chinese translation quality estimation task. Compared with QuEst++， Bilingual Expert， and TransQuest， the dominant models for quality estimation tasks， Pearson correlation coefficients improved by 0.226， 0.156， and 0.034， and Spearman correlation coefficients improved by 0.123， 0.038， and 0.026， respectively.

Exclusive lanes for connected and autonomous vehicles（CAVs） will emerge in order to ensure the safety and efficiency requirements in the process of traffic flow mixed with human-driving vehicles and CAVs. When the inner lane of the expressway is set as the exclusive lane for CAVs， it has important theoretical significance and practical value to study the strategy of guiding CAVs to merge from the ordinary lane to the exclusive lane. Firstly， the entrance area of exclusive lane was designed and vehicle control rules were proposed. Secondly， with the goal of making more CAVs change lanes to the exclusive lane， the strategy of selecting lane-changing signal actions was proposed based on deep reinforcement learning. Finally， the numerical simulation was carried out with Python language compilation. The results show that the proposed algorithm can converge very quickly under 9 scenarios constructed by different factors， such as the CAV penetration rates and the proportion of CAVs arriving at the exclusive lane； it can effectively guide CAVs to merge into exclusive lanes and ensure traffic efficiency； congestion in the second lane can be significantly reduced compared to the unsignalized control when the penetration rate changes from 20% to 40%； the proportion of CAVs changing to the exclusive lane is significantly higher under the two exclusive lane entrances scenario than that under the one entrance scenario. It shows that the proposed strategy has good applicability and can provide reference for engineering construction.

In order to accelerate post-earthquake repair of bridges and reduce earthquake losses， a systematic review， summary， and outlook were conducted on the research of resilient structural systems in bridges. Firstly， the development history of resilient structural systems was reviewed. Secondly， from the perspective of seismic mechanism and engineering application， the similarities and differences between ductile and resilient structural systems were elaborated. Emphasis was placed on the important way to achieve earthquake-resistant resilient in bridges the rocking structural systems， and the working mechanisms， hysteresis curves， and engineering applications of the current four typical rocking structural systems were discussed. Thirdly， the progress of the domestic and foreign cutting-edge researches in resilient structural systems was summarized and analyzed from the aspects of rocking structural systems， self-centering energy dissipation devices， and seismic design methods. Finally， the existing problems and development trends of bridge resilient structural systems were summarized and prospected.

The current deep learning-based hydrological prediction models mainly use the same network weights to simulate the spatio-temporal relationship between soil moisture and predictors， which has some limitations in reflecting the spatio-temporal specificity of soil moisture. In this paper， we propose a spatio-temporal prediction model with channel-agnostic and spatial-specific structures. Spatial-specific features are used to extract the spatial characteristics of soil moisture in different regions by adaptively assigning network weights at different spatial locations； channel-agnostic features are used to extract the spatial characteristics of soil moisture and predict in a single grid point channel using one-dimensional convolution. The channel-agnostic feature extracts the temporal characteristics of soil moisture and predict in a single grid point channel using one-dimensional convolution， accurately captures the transformed relationship between soil moisture and its predictor at a single point， and uses lagged and predictor as model inputs for prediction at a single grid point. The experimental results show that the prediction ability of the model in this paper is greatly improved compared with the traditional DL model， and most obviously in the northern region of China， the determination coefficient （R²） of the model is improved by 80% compared with that of CNN.

To study the constant flow mechanism of the constant flow control valve， the influence of flow area and gradient on the characteristics of constant flow control valve are analyzed based on the actual structure of a certain type of valve. Firstly， the mathematical model， the control theoretical model and the transfer function block diagram of the dynamic system are established to quantitatively analyze the influence of these factors. Then the AMESim simulation model is established， and the influence mechanism of these factors on the steady-state and dynamic characteristics of the valve is analyzed. The reliability of the simulation model is verified by the experiment. The results show that there exists negative feedback between stages in the constant flow control valve， which compensates the pressure difference between the two sides of throttle and reduces the flow adjustment deviation. Adjusting the flow area of the pressure compensator can improve the steady-state characteristics of the constant flow control valve. Adjusting the flow area gradient of the pressure compensator can significantly improve the static and dynamic characteristics of the constant flow control valve.

