The coordinated control of multiple power sources and dual clutches in the starting process of dual clutch plug-in hybrid vehicle is studied. Firstly， the dynamics of different starting modes of the system is analyzed， and the system mathematical model and starting performance evaluation index are established. On this basis， considering SOC， load condition， slope and other factors and the driver's starting intention， the system starting mode decision-making method is established. The coordinated control strategy of power source and dual clutch for each starting mode is formulated with the maximum impact as the constraint. The simulation results show that the selection of starting mode under different working conditions is reasonable. The maximum impact degree is 7.31 m/s³， and the maximum sliding friction work is 4.94 kJ， all within a reasonable range， which verifies the effectiveness of the strategy.

A thermodynamic model for the scroll compressor's working process is developed based on the thermodynamics of the variable mass system and the control volume method. The unsteady numerical simulation of the compressor's internal flow field was performed using the CFD （Computational Fluid Dynamics） method. The temperature， pressure， velocity， and mass flow of the fluid in the working chamber change were obtained. The test platform is built to test the inlet and outlet volume flow， discharge temperature， driving motor temperature and body vibration value of the prototype change trend. The results show that the thermodynamic model of leakage and heat transfer is more in line with the actual working process of scroll compressor. Because of the mass exchange between adjacent working chambers， the temperature and velocity distribution in the working chamber will be uneven. Under scroll compressor different discharge pressures， the maximum difference between the inlet and outlet volume flow of the scroll compressor is 0.15 {\mathrm{m}}^{\mathrm{3}}/\mathrm{m}\mathrm{i}\mathrm{n}， and the maximum temperature difference between the discharge temperature of the scroll compressor is 19 ℃.

In order to study the thermal dissipation performance of a liquid-cooled plate with parallel channels， the three-dimensional steady-state analysis was performed by using CFD method. The effects of coolant flow rate， channel width， depth and layout of enhanced heat transfer structure on the performances of a liquid-cooled plate were contrastively investigated， including the thermal dissipation， the uniform temperature and energy consumption. The results indicate that the thermal dissipation performance is improved by increasing the coolant flow rate， but excessive flow rate leads to increased energy consumption and limited improvement effect. Designs of decreasing channel width from the center to two sides， decreasing channel depth and adding enhanced heat transfer structure are all beneficial to the thermal dissipation and temperature uniformity of the liquid cooling system. In addition， the design of wholly added enhanced heat transfer structure （S1） reduces the average temperature and maximum temperature difference by 8.9 °C and 9.06 °C respectively， compared with the design of equivalent channels width （A5）. The conclusions provide a theoretical direction for structural design of battery thermal management system.

The existing regenerative braking strategy has insufficient constraints on the battery， resulting in low energy recovery efficiency， battery overcharge， and other problems. To protect the battery and improve the regenerative braking efficiency， the state of power of the battery has been fully considered. An optimized regenerative braking strategy has been proposed using the dynamic programming algorithm to obtain the optimal solution. Under fixed working condition and NEDC working condition,Compared with the method based on the I curve， the strategy proposed improves the economy by more than 36% and the stability by more than 48% compared with the single objective method. In addition， test experiments verify the practicability of the proposed strategy.

For the feature that the RFCS has different models in motion and positioning stage， an extended state observer assisted model predictive control （ESO-MPC） method was proposed. The time-domain discrete difference equation was used to predict the dynamic response of RFCS， and an adjustable parameter during the time-domain discretization process was introduced to deal with uncertain factors of the experimental model. The feedback information was obtained through the Expanded State Observer （ESO）， which can observe the position and speed of the RFCS as well as the spring and damping forces of the flexible hinge in real time. The performance of traditional PID， feedforward PID， LADRC and ESO-MPC was compared by 25 sets of RFCS point-to-point experiments and load experiments. The experimental results show that the four control schemes can achieve a steady-state error of ±0.1 μm， and ESO-MPC has the smallest setting time and the robustness of ESO-MPC is verified by load experiment.

