To solve the adaptability problem of adaptive cruise control system for different styles of drivers, the personalized multi-mode adaptive cruise control algorithm was proposed. Based on the next generation simulation(NGSIM) data set, the car-following data were clustered and analyzed by K-means algorithm. The driver driving style was divided into radical, general and conservative types, and the corresponding car-following distance model was constructed. The gradient algorithm based on the deep deterministic policy was designed to modularize the basic performance of adaptive cruise, and the personalized car-following distance model was integrated into the basic performance of each mode. Based on the maximum entropy inverse reinforcement learning, the weights of personalization before the reward function were designed. To verify the multi-mode adaptive cruise control, the system was run in Simulink/Carsim joint simulation environment, and the results were compared with real vehicle samples of three driving styles. The results show that the tracking performance is good, and the inter distance and speed are close to those of the real samples of three types of drivers, which conforms to the driving habits of drivers and meets the personalized driving needs.

To solve the problems of discontinuity in path curvature, low parking efficiency and low path tracking accuracy in autonomous parallel parking, the characteristics of the circular arcstraight linecircular arc initial parking path were analyzed, and the path planning method based on  fifthorder polynomial curve was adopted. To achieve the balance between path length and path curvature performance, based on the constraints of maximum path curvature, required parking space and obstacle avoidance, the weighted sum of the maximum curvature and the horizontal coordinate of parking start point was used as objective function. The particle swarm optimization algorithm with nonlinear dynamic adaptive inertia weight was employed to optimize the horizontal coordinate of the parking start point, and after optimization, the path became smooth and continuous in curvature. The path tracking control method based on model predictive control was established, and the genetic algorithm was used to optimize the prediction horizon and control horizon for reducing computational load while ensuring tracking accuracy. The real vehicle validation was conducted based on the autonomous driving developer kit of Baidu Apollo. The results show that the vehicle parking can be completed safely without collisions, which verifies the feasibility of the path planning method. Under the premise of reduced computation load, the average path tracking error is decreased by 4.348% compared to that before optimization, and more high tracking accuracy can be obtained by the proposed path planning method.

 To solve the problem of the existing underwater detection with low detection accuracy due to single variety and dense target, the improved YOLOv5 algorithm of RSC-YOLOv5 was proposed. The RepVGG Block module was used on Backbone to improve the recognition accuracy of targets of different scales and enhance the inference speed. The shuffle attention （SA）module was added to improve the feature extraction ability of the algorithm. Content-aware reassembly of features(CARAFE) upsampling was used in Neck to obtain larger receptive field. Varifocal Loss was introduced to pay more attention to the high-quality positive samples in intensive target sample training. The experimental results show that the average accuracy of the RSC-YOLOv5 fish and crab detection algorithm is 93.6%, which verifies that the proposed algorithm is suitable for the underwater fish and crab detection.

To improve the anti-wear and anti-fatigue ability of raceway surface of the shield main bearing, SYSWELD software was used to carry out the numerical simulation of laser transformation hardening under different parameters for the shield main bearing raceway material of 42CrMo steel. The effects of laser power, scanning speed and spot size on the phase transformation of 42CrMo steel for the raceway material of shield main bearing and the depth of hardened layer were investigated. The results show that with the increasing of laser power, the decreasing of laser scanning speed or the decreasing of laser spot size，the depth of raceway hardened layer is increased. When the laser power is 2 250 W with scanning speed of 20 mm/s and spot length of 5 mm, the depth of the hardened layer of the raceway can reach 1.08 mm, and the hardened layer is relatively uniform. 