Based on the special working condition requirements of hybrid electric vehicles， this thesis conducted a study on the vibration and noise characteristics of the timing silent chain system （TSCS） based on the structural design of the TSCS. The dynamics characteristics of the TSCS were analyzed and the boundary conditions required for vibration analysis were obtained by establishing the dynamics model of the TSCS， and based on the finite element analysis method， the vibration of the TSCS was studied through modal analysis and frequency response analysis， at the same time， using the vibration characteristics of the TSCS as input conditions， the acoustic boundary element simulation analysis of the TSCS was carried out， and the acoustic pressure-frequency response curve and the distribution of noise were studied. The analysis results show the excellent vibration and noise characteristics of the TSCS， and the analysis method will provide a reference for the vibration and noise reduction work of hybrid electric vehicles.

Aiming at the problem that deep learning-based iris segmentation models need a large number of parameters， computation cost， and space occupation， a lightweight iris segmentation model is proposed in this paper. First， the feature extraction network of Linknet is replaced with the improved lightweight deep neural network MobileNetv3. This design significantly improves the efficiency of the model while maintaining accuracy. Then， in order to reduce the loss of iris feature information， a multiscale feature extraction module is designed in this paper. Once again， an efficient parallel attention mechanism is introduced to suppress noise interference and enhance the weight of iris region pixels. Finally， the proposed model was compared with other iris segmentation models on three iris databases， and the results showed that the model achieved a better balance between iris segmentation accuracy and efficiency.

The Converge was used to study the spray characteristics of porous injectors under different fuel temperature conditions， and the influence on the spray collapse degree was studied by changing the number of nozzles， ambient pressure and temperature， oil beam injection direction and radial spacing of adjacent nozzles. The results show that with the increase of fuel temperature， the degree of flashing phenomenon is more violent， the spray cone angle and spray width increase， the spray center is difficult to exchange with the external gas， so that the low pressure condition of the spray center is difficult to improve， and the oil beam gathers to the middle under the action of pressure difference， and finally evolves into spray collapse. By changing the ambient pressure and temperature， it can be found that after appropriately increasing the environmental pressure， the spray collapse phenomenon will become violent， and increasing the ambient temperature has no obvious effect on the degree of spray collapse； Changing the oil beam spacing by changing the number of nozzle holes， the direction of the oil beam and the radial spacing of the spray holes has a significant effect on the degree of spray collapse.

Aiming at the problems of fixed network topology， low dimensionality， uneven data distribution and low correlation， the training effect of intrusion detection model in industrial distributed environment is poor. In this paper， Fedformer， a federated deep learning algorithm based intrusion detection model for industrial Internet of Things （IOT）， is proposed. Firstly， the encoder structure of Transformer network model is introduced and improved， and the convolutional neural network and gated cyclic unit are integrated， and the intrusion detection model for industrial IOT is constructed by using the attention mechanism. Secondly， the detection model is integrated with the federated learning framework， which allows multiple industrial IOT to jointly build a comprehensive intrusion detection model. Under the premise of protecting the privacy of local data， the detection accuracy of industrial IOT network attacks is improved and the false positive rate is reduced. Experimental results show that the detection accuracy of Fedformer in the industrial network environment is 98.09%， and the false positive rate is reduced to 8.31%.

In bolt connection， excessive pre-tightening force will make bolts fail in case of accidental overload. Traditional method of bolt pre-tightening analysis considers that the parameters are determined， but the parameters are random in actual working conditions， which causes larger analysis errors. To solve this problem， considering the influence of random factors， a finite element analysis model of bolt pre-tightening failure behavior is proposed in this paper， the reliability analysis is carried out by judging whether the stress value at the thread of bolt exceeds the allowable stress value， and the reliability sensitivity analysis is calculated to evaluate the influence of various parameters on the failure phenomenon in bolt connection. The results show that the change of bolt diameter has the greatest influence on the phenomenon of bolt static strength failure， and the Poisson's ratio of material has the second influence on it. The reliability of bolt connection will increase with the increase of bolt tooth angle， screw pitch and Poisson's ratio of material， and decrease with the increase of bolt diameter， material density and elastic modulus.

