To solve the lateral stability control problem of buses under high-speed and large-steering emergency maneuvers, the lateral stability analysis and optimal control method for buses based on piecewise affine (PWA) identification of tire cornering characteristics was proposed. To address the contradiction between the high-precision modeling of tire cornering characteristics and the real-time system analysis and control, the PWA identification method was employed to model the data-driven nonlinear tire cornering mechanical characteristics. Subsequently, the eight-degree-of-freedom lateral dynamics model of the bus with incorporated roll, yaw, longitudinal and lateral motions was established. Based on the saddle-node bifurcation theory of nonlinear systems, the stability region boundaries of the bus phase plane defined by sideslip angle and yaw rate were constructed. The evolution patterns of the lateral stability boundaries influenced by vehicle speed, front wheel steering angle and road adhesion coefficient were revealed for enabling precise quantification of the bus lateral instability degree. Integrating the nonsingular terminal sliding mode control algorithm, the direct yaw control strategy based on lateral instability degree was designed, and the distribution rules for differential braking torque were determined. The simulation analyses of the bus lateral stability control performance were conducted. The results indicate that under the step steering conditions, by the proposed control method, the peak sideslip angle is approximately reduced by 52%-70%, and the yaw rate oscillation amplitude is approximately decreased by 41%-83%. Under the sinusoidal steering conditions, the peak sideslip angle is reduced by 15%-16%, and the yaw rate oscillation amplitude is reduced by 41%-61%.

 To solve the problems of low detection accuracy, poor real time performance and difficulty for detecting small target vehicles in existing vehicle detection algorithms under limited computational power of vehicle hardware, the improved vehicle detection algorithm was proposed based on YOLOv8. The omni dimensional dynamic convolution (ODConv) was used to reconstruct the Backbone network to improve the ability of extracting local image features. The gather and distribute (GD) mechanism was introduced into the Neck part of the model to effectively collect and fuse shallow details and deep semantic information for enhancing the multiscale feature fusion ability of the model and improving the detection accuracy of small target vehicles. Slide Loss and MPDIoU Loss were used to optimize the loss function of the original algorithm for making the distribution of positive and negative samples more uniform and improving the convergence speed and accuracy of the model. The experimental results on the KITTI dataset and BDD100K dataset show that the running speed of the algorithm reaches 71 fps with the F1-score improved by 1.4% and 0.9% and the mAP50 improved by 0.5% and 1.2% on the KITTI and BDD100K datasets, respectively. The experiments on small target vehicle datasets show that the mAP@50:95 is improved by 0.7%. Although the improvements are relatively modest, the proposed method achieves stable accuracy gains under the challenging conditions of severe vehicle occlusion in the KITTI dataset and large target scale variation in the BDD100K dataset with maintaining real time detection speed, which has practical value for vehicle mounted platforms with limited computational resources.

To investigate the impact of home-to-school travel behavior on the physical and mental health of Chinese junior high school students and identify the causal relationships, the ordered logit regression and the double machine learning (DML) methods were employed based on the data of 2014-2015 school year of China education panel survey. The associations of home-to-school travel mode and duration with negative mood, self-reported body size, self-reported health, illness frequency and sleep duration were examined. The results show that compared with the private car travel, the active home-to-school travel of walking or cycling can significantly reduce negative mood and improve body size with prolonged sleep duration. The travel duration exerts more pronounced impact on health, and the long travel with travel duration more than 20 minutes exacerbates negative mood and shortens sleep duration. The causal inference via DML confirms the robust causal effect of travel duration on negative mood and sleep duration, while the causal impact of travel mode is relatively weak. The Lasso, random forest, neural network and best models all verify the reliability of these findings.

The monocular ranging method for forward vehicles in intelligent driving was proposed. By the proposed method， the vehicle distance could be measured more accurately under the condition of single camera， and the camera attitude could be estimated in real-time. The estimation of camera pitch and yaw angle was achieved by road vanishing points. The images collected by the camera were sequentially processed through the Roberts operator amplitude calculation, feature point extraction, feature line segment generation, road vanishing point voting and estimation of camera attitude to obtain the pitch and yaw angles. The distance estimation network was designed and divided into multiple levels based on image size, and the image feature was incorporated with integrating vehicle grounding point and vehicle width information for effectively improving ranging accuracy. The validation was conducted on KITTI dataset, TuSimple dataset and continuous driving scenarios. The results show that the relative error of AbsRel is 8.3%, and the performance is improved compared to those by the previous algorithms.