To reduce the temperature of heating wall and improve the uniformity of temperature distribution, the flow boiling heat transfer process of HFE-7100 in the wavy microchannel was numerically simulated, and the effects of microchannel structure and densification methods on the flow boiling characteristics were discussed. The results show that when the waviness throughout the whole flow direction is densified with increasing the ratio of wave amplitude to wave pitch from 0.04 to 0.08, the equivalent diameter of bubble at the outlet of microchannel is decreased to some extent，and the segmental annular flow is also suppressed. However, the pressure loss between the outlet and inlet of microchannel is increased significantly with the maximum increasing up to 140.54%. Increasing the waviness of the upstream, midstream or downstream in the microchannel can effectively inhibit the segmental annular flow from occurring at the outlet section and reduce the heating wall temperature for improving the temperature distribution uniformity of heating wall. The comparison illuminates that increasing the downstream waviness can more significantly reduce the bubble growth rate. The heating wall temperature and distribution can be effectively controlled as well, which is equivalent to that of the whole densified microchannel. However, the pressure loss between the outlet and inlet can be reduced by 34.23%. Increasing the downstream waviness in the wavy microchannel is  preferable method for improving the flow boiling heat transfer characteristics.

 When the vehicle works normally, the internal parts of transmission are often damaged due to high temperature caused by shifting and other behaviors, which affects the normal driving of the vehicle. Therefore, it is necessary to study the transient temperature field of the transmission. According to the thermal load characteristics of transmission components, the transient thermal network model of transmission was established by thermal network method in the forms of heat source and thermal resistance. The temperature of each transient node in the transmission was calculated, and the temperature rise of high-heat parts of wet clutch and planetary wheel during the whole process of vehicle starting to driving smoothly was analyzed. The results show that the temperature rise law of wet clutch is related to the gear shift. The temperature rise law of the gear is changed with the change of shift position. Temperature rise of parts in planetary gear is basically the same and changes steadily.

The video anomaly detection method based on fusion of attention with convolutional autoencoders was proposed to address the insufficient utilization of contextual semantic information in video sequences. The frame prediction strategy was adopted, and the convolutional autoencoder based on the Inception module was employed to extract the multi-scale features from input video sequences. To capture the interaction between moving objects and static backgrounds, the position attention and channel attention were incorporated. The memory enhancement module was integrated into the convolutional autoencoder to constrain generalization, and the latent loss function was introduced to enlarge the reconstruction errors of anomalous events. The calculation of anomaly scores was derived, and the dataset for anomaly detection was provided. To verify the proposed method, the qualitative analyses of abnormal behaviors, the model performance comparisons, the memory item update threshold experiments and the ablation studies were conducted. The results show that the proposed method can effectively detect the anomalies in videos with high detection accuracy. The AUC values on the UCSD Ped2, CUHK Avenue and ShanghaiTech datasets are 97.7%, 88.9% and 73.8%, respectively.

To provide more reliable safety recommendations for vehicles navigating curves, the lane curvature measurement method based on binocular stereo vision was investigated. The binocular camera was calibrated by the MATLAB camera calibration toolbox to obtain intrinsic and extrinsic parameters. The original left and right images from the binocular camera were input into the stereo matching network to generate disparity map. The Sobel edge detection and threshold detection were applied to the region of interest (ROI) in the original left camera image, and the results were fused. The perspective transformation was applied to the fused results within the ROI to obtain binary image. The positions and quantities of white pixels within the ROI were statistically analyzed by the histogram. The lane lines were fitted by the least squares method, and the fitting accuracy was enhanced by Kalman filter prediction algorithm. The pixel coordinates of the lane lines were determined based on the inverse affine transformation and the fitting results. The disparity values within the disparity map were obtained according to the pixel coordinates, and the three-dimensional coordinates of the lane lines relative to the main camera of the binocular system were derived by triangulation. The distance conversion scale between the measured real-world coordinates and the image pixel coordinates was calculated and applied to the lane curvature computation. The results show that by the proposed algorithm integrating Kalman filter prediction for lane line detection and binocular stereo vision for lane curvature radius measurement, the average deviation of 5.98% in the calculated curvature radius of curves is achieved, which can ensure the robustness of lane line fitting for enhancing the accuracy and reliability of the binocular stereo vision-based lane curvature measurement method.