A control strategy of active suspension systems was proposed for high-mobility rescue vehicles—model predictive control strategy based on ensemble Kalman Filter technology（EnKF-MPC）. Firstly， dynamic model of the active suspension system was completed for high-mobility rescue vehicles. The data of the vehicle dynamic system and the vehicle-mounted positioning system were fused through the Ensemble Kalman Filter technology in order to realize the accurate estimation of the vehicle's pose information； Aiming at the problem of vertical positioning error for the vehicle positioning system， a point cloud matching algorithm was designed to complete the accurate evaluation of the vehicle's vertical direction information； in addition， a model predictive control strategy was proposed， which used the vehicle's pose information obtained by the ensemble Kalman filter algorithm and the road profile information obtained by the on-board lidar as system inputs to control the active suspension system of the vehicle to improve the ride comfort and handling stability of the vehicle. Finally， a real vehicle test was carried out. The research results show that the vertical direction error of the proposed vehicle pose estimation algorithm is about \pm \mathrm{3.100}\text{?}\mathrm{c}\mathrm{m}， the pitch angle error is about ±0.175°， and the roll angle error is about ±0.210°. Compared with the passive suspension system， the proposed active suspension control method reduced the root mean square value of the vertical displacement by 37%， the pitch angle by 35%， and the roll angle by 35%， which significantly improved the ride comfort and handling stability of the vehicle.

In view of the problem that hyper parameters such as the number of hidden layers， the number of nodes in each hidden layer， the sparse coefficient and the dropout ratio of input data will directly affect the diagnosis performance of the network when deep neural network is used for rolling bearing fault diagnosis， an intelligent fault diagnosis method of rolling bearing based on optimized and improved stacked denoising auto encoders （SDAE） was proposed. The student psychology based optimization （SPBO） algorithm was used to adaptively select the hyper parameters of the denoising auto encoder （DAE） network to determine the optimal structure and parameters of SDAE network. Then the fault state features with stronger representation was extracted and input to soft-max classifier to achieve accurate diagnosis of rolling bearing operating conditions. Three open source datasets are used to verify the performance of the proposed network， the results illustrate that the diagnosis method based on SPBO-SDAE network is superior to support vector machine （SVM）， back propagation （BP） neural network， radial basis function （RBF） neural network， traditional SDAE network， SPBO based deep belief network （DBN）， genetic algorithm （GA） based SDAE network and particle swarm optimization （PSO） based SDAE network in feature extraction， diagnosis speed and fault diagnosis accuracy.

In order to solve the problem of traffic safety and traffic congestion caused by cut-in behavior on urban roads， the decision model of cut-in behavior was proposed in the traditional environment and network environment. Considering the safety gain， speed gain and lane-changing gain， the benefits of both parties in the game were quantified， and the benefit function of lane-changing vehicle and lag vehicle in cut-in behavior was given to obtain a game model and the Nash equilibrium solution in view of the driver's intention to change lanes. The model was calibrated and tested using NGSIM data. The results show that the model has a good fitting degree for the decision behavior of lane change vehicles and lag vehicles on urban roads. RMSE of the decision behavior variables of lane-changing vehicles and lag vehicles are 0.1385 and 0.4361 for mandatory lane change cut-in behavior， and 0.2278 and 0.1748 for discretionary cut-in behavior， respectively.In the game behavior， the safety gain has a greater influence on the driver's decision than the lane-changing gain， and the discretionary behavior is more obvious.However， the speed gain always has a greater influence on the decision-making of the drivers behind than the safety gain， and the gain weight is 0.9 and 0.1.

In order to better protect the important traffic nodes， a rail transit node importance evaluation algorithm based on rough set theory was proposed. For the uncertain passenger flow factors in each node of rail transit， the boundary domain concept was used to clearly describe the fuzzy data. The knowledge of the sorted concept sets and elements was reduced， decision rules was added to complete the information solution. The size of node passenger flow was judged，the node weight was determined by knowledge entropy, and the node importance was evaluated from three dimensions: node centrality， node proximity centrality， and node order centrality. The experiment proves that the evaluation time of the proposed algorithm is within 3 s， the evaluation time is short， and the evaluation result is accurate.

