To improve the stability and accuracy of the sound quality evaluation model, the vehicle interior sound quality evaluation model(ALSTM) based on long short-term memory (LSTM) network and attention mechanism was proposed. The steady noise samples of different brands of vehicles at the right ear of drivers under different working conditions were collected, and the subjective evaluation test of noise samples was carried out with annoyance as evaluation index to establish the evaluation data set of interior sound quality. On the basis of the data set, the sound quality evaluation model based on LSTM network was constructed with Mel-scale frequency cepstral coefficient (MFCC) of noise samples as feature input, and the attention mechanism was introduced to optimize the model. The experimental results show that the proposed evaluation model of sound quality can effectively evaluate the vehicle interior noise, and the accuracy in the test set is as high as 97.07%. Compared with other methods, the stability, convergence speed and classification accuracy of the ALSTM model are improved.

To solve the problem of vehicle dynamics state variations in autonomous vehicles due to the uncertainty of road conditions on seacrossing bridges, the vehicle roll control strategy was proposed based on the deep deterministic policy gradient (DDPG) algorithm, and the generalization capability under different speeds was discussed. The vertical model of the seacrossing bridge was constructed to provide dynamic road environment. The vehicle dynamics and vehicle tracking error models were established for incorporating the dynamic characteristics of vehicle roll, sideslip and yaw and for establishing the criteria for roll stability. The state space and action space for the DDPG algorithm were designed, and the reward function was formulated based on the vehicle roll state. The numerical simulation results show that by the DDPG algorithm, good performance is achieved in each episode with robust learning and problemsolving capabilities in complex and uncertain environments. The vehicle roll angle and lateral distance error are ensured within acceptable and minor fluctuation ranges to achieve safe vehicle control.

 To solve the severe miss detection problem in pedestrian detection caused by insufficient pixel information of distant targets and occlusion induced loss of human pattern information, the pedestrian detection method was proposed based on dual key points combinations. The discriminative semantic features of pedestrians were effectively extracted and fused by utilizing key points of the head and center regions for significantly reducing the pedestrian miss detection rate. The deformable convolution was introduced into the deep aggregation backbone feature network to enlarge the receptive field and enhance the semantic information of human pattern. The dual branch joint detection module based on key points combinations was designed, and the positive samples for different branches were redefined to strengthen the semantic information of small scale and occluded targets. The results of the dual branch detection were fused using the non maximum suppression (NMS) algorithm. The results show that on the CityPerson validation dataset, the average miss detection rates of the normal, small scale and heavily occluded subsets reach 8.24%, 11.81% and 30.59%, respectively. Especially, for the heavily occluded subset, the miss detection rate is reduced by 15.71% compared to the traditional method ACSP. By the proposed method, the detection speed reaches 16 frames per second. On the CrowdHuman dataset, the average precision and average miss detection rate are 86.30% and 45.52%, respectively. Compared with other state of the art methods, the proposed method exhibits superior performance in average precision, miss detection rate and detection speed, which demonstrates significant application value in complex scenarios with dense pedestrian crowds.

To explore the characteristics of traffic volume data and improve the accuracy of parameter estimation for traffic volume data fitting, the parameter estimation method for traffic volume data fitting was proposed based on the maximum correlation entropy criterion. The traffic volume data was preprocessed, and the probability distribution of the data was specified. The maximum correlation entropy criterion was used to estimate the model parameters, and the gradient ascending method was used to output the model parameters. The estimated values of the probability distribution parameters were obtained, and the results were compared with those of the traditional methods. Based on the measured data and the four assumed distribution models of Normal distribution，Logarithmic normal distribution，Webull distribution and Rayleigh distribution, the performance evaluation was conducted.The results show that the proposed method has the best fitting performance, and the comprehensive evaluation indexes are respective 0.979 00, 0.726 08, 1.397 69 and 1.494 50, with strong accuracy and reliability.

Based on the real vehicle configuration driven by parallel hub motor, the main factors affecting the braking energy recovery efficiency were analyzed, and the Federal Kalman filter longitudinal speed estimation algorithm was used to propose the three-layer compound braking control strategy. By the braking decision-making layer, the braking conditions were identified and entered into the corresponding braking mode according to the pedal input and driving state. According to the decision, by the braking control layer, the electric braking torque distribution of the front and rear wheels was optimized through particle swarm optimization (PSO) under conventional braking conditions to maximize the effective battery recovery efficiency. Under emergency braking conditions, the fuzzy self-tuning PID algorithm was used to realize the safe and effective control of the slip rate of each wheel on different adhesion roads, and the robustness of the anti lock braking control was optimized with improved braking safety. The command of the control layer was responded by the braking executive layer, and the electric braking compensation mechanism was designed to quickly compensate and adjust the pressure error of each wheel cylinder for improving the braking stability of the vehicle. The control strategy was verified by the real vehicle.