To improve the driving safety, the fault diagnosis model integrating local correlation constraints and signed directed graph (SDG) was proposed to address the challenge of fault localization in automatic emergency braking (AEB) systems. Taking the AEB system based on the time-to-collision (TTC) control strategy as research object, the SDG model of  braking system was constructed by combining vehicle dynamics model, brake pressure model and sensor system layout. To verify the proposed model, the cosimulation of Carsim and Simulink was conducted. The results demonstrate that the SDG model based on graph theory offers strong interpretability and effectively reflects fault propagation paths. Through the cosimulation tests of Carsim and Simulink, the fault diagnosis and localization during the active braking process can be efficiently achieved. By the propose method, the abnormal internal connection nodes can be detected within approximately 10 diagnostic function computations with accurate fault localization and propagation judgment.

Using salt and tap water as raw materials, the hypochlorite solution was produced by electrolysis for using as substitute disinfectant in public places, and the substitute disinfectant was integrated into local synthesis spray sterilizer. The device was composed of two parts with pneumatic electrostatic atomization unit and disinfectant local synthesis unit. The effective operating distance was 6 m with unit time spray volume of 303 mL/min. In the E. coli test, the ORP value of the test group was high, and the change range was small when the salt water with mass fraction of  2% was reacted for 3 min. Compared to the pure water control group, the sterilization rate of the selected test group was 100.00%, and sterilization effect was significant.The spray performance test was conducted to analyze the spray performance under different conditions by changing the air pressure with or without  charging.The results show that as the air pressure is changed, the amount of spray per unit time and the spraying width and distance are changed accordingly. At the air pressure of 2.5×105 Pa, the spray volume, spray distance and spray range per unit time are 303 mL/min, 6.00 m and 1.25 m, respectively. In the charged state, the spray amplitude, deposition density and deposition amount are larger than those in the same condition. At the air pressure of 2.0×105 Pa, the maximum spray amplitude is 1.21 m, and the spray amplitude is increased by 0.20 m compared with that at noncharged condition. The deposition density of droplets can be increased by charging, and the deposition amount is increased by 0.08 μL/cm2 at the upper position of 3.70 m. At the air pressure of 2.5×105 Pa, the average particle size of droplets is 47.93 μm. The experimental results show that the total number of bacteria is less than or equal to 300 CFU/（25 cm2） when the developed substitute disinfectant local synthesis spray disinfection device is used in laboratories and public places, which meets the national standard.

The intake flow rate and pressure in the air supply system of proton exchange membrane fuel cells(PEMFC) are two coupled control variables, and the coupling affects the stability of the fuel cell system. According to the decoupling control requirements of intake air flow and pressure, the system openloop response experiment was designed, and the system identification method was used to obtain the model parameters of the system transfer function matrix. On this basis, the double closedloop PI, feedforward decoupling and fuzzy decoupling control strategies were designed. The decoupling effects under different control strategies and the antidisturbance abilities of different control strategies to system parameter changes were compared and analyzed. The simulation results show that the coupling effect between flow rate and pressure is significant under dual closedloop PI control. By the feedforward decoupling control, the complete decoupling can be achieved, but the decoupling effect is affected by the parameters of the controlled object model. The fuzzy decoupling control has good decoupling effect and stronger adaptability to parameter changes compared to conventional feedforward decoupling control, which has better application prospect in the multivariable strong coupling system control.

To quickly and accurately detect the feature information of track cone buckets in the formula unmanned competition, the detection method fusing LiDAR point cloud data and camera image data was proposed. The region of interest (ROI) was extracted from the point cloud data, and the noise was filtered.The adaptive ground segmentation algorithm was introduced for the current ground undersegmentation and oversegmentation problems, and Euclidean clustering was used to extract the spatial location information of cone buckets. The image dataset was collected and labeled, and the posttraining detection was performed by YOLOv5 to complete the extraction of cone bucket color information. The bucket spatial position information detected by LiDAR and the bucket color information detected by camera were fused.The experiments were designed for validation. The results show that the proposed fusion detection method can solve the problems caused by single sensor detection of cone buckets, and it can quickly and accurately detect the cone buckets in the track, which provides guarantee for planning and decision making.

