The investigation on the rumor propagation model is of great significance in developing rumor-dispelling strategy and maintaining social stability. Considering the rumor propagation of college students on college network social platforms, the IG2D2 rumor model with six states is proposed, and then the rumor propagation model based on average field theory is established. Furthermore, taking the factors such as individual differences, conformity effect and trust degree into consideration, a novel non-consistent dynamical rumor propagation model is proposed. The simulation results show that the influence of the degree of the network on the peak value of rumor propagation decreases with the increase of conformity effect, and the retransmission degree of the rumor itself plays an important role in the propagation of rumor.

Research on product quality prediction for intermittent processes with limited data and nonlinear characteristics. In multi-source domain adaptation, the difference in the amount of data in the source and target domains leads to data imbalance, while there is nonlinearity in the process, which further leads to the inability to model accurately. To address this issue, the SMOTE method is first employed to eliminate the adverse effects of data imbalance on the model’s prediction results. Then, a nonlinear MDAJYPLS algorithm based on RBF networks is proposed to capture the nonlinear characteristics and make full use of the information from multiple similar source domains to assist the target domain process modeling. Further, a batch process quality prediction based on non-linear multi-source domain adaptive JYPLS via SMOTE is proposed, which can improve the prediction accuracy of the model. Finally, the effectiveness of the proposed method is verified by simulation experiments of penicillin fermentation process quality prediction.

For the limitation of existing control methods for obtaining intermediate states, a practical control method (stateless control) is proposed for suspension system by combining the linear quadratic regulator with the Luneberger observer. It directly establishes the relationship between current control input with historical control inputs and outputs. Using a quarter-car suspension system model as an example, the practical control performance of the stateless method is investigated. It incorporates sensor noise, modeled according to sensor precision, and accounts for suspension system parameter uncertainties. The results demonstrate that under both excitation and random excitation, the stateless control method effectively balances the three ride comfort performance indices of the suspension system, achieving excellent overall control performance. Crucially, the stateless control method only requires historical sensor outputs and control inputs to realize optimal control effectiveness even with limited sensors, thus significantly simplifying the control system design and offering significant practical value.

To address the difficulties in stochastic distribution control teaching, including high theoretical abstraction, the gap between simulation and real industrial constraints, and limited experimental resources, a pure physical experimental system based on a vertical disc grinding process was developed. The system integrates platform construction, data acquisition, modeling, and control into a unified framework. Centered on a self-developed vertical disc mill, it incorporates precise feeding, disc gap adjustment, and speed regulation, while real-time particle size distribution data are obtained through automatic intermittent sampling and a dry laser particle size analyzer. Radial basis function expansion and inverse integration are used to parameterize the output probability density function. Combined with iterative learning and subspace identification, a linear prediction model describing the dynamic relationship between manipulated variables and weight vectors is established. A constrained controller is further designed to achieve target distribution tracking under input constraints and disturbances. Experimental results show that the system is stable and interactive, and can effectively support experimental teaching in stochastic distribution control and related courses.

The problem of memoryless feedback stabilization for a class of linear systems with unknown input delays is studied. For a class of unstable linear systems whose open-loop poles are located on the imaginary axis, a finite-dimensional memoryless feedback control scheme is proposed when the input delay is unknown. By designing a memoryless truncated predictor feedback (TPF) controller, the implementation problems brought by traditional infinite-dimensional control schemes are avoided. When the unknown input delay is in a time interval, the global asymptotic stability of the closed-loop system is guaranteed. An explicit expression for the variation range of the unknown input delay is given. Finally, the designed memoryless truncated predictive feedback controller is applied to the delayed dual oscillator system and anti-pitching control of high-speed catamaran, and its effectiveness is verified through MATLAB.

A cooperative output-feedback secure control scheme based on observers is proposed for connected and autonomous vehicle systems with intermittent denial-of-service (DoS) attacks on communication. Firstly, the dynamic model of the longitudinal connected and autonomous vehicle system is analyzed, and the feedback is linearized to obtain the linear dynamic equations. Secondly, by using the common Lyapunov function, a secure control scheme is designed to make the cooperative tracking error asymptotically stable. Finally, for maximizing the duration of DoS attacks, appropriate parameters are selected to design an optimization algorithm, in order to ensure the safe operation of the connected and autonomous vehicle system. The experiment simulated a networked vehicle system consisting of 4 followers and 1 leader, and the simulation results verified the effectiveness of the proposed method. The experiment is conducted by simulating a connected and autonomous vehicle system consisting of 4 followers and 1 leader. The simulation results verify the effectiveness of the proposed method.