The robust backstepping sliding mode RBF network adaptive controller was proposed for the quadrotor unmanned aerial vehicle(UAV) attitude system with disturbance. Based on the backstepping sliding mode control, the RBF network was used to approximate and compensate the ideal control law. The minimum parameter learning method of the neural network was adopted, and the weight upper bound of the neural network was estimated as estimated value of the neural network. The adaptation law was used to replace the adjustment of neural network weights, and Lyapunov theorem was used to prove the stability of system. The simulation results show that compared with the backstepping sliding mode control method, the proposed method has shorter adjustment time and better tracking accuracy in the case of disturbance. It is verified that the proposed method has better anti-interference and robustness.

To investigate the stress distribution and safety reliability of the baffle exposed to high temperature in heater, the stress of baffle was analyzed based on the finite element method. The dangerous section path was selected for the stress linearization analysis, and the structure of baffle was optimized according to the safety assessment results. The results show that the stress concentration caused by sharp angle has great influence on the stress distribution of   baffle, and the excessive film stress of PL is the main reason of stress assessment risk. The overall stress of baffle is increased with the increasing of fuel inlet velocity, and the temperature is positively correlated with the thermal stress. The stress change trend is similar at different speeds, and the structure discontinuity at the connection of heat exchange tube bundle and folding plate can cause stress increase. In the optimization scheme, the maximum thermal stress of baffle is reduced by 16.7% when the sharp angle is passivated. The maximum thermal stress with thermal barrier coating is decreased by 44.5% as compared with that for passive tip angle, and decreased by 53.7% as compared with that for the original structure. The overall thermal stress is decreased, and the strength of the optimized baffle plate is assessed as qualified.

 For the system log data with the distribution characteristics of "group anomaly" and "local anomaly", traditional semi-supervised log anomaly detection method of anomaly detection with partially observed anomalies(ADOA) has poor accuracy of pseudo-labels generated for unlabeled data. To solve the problem, the improved semi-supervised log anomaly detection model was proposed. The known abnormal samples were clustered by k-means, and the reconstruction errors of unlabeled samples were calculated by kernel principal component analysis. The comprehensive anomaly score of sample was calculated from reconstruction error and similarity to abnormal samples, which was used as pseudo-label. Sample weights for the LightGBM classifier were calculated based on pseudo-labels to train the anomaly detection model. The impact of the proportion of training set samples on model performance was explored through parameter experiments. The experiments were conducted on two public datasets of HDFS and BGL. The results show that the proposed model can improve the pseudo-label accuracy. Compared to existing models of DeepLog, LogAnomaly, LogCluster, PCA and PLELog, the precision and F1 score are improved. Compared to traditional ADOA anomaly detection methods, F1 scores are increased by 8.4% and 8.5% on the two datasets, respectively.

To solve the problem that the existing in-vehicle voice navigation devices were susceptible to interference from the noise both inside and outside vehicle and could not accurately determine the source of sound signals, the voice navigation anti-interference system based on lip state recognition was proposed. Using a camera to perform real-time recognition of the driver lip state, the start and end points of the driver voice signal were accurately determined, and the activation and deactivation of the voice navigation input signal were controlled for enhancing the driver control over the voice navigation and reducing the interference from the noise inside and outside vehicle. To accurately assess the accuracy and robustness of lip state recognition, the multimodal lip state recognition network based on key point-appearance short-term feature fusion was proposed. The experiment of validating the effectiveness of key point short-term features, the ablation experiment of multimodal feature fusion in lip state recognition and the voice navigation anti-interference tests in both simulated laboratory environments and real in-vehicle environments were conducted. The results show that the proposed key point short-term feature operator can enhance the representation ability of lip state changes by more than 14%. The key point-appearance fusion lip state recognition network improves the recognition accuracy by 8.98% through feature complementation. The voice navigation anti-interference system based on this network exhibits high accuracy of 92.6% and good real-time performance with detection speed of 35 F/s. The interference from the noise inside and outside vehicle on the driver voice control authority can be effectively reduced even under the significant head pose changes of more than 70 degrees to the left or right, which demonstrates high robustness.

To optimize the segmental division, storage and hoisting scheme of segmental prefabricated corrugated steel web composite box girder bridge, taking the approach bridge of Nanjing Yangtze River Fifth Bridge-Lixin Road Crossing Bridge as research background, the numerical simulation was used to verify the reasonable partition length of composite beams with corrugated steel webs. The stress and deformation characteristics of standard beams during storage and hoisting period were analyzed, and the reasonable control measures were proposed. The results show that when the segment length is from 240 cm to 400 cm, it has little influence on the flexural capacity of beam, and the reasonable segment length of composite beams with corrugated steel webs depends on hoisting and transportation conditions. The number of storage layers of section beams should not exceed 2 layers. If double layers are used for storage, the temporary rigid bracing should be set between the concrete top and bottom plates of section beams with corrugated steel webs. Four hoisting points should be adopted for section beam hoisting with four hoisting points set near the web, and the temporary rigid bracing should be set between the top and bottom plates during hoisting to avoid the excessive deformation of roof and the increasing difficulty of on-site assembly control.