 To realize the diagnosis and classification of multiple types of maize leaf diseases, the image classification algorithm  of combining CNN and Transformer(ICNN-Transformer) was used to classify various maize leaf diseases. The Transformer embedding CNN was adopted, and the RReLU activation function was used to replace the ReLU activation function for importing the pre-training model through transfer learning. The 3×3 conv in bottlenecks was replaced by pyramid convolution, and the dilated convolution with different dilation rate was introduced into pyramid convolution. To construct the new feature map of combining deep and shallow information of network, the upsampling fusion operation on feature maps of different sizes was adopted in the network. The results show that the combination improves the feature extraction ability of the model, and the accuracy of disease classification and the robustness of algorithm are improved. Compared with CNN (ResNet) and Transformer (Vit), by the proposed ICNN-Transformer，the various performance indicators are improved， and the accuracies are increased by 3.97% and 4.50%，respectively. The anti-interference ability of ICNN-Transformer in complex environments of fog, rain and dark is significantly improved. The experiment results show that different numbers and positions of Transformer embeddings can affect the classification accuracy of the model, which enhances the feature extraction ability of model and improves the disease classification accuracy and algorithm robustness. A better intelligent classification method for maize leaf disease diagnosis is provided.

To solve the positioning problem of rotary table type silage wrapping machine with large rotational inertia for difficult stopping accurately at the specified phase during operation, the coupling relationship among the weights of round hay bales, the rotating speed of rotary table and the angle of positioning error was investigated, and a set of mechatronics hydraulic integrated control system was designed for the rotary table type wrapping machine. In the parameter learning module of the system，the weight of round bale and the rotary table speed were used as learning parameters, and the parking position control parameters were assigned by the regression program. The hydraulic oil replenishment buffer circuit and time variable were combined to make the control system realize the control effect equivalent to that of the proportional valve. The performance test of the control system was carried out with positioning stability and positioning accuracy as evaluation indices. The test results show that the rotational speed of rotary table has the most significant influence on the positioning accuracy. Under the low rotational speed conditions,the rotary table exhibits high positioning accuracy with low average positioning deviation rate (<0.000 1%). Under the high rotational speed conditions,the rotary table has low average deviation rate (<2%) for the test groups I and II with test deviation range (-6%, 2%). The average deviation rate of the rotary table under high rotational speed conditions is higher than that under low rotational speed conditions, but the higher repeatability (>96%) has higher positioning stability.

To meet the requirements raised in practical applications, the scheme of parallel operation of four diaphragm pumps was proposed. The internal flow of the pump unit was simulated by the computational fluid dynamics (CFD) method, the overlapping grid zero gap technique and the 6 degree of freedom (6DOF) equation. The effects of the reciprocating frequency and diaphragm deformation on flow characteristics and pump performance were investigated. The comprehensive comparison of the distributions of pressure and velocity in the pump chamber and the pulsation of flow rate was implemented. The results show that the most significant pressure drop in the pump unit occurs upstream and downstream of the ball valve, and it is increased with the increasing of reciprocating frequency and diaphragm deformation. The volumetric efficiency is decreased with the increasing of reciprocating frequency. Point sources and point sinks are evidenced in the flow field, which are developed with the increasing of reciprocating frequency. The increase in diaphragm deflection can induce vortices. Increasing the reciprocating frequency and diaphragm deflection can reduce the overall pulsation of flow rate. In comparison, the diaphragm deflection has  greater impact on the flow rate compared to the reciprocating frequency. The increase in the maximum deformation of the diaphragm causes the decrease of volumetric efficiency less than 3%, while changing the reciprocating frequency results in the decrease of 18%.