 To solve the problem of traditional Ethereum phishing scam identification and classification with high computational memory cost due to overlooking the importance of inter-subgraph relationships, the hierarchical graph attention framework was proposed to process subgraph classification tasks by leveraging graph attention technology for extracting behavioral patterns of account addresses. The hierarchical graph attention pooling encoder was constructed, and the node-level encoder was used to extract intra-subgraph node importance, while the subgraph-level encoder was employed to capture inter-subgraph significance for revealing potential associations both within and between subgraphs. Combined with graph contrastive learning techniques for joint training, the contrastive learning loss was introduced as regularize to alleviate label sparsity for improving subgraph classification performance. Comparative and ablation experiments were conducted on real Ethereum datasets, and the parameter analysis was carried out with F1 score as evaluation metric. The experimental results show that by the proposed method, the maximum improvement of 1.7 percentage points in F1 score is achieved on the real dataset with outperforming classical models of GCN, GraphSage and GAT, which requires less memory than other node classification approaches.

Based on the regulations of GB 13057—2023 "Strength of Seats and Their Anchorages of Passenger Vehicles", the optimization design was conducted for the anchorages of a certain bus seat. The side length of square tube, the wall thickness of square tube, the thickness of end cap, the thickness of side reinforcement component and the thickness of thinwalled parts used for upper and lower connections were selected as design variables. The adaptive Kriging (AKR) agent model for the seat forward displacement was established. The mass of the vehicle anchorage was calculated by the Catia parametric design method. With seat forward displacement and anchorage mass as optimization objectives, the multiobjective optimization of seat anchorage dimensions was performed by the particle swarm optimization (PSO) algorithm. The optimization results show that the seat forward displacement is reduced to 443.22 mm, which is decreased by 7.85% compared to that before optimization and can significantly enhance the seat rigidity for providing greater safety for occupants. The mass of the vehicle anchorage is reduced by 11.08% for achieving the goal of lightweight design.

To solve the problem of huge cost caused by the collision of agents with obstacles when the environmental information was unknown, the robust reinforcement path planning algorithm based on Bayesian optimization and quadrature was proposed. The grid map of the environment was established for setting environmental rewards. Bayesian optimization was used to establish Gaussian Process surrogate model for historical action sets, and the mean and variance of rewards were estimated by Gaussian Process. The upper confident bound (UCB) method was selected to balance exploration and exploitation, and the action sequence was selected to avoid over-exploration and over-exploitation. Bayesian quadrature method was used to actively learn the environment, and the uncertainty of environmental information was reduced by minimizing the expected posterior variance of Q value for avoiding collision and improving robustness. The Q table was updated iteratively, and the Q learning method was used to plan the path. The simulation experiment was carried out to compare the proposed algorithm with the classical Q-learning path planning algorithm. The results show that compared to the classical algorithm, the proposed algorithm has higher learning efficiency for the environment, smaller collision probability, faster convergence speed of the optimal path steps and more effective path planning.

To solve the problems of poor generalization and defective fault dataset in the existing fault diagnosis research of vehicle steering system, the fault diagnosis method of intelligent steering system based on the random forest (RF) algorithm was proposed. The physical model of the intelligent steering system was built based on the joint simulation of Simscape and CarSim. The several critical faults of steering system were simulated, and the fault datasets under various driving conditions were collected. The fault diagnosis model based on RF algorithm was established. The input data was classified to realize the fault diagnosis of intelligent steering system, and the proposed method was compared with several typical algorithms. The steering system bench was built to collect the imbalance sample data set for further validation. The experimental results show that using the RF algorithm for fault diagnosis of the simulated fault dataset, the fault diagnosis accuracy rate is approximate 87.43%. In the experimental verification, the fault diagnosis accuracy rate is 99.93% with fast diagnosis speed and good generalization.