Accurate trajectory tracking is the key for the intelligent vehicle to achieve autonomous motion control. A novel robust adaptive sliding mode control strategy is proposed to dealing with the issue of system uncertainty affecting the accuracy of trajectory tracking control. Firstly, a two-degree-of-freedom vehicle dynamics model is established based on the principles of vehicle kinematics. Secondly, a sliding surface of proportional integral derivative (PID) type with adaptive properties is designed based on trajectory tracking errors, and the accuracy and robustness of trajectory tracking control are improved by designing adaptive update laws to estimate the sliding mode control gains and the upper bound of system uncertainty in real-time. Thirdly, the control parameters of the controller are optimized by the particle swarm optimization algorithm, which further improves the trajectory tracking control performance. Finally, the proposed control strategy is verified by the simulation under different road conditions and vehicle speeds. The simulation results show that the proposed control strategy can ensure that the intelligent vehicle tracks the target trajectory under the influence of system uncertainty, and its control performance is superior to the fractional order PID control.

In order to address issues of low simulation accuracy, poor universality and poor visualization effect, a universal high-fidelity air-to-air missile simulation system based on Unity3D and CADAC is designed. On the one hand, the designed simulation system uses a high fidelity CADAC software package, which comprehensively considers the missile kinematics and aerodynamics, different guidance laws and target maneuvering modes, fuel consumption and center of gravity changes, and other factors, thus improving the fidelity of the simulation system. On the other hand, the designed system can be applied to different category of missiles such as air-to-air missiles, air-to-ground missiles, etc., which improves the universality and adaptability. Finally, Unity3D is used to conduct visual simulation, and visually display the whole process from launching, searching, guidance and hitting the target. Typical simulation cases and performances analysis is given to verify the effectiveness of the simulation system.

Simultaneous localization and mapping (SLAM) is a key problem in the research and application of mobile robots, which is used to realize autonomous and accurate localization of mobile robots in complex environments. The system composition, key technologies and applications of SLAM are briefly introduced. Focusing on five aspects: feature point method, filtering method, graph optimization method, multi-sensor fusion and dynamic scene, the key technologies, domestic and foreign research status and symbolic application progress of SLAM system are reviewed. Combined with representative systems, the advantages and disadvantages of different methods are compared and analyzed, and the multi-sensor fusion SLAM systems are elaborated in detail, and the SLAM technology in complex scenes is prospected.

To address the challenge of achieving low-power consumption and rapid online measurement of chemical oxygen demand (COD) in small-scale wastewater treatment plants, an online self-organizing neural network (OSNN) prediction method based on radial basis function (RBF) is proposed. This method realizes accurate prediction of COD by dynamically controlling the number of neurons and their update rate. By leveraging the excellent continuous function approximation capability of RBF, combined with the flexibility and adaptability of self-organization, the accuracy and adaptability of the measurement model are improved. The proposed method for controlling the update rate of neuron number can maintain the compactness of the neural network and reduce the extension of training time caused by frequent and substantial changes in neuron number. Experimental results demonstrate that the RBF neural network with self-organizing capability can reliably predict the COD parameter values.

The steam specific enthalpy and dryness of steam turbines in small-scale cogeneration units are important indicators for evaluating turbine safety and economy. However, they are usually not measurable online. An approach for online calculation of these parameters is proposed. First, the superheated state of the steam at the outlet of the steam turbine stage is determined. For the superheated steam, the specific enthalpy of the stage outlet steam is calculated by solving the isentropic enthalpy drop and internal efficiency of the stage. For the saturated steam, the specific enthalpy and dryness of the stage outlet steam are calculated by means of a comprehensive calculation model incorporating the isentropic enthalpy drop, internal efficiency and saturated steam specific enthalpy. Verification results for a 15 MW cogeneration steam turbine demonstrate that the proposed approach is feasible for online application and achieves high precision.

Taking a six-degree-of-freedom (6-DoF) serial manipulator as the research object, the concept of key joints is first introduced. By implementing torque control at these key force-controlled joints while simultaneously incorporating their output angles into the manipulator’s position control, a hybrid force/position control algorithm based on active force-controlled joints is proposed. Leveraging the theories of manipulator dynamics and inverse kinematics, the proposed algorithm achieves force control along a predefined positional trajectory. Finally, for the RM65-B dexterous 6-DoF manipulator, a MATLAB-based simulation platform is established to conduct experiments and analyze the tracking performance of the hybrid force/position control scheme. The favorable simulation results validate the feasibility of the proposed control method.