To realize the classification and recognition of entering and leaving stops′ driving styles， based on the entering and leaving stops data in the natural driving process of pure electric bus， 14 driving behavior characterization indexes were selected， the dimension of the indexes is reduced by principal component analysis， and a K-means clustering model was established to cluster the entering and leaving stops segments into three categories. Taking economy， dynamic and comfort as the three dimensions of semantic interpretation， the three types were interpreted as high energy consumption & aggressive style， general style and energy-saving & comfort style. A three-layer BP neural network model was established to realize the on-line recognition of driving style. The model verification showed that the evaluation index values of the recognition model are near 0， and the average recognition rate of the model is 93.52%， which can better realize the driving style recognition of any entering and leaving stops segment.

A 3D model of Diesel Particulate Filter（DPF） was established by CFD software， the influence of soot and ash distribution on the airflow movement of dissymmetric DPF ducts was studied. A one-dimensional thermodynamic model of engine was used to analyze influence of the DPF carries size， wall thickness， and the inlet/outlet aperture ratio on exhaust noise under plateau environment. The results showed that： the flow velocity of DPF inlet port has been increases first and then decreases axially from the beginning end. The flow velocity of the DPF inlet port has been decreases radially from the center. The carrier diameter has significant impact on noise than the length under plateau environment. The carrier diameter of 190 mm has a better noise attenuation effect than the diameter of 170 mm， reaching more than 3 dB. The inlet/outlet aperture ratio is 1.1， the end of DPF noise is the lowest， the attenuation amplitude increases with the increase in carbon loading （0~8 g/L）， the noise attenuation effect is better more than 3 dB.

In order to improve the microstructure and mechanical properties on surface of magnesium alloys， the surface modification of Mg-3.4Al-0.8Zn-0.4Sn magnesium alloy was carried out by friction stir processing. By using a pin-less tool， setting a series of travel speeds and rotation speeds， analyzing the thermal cycle process， observing the macroscopic morphology of the surface modification zone， characterizing microstructure and measuring microhardness， the influence of process parameters on the surface modification layer was studied. The results show that， the surface modification zone can be divided into four layers including stir layer， rotation flow layer， transition layer and themo-mechanical affected layer. As the travel speed increases， the time and the peak temperature of the thermal cycle， the thickness of the surface modification layer and the average grain size decrease， while the microhardness increases. By contrast， as the rotation speed increases， the peak temperature of the thermal cycle， the thickness of the surface modification layer and the average grain size increase， while the microhardness decreases. The modification layer obtained at each parameter showed significant grain refinement and microhardness increment compared to the base material.

In this study， the recognition test of bok choy and its associated weeds at the seedling stage was carried out， and a novel method based on recognizing vegetables and then indirectly recognizing weeds was proposed. By combining deep learning and image processing technology， this method can effectively reduce the complexity of weed recognition， and at the same time improving the accuracy and robustness of weed recognition. First， a neural network model was used for detecting the bok choy and drawing bounding boxes. The green targets outside the bok choy bounding boxes were marked as weeds， and color features were used to segment them. Besides， an area filter was used for eliminating the noises and extracting weed regions. In order to explore the effects of different deep learning models on bok choy recognition， SSD model， RetinaNet model and FCOS model were selected， and three evaluation metrics of F₁ value， average accuracy and detection speed were used for comparative analysis. The SSD model was the best model for bok choy recognition， with the highest detection speed and excellent recognition rate. Its F₁ value， average accuracy and detection speed in the test set were 95.4%， 98.1% and 31.0 f/s， respectively. The improved MExG index can effectively recognize weeds， and the segmented weeds have complete shapes and clear outlines. Experiment results show that the proposed method for recognizing weeds in vegetable fields is highly feasible and has excellent application effects， which can also provide technical reference for weed recognition in similar crop fields.