 To address the classification challenges of high-dimensional small-sample gene expression profile data, and solve the difficulty of conditional Wasserstein  generative adversarial network with gradient penalty (CWGAN-GP) in maintaining gradient consistency and ensuring sensitivity and stability of generated data in high-dimensional spaces, the likelihood-constrained CWGAN-GP data enhancement method (CWGAN-GP-LC) was proposed. To ensure the sensitivity of generated samples, the generator of CWGAN-GP was augmented with original data and minimal multivariate noise. The latent space was partitioned by the joint probability distribution of generated and original samples for constraining the space of generator. The likelihood between generated and original samples was used to judge the authenticity for shortening the time to achieve the high-quality synthetic data. The basic CGAN and CWGAN-GP methods were introduced, and the generation space constraint and sample selection strategy of CWGAN-GP-LC were provided. The experiments of Wasserstein distance, sample distribution and classification accuracy were conducted on six gene expression profile datasets of Colon, DLBCL, MLL, SRBCT, CNS and Leukemia2, which were compared with those by CGAN, CWGAN-GP, CLSGAN-GP and Gene-CWGAN methods. The results show that on the Colon dataset, the combination of CWGAN-GP-LC with MLP classifier achieves classification accuracy of 91.12%±3.26%, outperforming that of 87.24%±5.36% of Gene-CWGAN. On the MLL dataset, the accuracy reaches 99.12%±2.16%, and on the Leukemia2 dataset, the accuracy reaches 99.72%±3.35%. The standard deviations are generally lower than those of comparative methods, which illuminates that the data generated by the proposed method exhibit higher sensitivity and stability.

To solve the problems of the difficulty in distinguishing actions from background due to dynamic drift of background features and the prediction uncertainty caused by lack of temporal boundary supervision in weakly-supervised temporal action localization, the weakly-supervised temporal action localization method based on background drift constraint and deep evidence constraint was proposed. The context attention branch was introduced to establish the background drift feature attention learning constraint mechanism, and the attention modeling on action start and end segments was performed to optimize the feature representations of actions, context and background in video segments. The deep evidential learning framework for actions, background and context was designed to correct the prediction results of the three semantic categories and action segment localization decisions. The experiments were completed on the THUMOS14 and ActivityNet1.3 dataset. The results show that by the proposed method, mAP value reaches 43.5% on THUMOS14 with averaged over t-IoU thresholds from 0.10 to 0.70， and mAP value reaches 24.9% on ActivityNet1.3 with averaged over t-IoU thresholds from 0.50 to 0.95. The results demonstrate the effectiveness of the proposed method.

A radar dynamic gesture recognition method based on multi-dimensional spatio-temporal features was proposed to solve the problems of radar gesture recognition with insufficient representation ability of single feature, high overhead of traditional image conversion and inadequate utilization of temporal correlation of dynamic gestures. The micro-Doppler characteristics were utilized, and the range, velocity and angle features of gestures were estimated through signal preprocessing. The multi-dimensional spatio-temporal features including range-Doppler-time (RDT) and angle-range-time (ART) were constructed, and the feature concatenation and fusion were completed. The hybrid ResNet18-LSTM network was designed to jointly extract the spatial features and temporal correlation features of gestures. The dedicated action detection algorithm was proposed to realize automatic acquisition and detection of single-group gesture data only by radar sensors. The multi-dimensional spatio-temporal features and the proposed network were verified through comparative experiments and ablation experiments, and the real-time performance of the system was evaluated via online testing. The results show that by new method, the test accuracy reaches 98.96% on the collected dataset, and the recognition accuracy achieves 97.0% in the constructed real-time system. The multi-dimensional spatio-temporal features can effectively improve the performance of dynamic gesture recognition, and the proposed model and detection algorithm exhibit good practicability and reliability.

To address the issue of lithium plating during fast charging of lithium-ion batteries (LIBs), the electrochemical-thermal coupling model was developed by COMSOL, and the accuracy was validated through low-temperature lithium plating experiments. On this basis, the MATLAB & COMSOL co-simulation platform was established, and 5-stage constant-current charging strategy was adopted. The elitist nondominated sorting genetic algorithm was employed to optimize the fast charging protocol without lithium plating. The results show that by the optimized charging scheme, the LIB can be charged to 80% state of charge within 2 377 s without lithium plating. Compared with the average charging rate scheme, the optimized scheme reduces lithium plating by 107.200 0 C (Coulomb) in single charging process and enables more capacity to be charged before the cut-off voltage.