2025年8月21日至8月24日，东北大学流程工业综合自动化全国重点实验室承办的第七届工业人工智能国际会议在沈阳召开。大会举办了关于学术服务的重要性及其如何开展的圆桌论坛，论坛由东北大学贾同教授主持，邀请到领域内六位知名学者：金耀初教授（西湖大学）、陈俊龙教授（华南理工大学）、宋永端教授（重庆大学）、黄廷文教授（深圳理工大学）、杨双华教授（英国雷丁大学、广州南方学院）、丁正桃教授（英国曼彻斯特大学），围绕学术服务的核心价值、实践路径、职业赋能、平衡策略及跨领域协同等关键问题展开深入研讨，为广大科研工作者，尤其是青年学者提供了兼具理论高度与实践指导的宝贵经验。

In recent years, inspection robots have greatly improved the safety of mine operations. However, the unstructured environment of mines poses difficulties in the high cost and maintenance to track inspection robots, and the explosion-proof requirement of mines limits the hardware design of inspection robots. A mine inspection unmanned aerial vehicle (UAV) system that meets intrinsic safety standards is proposed for the first time, and designs low-power UAVs and charging ground cabins for the explosion-proof requirement. In addition, the UAV inspection scheme is adopted to overcome the difficulties caused by the unstructured environment of mines. Furthermore, a scheduling strategy model for unmanned inspection tasks in underground mines based on automata is established, and synthesized a complete scheduling strategy for unmanned inspection tasks in underground mines based on the supervised control theory. Finally, the effectiveness of the scheduling strategy is verified through scheduling experiments with a UAV and a ground handling cabin.

PID control algorithm is a widely used control strategy with reliable control performance, and it has many advantages such as simple structure, easy tuning and easy implementation. In order to solve the parameter tuning problem of PID controller with robustness constraint, we apply non-dominated sorting genetic algorithms-II (NSGA-II) to tune the PID controller parameters and applies the optimized PID to the control of flight attitude and altitude of quad-rotor aircraft. Firstly, the quad-rotor aircraft model is linearized. For the altitude and attitude channels of full drive control, the linear model of the sub-channel is obtained. Then, in the NSGA-II, PID parameters are selected as decision variables, the rejection ability performance and tracking performance are taken as the two objective functions, and the maximum sensitivity function is taken as the robustness constraint. Finally, by comparing the parameter control performance of the simulations with that of the traditional method, the superiority of multi-objective genetic algorithm tuning PID in tracking and jamming is verified.

For the lot-sizing scheduling problem of uncorrelated parallel machines considering the adjustment time of sequence-dependent machines, a mathematical model of the problem is established with the optimization objectives of minimizing the maximum completion time (Makespan) and the total number of machine switching times, and an improved multi-objective biogeography optimizer (IMOBBO) is designed to solve the problem. In order to meet the needs of lot-sizing scheduling, a partitioned matrix coding method is designed in the first stage, and the initial population is generated by the fusion strategy of random strategy, Logistic mapping and reverse learning mechanism, and wandering mechanism of wolves are introduced into the process of species migration, adaptive catastrophe operator and two-stage neighborhood search strategy are used in the catastrophe process; in order to minimize the adjustment time of each machine, the second stage adopts machine-based matrix sequence coding to optimize the processing sequence of each batch, and finally, the individual evaluation method based on hypervolume is introduced to Pareto non-dominated rank ordering. The effectiveness and superiority of the algorithm proposed are proved through simulation experiments of different scales and examples and comparison with related algorithms.

To solve the problems of low accuracy and poor generalization ability of the prediction model caused by outliers or noise in data, a robust prediction interval method based on stochastic configuration network and Bayesian quantile regression is proposed. Firstly, the stochastic configuration network (SCN) algorithm is employed to determine the number of the nodes in the hidden layer, as well as the input weights and biases. Then, the Bayesian quantile regression is embedded into the SCN to replace the classical least squares regression, the asymmetric Laplace distribution is used as the prior distribution of the SCN noise, and the maximum posterior estimation is used to convert the prior distribution of the SCN noise into the posterior distribution of the output weights. Finally, the expectation maximization algorithm is used to iteratively optimize the SCN noise and hyper-parameters of the hypothesis distribution on the output weights. The experiment is conducted based on the standard datasets and historical data of the municipal solid waste incineration process to test the proposed method, and it is compared with other prediction algorithms based on SCN and quantile regression. The experimental results show that the proposed method has advantages in terms of the accuracy and generalization ability of point predictions, the reliability of prediction intervals, robustness, and computational efficiency.