To improve the mechanization harvesting level of quinoa and solve the problems of a significant loss rate， high impurity rate， and poor feeding during harvesting of ordinary rice-wheat combine harvester， according to the characteristics of quinoa plants during the harvesting period， a sizeable self-propelled quinoa combine harvester was designed. The machine used flaring type small row spacing chain teeth feeding into the header， combined longitudinal axial flow threshing cylinder， unique woven screen concave， double reciprocating vibrating screen and other devices， cooperation wide bridge， threshing cylinder CVT regulation， and excellent space design to realize the smooth quinoa feeding and efficient threshing separation. Critical components for the design were analyzed， and a field experiment was carried out. The results of the field experiment showed that when the water content of quinoa grain was 14.42%， the threshing rate was 96.83%， the impurity rate was 4.41%， the damage rate was 0.2%， the header loss rate was 1.14%， the entrainment loss rate was 1.73%， the cleaning loss rate was 1.09%， the splash loss rate was 0.16%， the total loss rate was 4.13%. During the experiment， the machine runs smoothly and meets the mechanical harvesting of quinoa. This paper can provide some reference for the design and test of quinoa combine harvester.

The vibration characteristics of the whole wheel tractor and the symmetrical parts of the structure were collected and analyzed by using the vibration test system， and the vibration similarity of the symmetrical structure was compared with the vibration intensity. The results show that the vibration characteristics of the symmetric parts of the front axle and the rear axle are symmetrical. The main vibration direction of the symmetric measurement points is the same， and the vibration similarity of the symmetric positions of the front axle and the rear axle is 93.91% and 94.29%， respectively. It is concluded that the vibration characteristics of the whole tractor and its symmetrical parts are affected by the working state of the whole tractor， the defects of parts and the assembly quality. The difference of the parts and assembly quality of the symmetrical parts of the tractor changes the vibration transmission path and affects the similarity of the vibration characteristics of the symmetrical parts. This paper presents a method for the comparison and evaluation of the vibration characteristics of the tractor symmetric structure， which can provide a fast-screening method for the on-line detection of tractor working state and the discovery of mechanical structure defects and faults.

To effectively improve the active safety of autonomous vehicles， a single parent genetic algorithm based active collision avoidance method for autonomous vehicles is proposed. Construct a dynamic model for autonomous vehicles using normalized azimuth evaluation functions. Based on the predicted behavior trajectory， pre braking is taken in advance， and a single point gene replication strategy is introduced to implement genetic operations with the goal of minimizing the path from the initial point to the endpoint， in order to obtain the optimal solution for autonomous vehicle active collision avoidance. The simulation results show that this method can plan a reliable turning collision avoidance path under emergency conditions， with stable vehicle state parameters， small path deviation， and strong robustness to different vehicle speeds and turning paths.

The wind speed law in the simplified U-shaped valley was studied by theoretical deduction and terrain model wind tunnel test， and the buffeting response of a long-span suspension bridge in the valley was evaluated. The results demonstrate that the wind speed in the simplified U-shaped valley is the result of combined effects of pressure gradient and width to depth ratio. The acceleration effect of wind speed grows with the increase of the width to depth ratio. The spanwise distribution of wind speed conforms to the parabolic model， and the result of the terrain model wind tunnel test also verifies its rationality. Compared with the parabolic distribution model， the uniform distribution wind speed model recommended by the Code overestimates the buffeting displacement response at the main girder position and the internal force response at the key position of the bridge. The research can provide a reference for bridge wind-resistant design in the valley.

Aiming at the problems of high measurement cost， complex data fusion， and few extracted features in traditional 3D motion parameter measurement methods， a non-contact measurement method of 3D motion parameters based on pose estimation is proposed. First of all， the use of high frame rate synchronous acquisition equipment to capture object motion video from different perspectives makes up for the shortcomings of blind spots that are prone to single-view capture targets. Secondly， using the 2D pose estimation model of the training target based on the deep transfer learning 2D pose estimation framework DeepLabCut， and using the model to estimate the 2D pose of the object from different perspectives， makes up for the shortcomings of the traditional vision method with fewer features. Third， the 2D pose estimation of objects in different directions is merged into a 3D pose estimation based on the principle of triangulation. Finally， according to the measurement requirements， the 3D motion parameters such as the displacement， speed， pose angle， frequency of the specified key points of the object can be calculated using the results of the 3D pose estimation. Taking the motion measurement of the planetary pendulum and the human hand as an example， we measured the 3D motion parameters of the two and analyzed the sources of error and the advantages and disadvantages of this paper. The experimental results show that the system measurement is accurate， and the visualization results correctly reflect the movement of the experimental subjects. This method provides a new idea for the non-contact measurement of 3D motion parameters of objects.