The direct data-driven control problem of an unknown class of Lorenz-type chaotic systems was investigated. The systems were transformed into linearly dependent form to construct the non-parameterized model. The state feedback controller was designed for the reconstructed system. The input and output data of the system were sampled offline, and the collected data were directly used to design and give the expression of the state feedback controller. The stabilization problem of the system under state feedback control was verified,and the stability of the closed-loop system was proved. Through numerical example simulation verification, the results show that the proposed direct data-driven control strategy is effective.

 The behaviors of adjacent lane target vehicles cutting into the own lane were investigated during the commercial autonomous vehicles cruising or following in the current lane. By the lateral safety distance threshold method, the lateral entry behavior of the target vehicle was predicted and identified. Four longitudinal distance models of active fuzzy alternative safety measurement, responsibility sensitive safety model, headway and collision time were investigated and simulated. The results of simulation testing and open source data environment testing show that the active fuzzy replacement safety measurement model can fully utilize the vehicle motion performance limitations and the road environment information transmitted by the perception module, which ensures driving safety with relatively small impact on traffic flow. The algorithm with entry recognition can adopt gentle braking before the vehicle enters the self driving lane for avoiding the dangerous behavior of emergency braking in the original algorithm and improving the driving safety and riding comfort of autonomous driving.

To solve the problems of high false detection rate of interferential point clouds and high missing detection rate of distant sparse point clouds in automatic driving scene, the obstacle point cloud detection algorithm based on improved PointPillars was proposed. The point cloud in pillars was encoded by the aggregation module and shared multilayer perceptron（MLP）. The salient and detailed features were mapped into pseudo-image features by stacking the maxpooling and average-pooling. To solve the problem of pseudoimage feature with insufficient attention and utilization, the deep and shallow feature maps were fused by attention and residual second block（ARSB） module to optimize the gradient and enhance the coordinate attention(CA) to effective targets. The results show that the improved algorithm has high detection accuracy for global point clouds. The detection precision of the improved algorithm is better than those of the classical 3D detection algorithms of PointPillars and STD methods, especially for the detection of car category. The detection speed is fast, which meets the requirements of realtime. 

To quickly construct small sample dialect corpus dataset and improve the performance of Chinese dialect speech recognition, the dialect speech recognition method based on few-shot learning was investigated. Based on the selected high-quality public corpus, the dialect data set was constructed by extraction, correction, supplementation and recording. Combining the complementary characteristics of connectionist temporal classification network and attention architecture and the supplementary role of language model to end-to-end speech recognition model, the end-to-end speech recognition model based on hybrid CTC/Attention with additional language model was constructed. On this basis, the few-shot learning method based on data augmentation and transfer learning was used to complete the training of the dialect speech recognition model. The experiments were completed based on the Guanzhong dialect corpus. The results show that the fine-tuning speech recognition base model with character error rate of 4.9% can result in the dialect model with character error rate as low as 6.9% on the test set. The construction scheme of small sample dialect corpus is feasible, and the dialect speech recognition scheme based on small sample learning is effective. According to the cross-validation test, the character error rate of Mandarin in the test set of the fine-tuned dialect model is 27.2%, indicating that the fine-tuned model supports cross-lingual recognition. 

In response to the national "double carbon" plan and to explore the potential of new carbon-free fuels in micro-power systems, the experiments were conducted to systematically investigate the flame characteristics of ammonia/hydrogen/air premixed flames in planar micro-combustor. The effects of operating parameters of flow rate and equivalence ratio on flame stability and combustor working performance were assessed. The formation mechanisms and transition laws of different flame structures were discussed in detail according to the internal flow heat transfer and external heat dissipation. The results show that as the flow rate is increased, the flame patterns appear repetitive extinction and ignition, planar, anchored, U-shaped and inclined flames in sequence. The velocity ranges of various types of stable and unstable flames are also determined. The flame stability at stoichiometric equivalent ratio is optimal when the equivalent ratio is 1.0, and the blowout limit can reach 0.35 m/s. The overall output characteristics of U-shaped flames are superior compared to other flame structures, and the maximum wall temperature can reach 900.0 K with the minimum wall temperature difference.