For two lane traffic scenarios， a decision-making and optimization-based lane changing trajectory planning method for autonomous vehicles was proposed. Firstly， a risk assessment method based on the Bayesian probability theory was designed to obtain the conditional probability of the lane safety in the current scenario； then， a behavior decision-making method based on the safety utility was designed. According to the risk assessment Bayesian network and decision graph， the behavior decision of lane keeping or lane changing was made. An optimization-based trajectory planning method based on the nonlinear MPC was proposed at the trajectory planning layer， which imitates the excellent driver to give the weight coefficient of each optimized objective function to solve the optimal desired lane changing trajectory. At last，The effectiveness of the decision-making and trajectory planning method was verified by the simulation. The simulation results show that the risk assessment， behavior decision-making and optimization-based trajectory planning method can make the safe behavior decision and plan the optimal lane changing trajectory for autonomous vehicles in different risk scenarios， so that the autonomous vehicle can change the lane safely and quickly.

The layout process of auxiliary stopping areas （ASAs） for normal high-speed maglev is facing the challenges of various demands and complex constraints. Focusing on ensuring the safety of maglev train operation with fewer ASAs， a scenario division-based method for ASA layout was proposed. Firstly， the general scenarios of ASA layout were divided into train operation direction， train speed profile， demand priority section and demand restriction section， and the optimization models of ASA layout were constructed. Then， taking that each ASA is as far as possible from the terminal station as the criterion， the optimization rules for different scenarios were designed， and then a heuristic algorithm for solving the optimization model was proposed. Finally， numerical experiments were carried. The results illustrate that the proposed method can efficiently obtain reliable and economical layout schemes. The demand of ASAs increases as the train stepping redundancy time increases. When setting up ASAs for bidirectional operation， the demand for ASAs can reduce by using the ASAs in the opposite direction.

Through the unconfined compressive strength test， the stress-strain curve， unconfined compressive strength， and elastic modulus of cement-modified slurry （CMS） were analyzed， and the time effect model of CMS elastic modulus was established. Based on damage mechanics， a particle swarm optimization （PSO） algorithm was used to identify random field parameters， a CMS mesoscopic random damage model was established， and a random field parameter time effect model was proposed. The results show that： ①the stress-strain curve of CMS is a softening curve， the optimal cement content is 20%， and the elastic modulus is a function of cement content and curing time. ②The mesoscopic random damage model can describe the stress-strain relationship of CMS， and the random field parameter λ has a functional relationship with the curing time. ③The macroscopic stress-strain characteristics of CMS can be explained from the perspective of meso-random damage through the damage variable D and strain evolution.

Pneumatic impact technology was used to repair the fatigue crack which initiates from the arc profile of rib-to-diaphragm， and the maintenance effect was evaluated by fatigue test. The bending fatigue tests of 11 specimens were carried out. The fatigue properties and the crack propagation law of the arc profile were discussed. The variation of fatigue crack growth rate and local bearing capacity after pneumatic impact maintenance was analyzed. The maintenance effect of pneumatic impact technology on fatigue crack of rib-to-diaphragm was revealed. The results show that the fatigue strength of the rib-to-diaphragm arc profile is about 46 MPa， which is lower than the recommended value in the Chinese code （JTG D64—2015）， under the action of out-of-plane deformation. This kind of fatigue crack is easy to initiate at the weld cover foot， and mainly propagates along the diaphragm base material， showing the characteristics of composite crack propagation. The cracks of all specimens have obvious growth lag after pneumatic impact maintenance， which indicates that the technology has a significant effect on the suppression of fatigue crack growth in this part. The results of more than 4 years engineering application also verify the effectiveness of the proposed method in the fatigue crack repair of rib-to-diaphragm.

In view of the limitations of foundation pit engineering investigation and the temporary nature of the project， a risk identification method of high-rise building foundation pit engineering based on fuzzy clustering maximum tree algorithm was proposed. The fuzzy similarity matrix was established by the absolute value subtraction method， the risk index data was standardized， and the risk evaluation index was used to divide the risk level of foundation pit construction. Using fuzzy clustering to quantitatively influence the coupling relationship between elements， the maximum tree structure in the form of undirected connected weight graph was described， and the risk identification was realized through the connected vertices in the connection tree structure. The simulation results show that the proposed method can accurately describe the specific location of the risk while determining the risk level of foundation pit engineering， and the recognition accuracy and efficiency can meet the expected requirements with strong robustness.