In response to the problem that the landslide displacement prediction models are mainly based on one-step prediction combining dynamic variables such as rainfall and deformation， and lack of consideration of static variables such as time period related to multi-step displacement influence factors. A multi-step landslide displacement prediction model integrating dynamic and static variables was proposed. First， the variable selection network was used to select the initial input variables， excavate the highly correlated variables with the daily displacement of the landslide， and weaken the influence of redundant variables on the model. Then， the static variables were integrated into the network， and the dynamic correlation of time was adjusted by encoding the context. Finally， the multi-step displacement prediction of landslide was realized by capturing the long-term dependence of time series with multi-head attention module. Taking Xinpu landslide in Chongqing as an example， the method is compared with DeepAR and Long Short-Term Memory（LSTM） models. The experimental results show that the method can achieve more robust and high-precision multi-step displacement prediction of landslide.

An improved collaborative filtering recommendation model is proposed to recommend the new product service system for users quickly and accurately. Firstly， the target user neighbor set is determined. Aiming at the problem of data cold start in the traditional collaborative filtering algorithm， a method combining user attribute similarity and user experience similarity is proposed， and Jaccard coefficient， average score correction coefficient and popular coefficient are introduced to improve the accuracy of user experience similarity. Then， the similarity set of new product service system is determined. An improved Pearson cosine similarity algorithm based on product service system attributes is proposed to solve the problem that the traditional project-based collaborative filtering algorithm ignores the similarity constraints of project attributes， and BP neural network is used to obtain the objective weight of each attribute under different product service systems， which improves the reliability of attribute importance. Finally， the recommendation guideline to judge whether the new product service system can be recommended to target user is constructed. Taking the tractor service system recommendation as an example， the feasibility and effectiveness of recommendation model are verified.

The point-of-interest （POI） recommendation services provided by the location-based social network （LBSN） have become an important means of mining users' preference for POIs. The sparsity of user-POI matrix is the primary problem to be solved， and a large number of unknown values in implicit feedback cannot reflect user preferences. To improve recommendation precision， this paper proposes a point of interest recommendation algorithm integrating social geographical information based on weighted matrix factorization （SGWMF）. The social information is modeled through the power-law distribution. The check-in information of the user's friends is converted into the user's visit location preference. Secondly， the power-law distribution of geographical information is used to construct the user's visit location preference matrix to alleviate the data sparsity problem. Thirdly， in order to extend the effectiveness of the model， we improve the objective function by adding implicit feedback term. Finally， the experimental results on two datasets show that it has better performance than other POI recommendation algorithms and can improve the accuracy of recommendation results.

The tire low-frequency handling model and the high-frequency durability model were unified， and the UniTire tire model including in-plane dynamic characteristics was proposed. The steady-state UniTire is the zero-frequency output of the tire dynamic system. When the frequency of the tire motion input signal is infinite， the output of the wheel center force is zero. Using these as boundary condition， the transfer characteristics of the tire dynamic system are established. Considering that in the simulation of the vehicle virtual proving ground （VPG）， the load history of the wheel center is mainly affected by the rigid ring modes of the tire， that is， the in-plane frequency is less than 100 Hz， and the out-of-plane frequency is less than 60 Hz. Higher frequency components contribute to the load analysis of the wheel center indistinctively， so a second-order transfer model that satisfies the boundary conditions is established. Finally， using typical tire dynamic test data， the tire model parameter identification and test verification are carried out. The comparison results show that the proposed model can accurately express the in-plane tire dynamic characteristics.