Taking a four-cylinder diesel engine as research object, the effect of methanol blending ratio on in-cylinder combustion roughness under methanol/biodiesel blended fuel combustion conditions was systematically investigated. The particulate samples of blended fuels were collected at different positions of the exhaust pipe. Combined with transmission electron microscopy and particle size spectrometer tests, the evolutionary characteristics of particulate morphology during the exhaust process were analyzed. The results show that with the increasing of methanol blending ratio, the peak in-cylinder pressure is decreased, while the peak rate of pressure rise is increased, which can aggravate the in-cylinder combustion roughness. Under the maximum torque condition of the diesel engine, as the distance between the sampling point and the exhaust outlet is increased, the primary carbon particles exhibit obvious overlapping and stacking phenomena. The proportion of branched particles in the nucleation mode is decreased, whereas the proportion of clustered particles in the accumulation mode is increased gradually. The peak particle size and average particle size are shifted toward larger values, leading to the increased proportion of accumulation-mode particles.

Based on the experimental platform of electrically heated fluidized bed, the co-combustion experiments of furfural residue(FR) and furfural wastewater(FW) were conducted. The effects of blending ratio and fluidizing velocity on furnace temperature and CO concentration were investigated, and the coupled influence of fuel moisture content, blending ratio and fluidizing velocity on SO2 and NOx emission characteristics was also explored. Additionally, the ash deposition behavior of combustion products was analyzed. The results show that the CO emission is decreased with the increasing of FW blending ratio. The SO2 and NOx emissions from FR combustion alone are higher than those from FW combustion alone. Increasing the FW blending ratio significantly reduces both SO2 and NOx emission with measured values lower than theoretical predictions, confirming the synergistic inhibition effect between the two fuels. As fluidizing velocity is increased, the SO2 emission is decreased, while the NOx emission is increased. When the FW blending ratio reaches 50%, the noticeable ash deposition occurs in the furnace. X-ray diffraction analysis reveals that the deposits contain low-melting-point eutectic compounds of potassium chloride and sodium sulfate, which can pose the risks of slagging and corrosion.

 To obtain the manufactured sand mortar with good workability and high early-strength, the performance tests were sequentially conducted on cement paste and manufactured sand mortar. The effects of the water-to-cement ratio, the proportion of sulfoaluminate cement replacing ordinary Portland cement in composite system and the dosage of aluminum sulfate as early-strength agent on the workability and mechanical properties of paste and mortar were investigated. The experimental results indicate that increasing the amount of sulfoaluminate cement in the composite system shortens the setting time of paste with reduced fluidity and increases the fluidity loss over time. To ensure good workability, the dosage of sulfoaluminate cement should be controlled within the range of 10% to 20% of the total cement content. Under the same water-to-cement ratio, the low dosage of sulfoaluminate cement can enhance the compressive strength of the mortar at various curing ages. The incorporation of aluminum sulfate improves the compressive strength of mortar with the most significant strength increasing observed at the dosage of 0.75%. The optimal mix proportion for early-strength manufactured sand mortar is determined with water-to-cement ratio of 0.45, sulfoaluminate cement content of 10% of the total cement and early-strength agent dosage of 0.75%.

 To improve the service performance and expand the application scope of conventional emulsified asphalt binder, the waterborne epoxy resin styrene butadiene rubber (WER-SBR) composite modified emulsified asphalt with various dosages was prepared by the emulsification with later modification process. Three preparation methods for emulsified asphalt evaporation residues of high temperature evaporation, room temperature evaporation and low temperature evaporation were compared, and the optimized preparation method was proposed. The high  and low temperature rheological properties were characterized by dynamic shear rheometer (DSR) and bending beam rheometer (BBR). The synergistic mechanism between modifiers and emulsified asphalt was revealed via scanning electron microscopy (SEM) and fluorescence microscopy (FM). The results indicate that WER significantly increases the complex shear modulus and rutting factor of emulsified asphalt and remarkably enhances the high temperature stability with the optimal dosage around 15%. SBR latex effectively reduces the stiffness modulus and increases the creep rate, which significantly improves the low temperature cracking resistance with the optimal dosage around 4%. Microstructural observations confirm that the three-dimensional network cross-linked structure formed by WER and SBR in emulsified asphalt is the key mechanism for improving rheological performance.