In order to solve the problems of slow speed and low accuracy in parking space status detection algorithm， a detection algorithm for parking space status based on convolution network structural re-parameterization was proposed. The algorithm uses structural re-parameterization to decouple the training network and inference network. At training time， a multi-branch structure was formed using small convolutional kernels of different scales for simultaneously acquiring local detail features in parking space images， so that the network achieves high detection accuracy. After training， the training-time multi-branch structure was equivalently transformed into a single-branch structure for inference using structural re-parameterization， which significantly improves detection speed without loss of detection accuracy. In this way， the network can have both higher detection accuracy and faster inference speed during inference. The experimental results show that the proposed algorithm has obvious advantages in prediction accuracy and algorithm inference speed compared with other car position state detection algorithms.

Aiming at the problem of poor robustness in visible light and thermal infrared image fusion tracking using Siamese network architecture， a real-time RGB-Thermal （RGB-T） infrared target tracking method based on feature fusion of dual Siamese network was proposed. Firstly， dual Siamese network was used to extract the features of template branch and the search branch of the visible light and infrared images， and two different modalities feature layers were obtained. Secondly， the self-attention feature enhancement module （SFEM） was used to enhance the features of the two modalities， and the dual-modal feature fusion （DMFF） module was used to fuse the features of template branch and search branch respectively. Finally， the template branch and search branch were used for cross-correlation operation， and the target position was obtained by classification and regression branches， so as to complete target tracking.

The test results on the grayscale thermal infrared target tracking dataset (GTOT) show that the precision rate (PR) of the proposed method is 91.8%, the success rate (SR) is 78.1%, and the running speed is 60 \mathrm{f}/\mathrm{s}. It shows that, compared with other RGB-T fusion tracking methods, the proposed method has higher robustness while maintaining real-time processing speed.

Aiming at the problems of low localization accuracy and classification accuracy in medical imaging disease detection by existing target detection algorithms， a multi-target detection method based on dynamic weighted Bagging ensemble learning was proposed. Taking chest imaging disease detection as an example， the Coordinate Attention （CA） module was introduced to enhance the receptive field of the region and improve the weak learner's ability to locate the target region. The dynamic weighted Bagging ensemble learning method was used to give the weak learner voting weight according to the confidence. The variance of the model was reduced， the generalization error was improved， and the classification accuracy was improved. The experimental results show that in the chest imaging disease detection task， the average detection accuracy of the proposed algorithm reaches 41.9%， which is 2.5% higher than that of the YOLOv5 prototype， and G-mean is improved by 1.3%； after the model is weighted and integrated， the average detection accuracy rate reaches 81.06%， which is 1.58% higher than the original model， and has high positioning accuracy and classification accuracy. Therefore， the proposed algorithm can better complete the task of chest imaging disease detection.

An enhanced deep iris classification model EnhanceDeepIris was proposed， with the help of generating network， second trains a iris classification network of deep learning which has already converged on the original training set， to make it can continue be trained and get better generalization ability on the test set. Three most advanced image classification networks VGG16， ResNet101 and DenseNet121 were used to verify the improvement effect of EnhanceDeepIris on deep learning classification networks. The method was tested on two iris datasets CASIA-Iris-Thousand and JLU6.0. Compared with the traditional data augment method， the classification model trained by EnhanceDeepIris has higher correct recognition rate and more stable test effect.

Aiming at the limitations of automatic classification method for retinal artery and vein vessels（\mathrm{A}/\mathrm{V}）， an automatic retinal A/V classification method based on attention U-Net （Attention U-Net，AU-Net）was proposed. The retinal \mathrm{A}/\mathrm{V} feature information was enhanced by using vascular structure information， topological relationship and edge information. The attention block was introduced into the VC-Net network model， which improvemented on U-Net， combining the local and global information， adjusting weight to restrict the retinal A/V features， such as inhibiting the background tendency features and enhancing the vascular edge and end features， so as to realize the accurate classification of retinal A/V. The method was tested in the DRIVE data set. The retinal A/V classification accuracy is 0.9685， F₁ value is 0.9886， sensitivity is 0.9803 and specificity is 0.9957. The experimental results show that compared with the classical U-Net， the performance indexes of the proposed method are significantly improved， which can be used for clinical reference.