To realize the high-precision motion trajectory tracking control of pneumatic actuators, the proportional directional valve-controlled pneumatic cylinder system was taken as research object, and the robust controller with online adaptive parameters was designed based on the  control strategy. The proportional directional valve flow model, the cylinder kinematic model, the cylinder friction model and the cylinder thermodynamic model of system were established. The robust controller was designed based on the backstepping method to suppress the effects of parameter estimation errors, uncertainty nonlinearities and external disturbances of system, and the online parameter adaptive rate was designed for the controller to effectively estimate the parameter uncertainty term in the system model in real time. The real-time control system for adaptive robust tracking control of the pneumatic motion trajectory was built by the XPC-Target module in MATLAB/Simulink. The experimental results show that by the designed controller， good trajectory tracking control performance is achieved. The maximum tracking error is 0.80 mm for tracking 0.4 Hz sinusoidal trajectory, which is about 2.67% of the amplitude.

To address the inefficiency of existing Fine-Tuning methods in handling minor visual variations, the transfer reinforcement learning model based on semantic selection was proposed. Inspired by the inattentional blindness that irrelevant background stimuli were overlooked during attentional shifts, the model was designed to decouple visual task transfer from policy control tasks through unsupervised semantic segmentation and semantic weight selection for preserving only policy-relevant visual features. The Flappy Bird game variant was utilized as experimental environment, and the transfer tests were conducted under varying visual interference conditions. The multiple soft attention mechanisms were compared in terms of transfer performance. The results show that the proposed method outperforms attention-based transfer approaches in efficiency, interpretability and adaptability to complex backgrounds. The significantly enhanced robustness is demonstrated, particularly in environments with substantial visual interference.

To solve the problems of high common mode voltage (CMV) and severely conducted electromagnetic interference in five-phase permanent magnet synchronous motor (PMSM) with near-four-vectors-based space vector pulse width modulation (SVPWM), the novel CMV-reduction-based spread-spectrum SVPWM strategy was proposed. To reduce CMV, according to the CMV generation mechanism of the five-phase PMSM driven by voltage source inverter, the vectors in opposite direction were employed to replace the zero vector in the near-four-vectors-based SVPWM strategy. To restrain current ripple and conducted electromagnetic interference, the frequency of SVPWM was adjusted on-line based on the mechanism of current ripple and high-frequency harmonic distribution. The five-phase PMSM experimental platform was constructed to compare the proposed RCMV-3SVPWM strategy with the conventional NFV-SVPWM strategy, and the dynamic tracking performance tests were carried out on the five-phase PMSM. The results show that by the proposed strategy, the CMV can be reduced by 40% with decreased current ripple of 0.30 A, and the conducted electromagnetic interference is effectively suppressed with maintaining good steady-state and dynamic performance.

To investigate the formation mechanism of central sagging in glass during the preheating stage of the glass forming process and the effects of two geometric parameters of glass thickness and aspect ratio, the numerical model of the fluid-thermal-structural coupled multiphysics was established based on the process flow of industrial preheating furnace. The high-temperature infrared thermography was employed to obtain the surface temperature distribution of the glass at the end of the preheating stage for verifying the fluid-thermal coupled numerical model. The developed fluid-thermal-structural coupled multiphysics model was used to simulate the formation process of central sagging during preheating for elucidating the underlying deformation mechanism. The effects of the two geometric parameters on glass sagging deformation were systematically investigated. The results indicate that geometric parameters play a critical role in glass deformation. The thinner glass exhibits more severe central sagging due to the reduced bending stiffness. The aspect ratio significantly affects the deformation behavior. When the aspect ratio is decreased from 6.32 to 1.58, the sagging depth is increased from 4.65 mm to 9.34 mm, which is attributed to the expansion of thermal strain region in the transverse direction.