Measuring cup to disk ratio（CDR） by segmenting optic disc（OD） and optic cup（OC） is an effective method for diagnosis of glaucoma. Compared with OD segmentation，OC segmentation still faces difficulties in segmentation accuracy. In this paper，a deep learning architecture MS-HRNET is proposed for joint segmentation of OC and OD. It is an improved architecture based on HRNET. By adding multi-scale input to HRNET，information loss during feature extraction can be compensated. Combined with multi-scale spatial and channel attention mechanism，deep image features are extracted. By adding a side output layer，the early training of the network is guided. Experimental results show that the proposed model performs better than the existing OC and OD segmentation methods on Drishti-GS1 and REFUGE datasets.

In object-level hyperspectral image modeling with limited leaf samples， a multi-scale three dimensional-one dimensional cascaded convolution neural network model was proposed. Firstly， in 3D convolution neural network （3D-CNN）， the dilated convolution was embedded to increase the receptive field of convolution kernel， and a multi-scale 3D-CNN network was constructed to extract and fuse spectral-spatial joint features of different scales， to improve the model performance without increasing the number of network parameters. Then， the one-dimensional convolutional neural network （1D-CNN） was cascaded to the optimal multi-scale 3D-CNN network， to further reduce the computational complexity and overfitting degree of model. Finally， on two datasets of chlorophyll content regression for basil leaf and drought stress recognition for pepper leaf， the optimal network architecture was explored， with comparison to a series of baseline CNN models. Experimental results showed that the proposed model can effectively improve the generalization performance and reduce the computational complexity for both regression and classification tasks of leaf hyperspectral images under the condition of small samples.

Based on the health monitoring information of the in-service bridge， it is very important to reasonably analyse the structural reliability for the safety and serviceability assessment. In this paper， the time series analysis method， the moving average method and the cubical smoothing algorithm with five-point approximation are used to reduce the noise of monitoring information （extreme stress signal）. Considering that some time series is not enough to predict with a single dynamic linear model， these time series need to be predicted by a combination of multiple dynamic linear models. This study introduces multi-process model and establishes a multi-process Bayesian dynamic linear model to predict and analyse the extreme stresses； Further， the time-dependent reliability of bridge members is predicted by combining multi-process Bayesian dynamic linear model （BDLM） and first-order second moment （FOSM） reliability method. The research results of this paper will provide the theoretical basis for structural reliability prediction.

To reduce the collision probability between robots and obstacles and improve the efficiency of robots， a deep learning based obstacle avoidance planning algorithm for indoor navigation robots was proposed. Firstly， through the indoor navigation robot navigation system， combined with deep learning， the detection and recognition capabilities of moving and non moving obstacles in the robot's environment were improved， thereby obtaining practical reactive obstacle avoidance navigation information that is more in line with the actual scene. Then， using this information， a simulation map was constructed， and an optimal task execution route was selected within the simulation map， the problem of difficulty in obstacle avoidance planning caused by disorderly and irregular obstacles was solved， and obstacle avoidance planning for indoor navigation robots was achieved. The experimental results show that the proposed method has a more reliable obstacle avoidance path， and the obstacle avoidance planning time does not exceed 1.2 s， effectively improving the obstacle avoidance accuracy and work efficiency of indoor navigation robots.

In order to improve the driving stability of the narrow vehicle on a bend， a tilt-driving mechanism is designed according to balancing the gravity torque and the centrifugal torque generated by vehicle tilting to avoid vehicle roller. The kinematics model of the tilt-driving mechanism is established by vector polygon method， which analyzes the relationship between tilt driving and tilt executing parameters. An active tilt reverse three wheels prototype vehicle with the tilting-drive mechanism is developed. And the turning and tilting experiment of S-track is conducted. Taking the theoretical tilt execution parameters as the target parameters， the theoretical value of motor is solved and used as the input. The theoretical and experimental data of motor and vehicle tilt parameters are compared and analyzed. Finally， the experiment results show that the experimental data are consistent with the theoretical data and the root mean square error of tilt angle is 0.89°. It verifies that the design scheme of the tilt-driving mechanism is feasible and the kinematics analysis is correct. At the same time， the vehicle tilt-driving control is verified to be reliable.