In non-cooperative scenes， to solve the problems of the poor reconstruction of faces and the low accuracy of face recognition due to the bad condition of posture deviations and the low image resolution， a triplet loss constrained generative adversarial network for reconstruction and recognition was proposed. The input faces were reconstructed firstly by encoder-decoder network， and then the reconstructed faces were recognized. In terms of training， a shortest distance triplet loss function was designed jointly with the adversarial mechanism， which increases the similarity intra-identification while enlarge the discrepancy inter-identification on the extracted feature representations. It is shown by the experiments that even under the condition of large angle deviations and low image resolution， the proposed algorithm performs better than currently leading face pose correction algorithms.

Aiming at the ambiguity of attribute information partition of social users in network， a multi-granularity social user image construction method based on knowledge graph was proposed. Using knowledge graph to map user information and classify multi-dimensional attributes of mapped data， a set of social users was established. The same weight value of each user's corresponding attribute tag in the set was given， and the weight parameters and the frequency of interest tag were calculated. Then， according to the click-frequency information of users， a sequence of continuous active locations and active ranges of users in a multi-granularity social network was constructed， and the features of each user in the sequence were calculated and divided in a unified manner to complete the establishment of user images. Simulation experiments show that the user profile construction time under the proposed method does not exceed 25 ms， and the maximum number of matching overlaps of corresponding attributes is 705. The matching degree is high， indicating that the constructed user profile information is clearly divided.

To solve the problem of the large inter-modality discrepancy and the intra-class variations in cross modality person re-Identification （CM-Reid）, a novel CM-Reid framework based on semantic coupling and identity-consistence constraint was proposed. In the semantic level， the semantic representations bi-directionally and fuse the semantic information between different modalities were coupled to alleviate the inter-modality discrepancy. In the identity level， the cross-modality triplet loss and identity loss to maintain the identity consistence were optimized to alleviate the intra-class variations. The experimental results show that the proposed method can effectively improve the performance of CM-Reid. Compared with the baseline method, the accuracy of Top-1 and mAP indicators is improved by more than 10%.

In the process of building a large-scale quantum key distribution network， a routing algorithm based on bucket weight calculation for quantum key distribution network is proposed to improve the success rate of key relaying in trusted relay mode， reduce the link key consumption and balance the network load， and a network architecture model based on software-defined network SDN is constructed. Among all the paths from the source node to the destination node， the controller selects the optimal path by calculating the Bucket weight， that is the bucket weight as the path priority. The simulation results show that the proposed routing method is reasonable and feasible.

Aiming at the poor generalization ability of dense correspondence computation algorithms of the 3D shapes using deep learning， a method for computing correspondence based on unsupervised deformable networks was proposed. Firstly， the local feature descriptors of non-parametric templates were extracted by the template feature extraction network based on spiral convolution. Secondly， the global features of the input shape were obtained through the encoder in the template deformation network， and spliced with the local feature descriptors of the template. Then the joined features were input into the decoder to calculate the deformation model of the template. Finally， the dense correspondence between the deformation model and the input model was calculated by the nearest neighbor search algorithm. The experiment results show that the proposed algorithm can obtain more accurate dense correspondence under the same training dataset and has a stronger generalization ability than that of the current mainstream algorithms.

Based on the slime mold algorithm （SMA）， an effective matching model （SMA_MSA） was developed to assist bioscientists in judging whether different sequences have homology， so as to predict protein structure. SMA_MSA and some other well-known competing algorithms were tested in the six data sets of BAliBASE 3.0. The results clearly show that SMA_SMA has excellent matching ability in the 31 data sets， indicating that the proposed model has great development potential in the problem of protein multiple sequence alignment.