To solve the issue of point cloud data loss in reflective workpieces, the three-dimensional measurement method based on adaptive multiple exposures was investigated. The binocular structured light measurement system was designed, and the workpiece point cloud was generated by phase retrieval, phase-based stereo matching and three-dimensional point calculation. The initial sequence of exposure times was determined, and the corresponding initial images were acquired to conduct the statistical analysis of overexposure ratio in the initial images. The relationship between initial exposure time sequence and overexposure ratio was constructed for enabling the adaptive calculation of  final exposure time sequence for point cloud generation by the predetermined overexposure ratio. By fusing the multiple exposure images, the point cloud of the reflective workpiece was computed. The system accuracy was evaluated using stepped calibration block, and the experiments on aluminum guide plate and aluminum side plate were conducted. The results show that the average planar distance error of the binocular structured light measurement system is 0.126 9 mm. Compared with the single-exposure imaging method, the proposed approach using fixed overexposure ratio can generate more complete point clouds for different aluminum alloy workpieces with the number of valid points increasing by 55% and 42% for aluminum guide plate and aluminum side plate, respectively. The proposed method effectively enhances the completeness of point clouds for reflective workpieces.

 Traditional vegetation analysis based on remote sensing imagery is often characterized by substantial manual intervention with low efficiency, suboptimal accuracy, slow updates and high resource consumption. To address these limitations, the novel fully convolutional YOLO (F-YOLO) network architecture was proposed. Derived from the YOLO framework, the target detection accuracy was improved by modifying the loss function to deliver pixel-level classification outcomes. The automated end-to-end pixel segmentation algorithm was designed to quantify the urban vegetation coverage. The Zijin Mountain area in Nanjing was selected as the study region. The images were processed through convolutional and pooling operations by the proposed F-YOLO to extract feature maps. These highly abstract features were subsequently upsampled to original dimensions through bilinear interpolation and followed by per-pixel vegetation classification for segmentation and coverage calculation. The experimental results show that by the proposed method, the vegetation segmentation accuracy of 94.76% is achieved with the vegetation coverage estimation precision of 96.72%. For object detection, the recall rate of 97.33%, the precision of 96.34% and the F1-score of 96.83 are significantly higher than those of SSD and Faster R-CNN. In terms of pixel-level segmentation, the accuracy of 94.76% also outperforms those of FCN-16, U-Net and FCN-32, which confirms the overall superior performance of the proposed approach.

 Based on ABAQUS software, the secondary development of spatially mobilized plane(SMP)strength criteria was realized by external subroutine method. The saturated and unsaturated regions of slope were distinguished by Boolean variable, and the influence of soil moisture content variation on capillary cohesion, friction angle and soil unit weight was considered. The strength reduction method was employed to investigate the stability of three-dimensional unsaturated soil slope under different rainfall conditions. The pore water pressure distribution and the safety factor of slope were obtained. The results indicate that soil types and rainfall patterns influence the pore water pressure distribution and slope safety greatly. When the wetting front is disappeared, the impact degree from large to small is from delayed rainfall pattern, average rainfall pattern, normal rainfall pattern to advanced rainfall pattern. The slope stability is conservatively estimated when the three-dimensional effect and the middle principal stress are not considered in the slope stability analysis.

The internal flow characteristics of double-inlet nozzle are extremely complex and have significant impact on the cavitation occurrence capacity of jet. The distribution characteristics of jet velocity, the vapor-liquid two-phase and pressure during the generation process of artificially submerged cavitation jet were investigated, and the influence law of the outlet width of outer nozzle on the flow field was explored.The results show that when high-speed jet flows through the inner nozzle contraction section into the throat, the jet velocity is suddenly increased due to the contraction of the outflow section, and the high-speed low-pressure zone is formed at the throat and near the throat outlet. By the low-speed jet wrapping effect, the velocity distribution in the region eventually presents droplet shape, and the liquid phase in the region shows the phenomenon of first diversion with latter polymerization under the action of vapor phase. As the width of outer nozzle outlet is increased, the jet velocity attenuation is increased with decreased distribution range of vapor phase, and the maximum vapor phase volume fraction is decreased.