In order to improve the safety of longitudinal follow-up control algorithm for unmanned vehicles， an indirect adaptive brain emotion neural robust controller （iARBERC） was proposed based on the brain emotion learning model and the good nonlinear approximation characteristics of radial basis function neural network. The stability of the control system was proved by Lyapunov analysis method. The simulation results show that iARBERC has the fastest response speed， the smallest tracking error and the best robust performance. Although the total control energy was improved， the deviation is less than 4%. Finally， iARBERC was applied to the semi-physical simulation platform of the following control system of the unmanned vehicle. The results show that the vehicle equipped with iARBERC has good following ability to the frequently changing longitudinal speed.

The acoustic-to-seismic coupling landmine detection system based on a parametric acoustic array was constructed to study the application method of the parametric acoustic array in acoustic landmine detection technology. Based on the theory of acoustic-to-seismic coupling landmine detection， the parametric acoustic array was used as an acoustic energy source to excite the surface vibration which can be detected by an accelerometer. The three-dimensional characteristic maps of surface vibration under different types of landmines and buried soil conditions were constructed. The experimental results show that the acoustic-to-seismic coupling efficiency of the soil above the landmine is obviously better than that of the bricks and other disturbances， and it is affected by the type of the landmine and the surrounding soil conditions， which indicates that the parametric acoustic array can be used for further research on the development of acoustic landmine detection engineering system.

To overcome the shortcomings of current DV-Hop algorithm， such as low positioning accuracy， low positioning efficiency and poor anti-interference ability， a sensor node positioning mechanism based on improved DV-Hop algorithm was designed to achieve ideal positioning effect. Firstly， the factors of large sensor node location error in DV-Hop algorithm were analyzed， and corresponding improvements were made. Then， DV-Hop algorithm was used to estimate the location of sensor nodes preliminarily， and DV-Hop algorithm was used as the initial value of Steffensen iteration model. The optimal sensor node location was obtained through Steffensen iteration. Finally，the simulation results were compared with other sensor node localization algorithms. The results show that the positioning error of the proposed mechanism is small. The average time of sensor node positioning is significantly shortened， and the efficiency of sensor node positioning is significantly improved.

Aiming at the issue of microwave modeling based on neural network， a structure parameter estimation method for microstrip bandpass filter based on convolution neural network is proposed. A dual band， three band and four band microstrip filter was designed and analyzed by adopting single open stub loaded rectangular ring resonator and transverse signal interference technology. Applying electromagnetic simulation software extracted the model training data and the S-parameters and structure parameters of microstrip filter are designed as the input and output of the proposed convolution neural network， respectively. Furthermore， the trained model was used to predict the structure parameters of microstrip filter. The proposed method adopts convolutional neural network to design the microstrip filter， which can effectively solve the problems of many input parameters， complex model caused by full connection and long time-consuming when using neural network to design microstrip filter. The simulation results show that the S-parameters of the microstrip filter designed by this method have high accuracy.

An improved ant colony algorithm based on physical electric field model was proposed in order to solve the problem of a single UAV visual navigation for ground moving target search. The complex search area was defined， the grid was divided， and the grid node was defined as an activable pheromone point. The ant colony algorithm was optimized based on the potential field rules of the physical electric field model， and the probability model was introduced. An improved particle swarm optimization algorithm for UAV search based on electric field model was established to control the pose and speed of UAV. The simulation results show that the average success rate of the improved ant colony algorithm for searching moving targets is 67.3%， and the average operation time of the example is 8.33 s. The simulation results show that the improved ant colony algorithm has higher success rate and less time-consuming than the particle swarm optimization algorithm.

Aiming at the problems of difficult field harvest, high labor intensity and high production cost of domestic forage ramie, a 4QM-4.0 fibre crops green fodder combine harvester was designed. The machine adopted a crawler-type walking device， which can complete field harvesting，conveying， shredding， collecting and automatic unloading at one time. Through theoretical analysis，the reel and horizontal cutting header，floating clamping device， shredding device and unloading device and other key components were designed. And the influence of drum diameter and speed on the formation of ramie fiber winding was analyzed. Performance and field texts showed that all working parts of the harvester operated stably， the work adaptability and reliability are good. The field test results show that the total loss rate of the harvester is 3%， the standard grass length rate is 91%， and the hourly productivity is 0.31 hm²/h， which meet the design requirements and the DG/T052—2019 green fodder harvester standard， and can meet the requirements of the feeder.