The current elevator group control scheduling system has insufficient debugging safety, high difficulty in design evaluation, and high debugging costs. Therefore, a simulation model for elevator group control is developed. Based on this model, the elevator group control scheduling scheme is dynamically evaluated according to indicators such as the average waiting time of passengers, long-time waiting rates, and the energy consumption of the elevators. To address the issues of low elevator operation efficiency and poor passenger comfort during elevator rides, a deep Q network (DQN) is developed to realize optimized scheduling for elevator group control. The corresponding state space, reward signals, and agent structure are designed by the primary factors affecting elevator transportation efficiency. Considering the lengthy duration of online training for reinforcement learning agents and the challenge of providing real-time decision support, the operational data from the elevator group control simulation model is utilized to train a feedforward neural network, and an elevator group control environment prediction model is developed to serve as the agent training environment. Simulation experiments are conducted by using the elevator group control simulation model. The results showed that, compared with the prevalent minimum response time strategy, the proposed strategy reduces the average waiting time of passengers, average time that passengers spend in the elevator, long-time waiting rate, the number of elevator starts and stops, and increases the average number of arrivals within 5 minutes.

For the problem of positioning and anti-swing of underactuated bridge crane system, a complementary sliding mode control PD control strategy is proposed. On the basis of the linear sliding mode surface composed of displacement, swing angle and respective first-order differential state errors as variables in the system, the error integral of displacement and swing angle is introduced to form an integrated sliding mode surface, and a supplementary sliding mode surface is designed correspondingly, so that the variable space of the system is limited to the intersection of the sliding mode surfaces, and the distance between the variable and the sliding mode surface is shortened. The speed of the system is improved. At the same time, the introduction of PD compensation control of load swing angle enhances the ability to suppress the swing angle and makes the car run more smoothly. The stability of the system is proved by constructing the Lyapunov function. Simulation and experimental results verify the effectiveness of the control strategy for positioning and anti-swing.

Short-term power load forecasting is very important to the stable operation of power system, in order to further improve the accuracy of load prediction, a combined short-term power load prediction model based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and the improved northern goshawk optimization (INGO) algorithm is proposed to optimize bidirectional long short-term memory neural network (BiLSTM). Firstly, the original load sequence is decomposed by CEEMDAN to obtain more stable load data. Then, through Arnold chaotic reverse learning initialization, adaptive Cauchy-Gaussian mixture mutation strategy and nonlinear convergence factor, the problems in the northern goshawk optimization algorithm were improved, and its optimization ability and convergence speed were significantly improved, so as to optimize the BiLSTM related hyperparameters. Finally, the CEEMDAN-INGO-BiLSTM power load prediction model is obtained by integrating and reconstructing each subsequence. The simulation results show that, compared with the comparison algorithm, the model effectively improves prediction accuracy

To solve the limited driving range and long recharging time during the distribution cost process of electric vehicles, the electric vehicle routing problem considering recharging mode decision is proposed, a mixed integer programming model is constructed with the goal of minimizing the total distribution cost. In response to the characteristics of this problem, an improved adaptive large neighborhood search algorithm driven by recharging and swapping features is designed. Based on the flexibility of charging time and the close correlation between charging stations and customers, neighborhood operators such as affiliation destroy and comparison repair for recharging stations are introduced. The experimental results show that compared with the large neighborhood search algorithm, the proposed algorithm can obtain better solutions when solving large-scale examples. Reasonable selection of the recharging mode decision can effectively shorten the recharging time of electric vehicles and reduce the total distribution cost.

Due to the problems of parameter uncertainty, actuator failure and external disturbance in the quadrotor UAV, in the case of limited input, the use of adaptive PID sliding mode control alone will not fully compensate the total disturbance, resulting in poor control effect. An adaptive sliding mode controller based on disturbance observer is proposed. Firstly, the adaptive sliding mode control is used to compensate the internal disturbance of the system such as parameter uncertainty, and the external disturbance is estimated and compensated by the disturbance observer to reduce the controller output; Then, the RBF neural network is used to optimize the control parameters by self-learning and self-adaptive ability, so that the tracking effect is better; the stability of the closed-loop system is finally proved. The simulation results show that the proposed method can reduce the controller output and have higher control quality when the input is limited.

In order to solve the problem of pose estimation accuracy degradation caused by error accumulation in mobile robot localization, a deep learning-based visual odometry method is proposed. Firstly, a convolutional neural network (CNN) is designed to extract more detailed features of the image sequences by optimizing the size of the convolutional kernel layer by layer. Then, an adaptive memory network records historical poses, while a bi-directional long short-term memory (Bi-LSTM) predicts future poses. By fusing both past and future information, the method reduces error accumulation in pose estimation. Finally, experiments on KITTI and TUM datasets show the method outperforms existing approaches in pose accuracy, absolute and relative trajectory error.

In order to better solve the flexible job-shop scheduling problem, a mathematical model is established with the goal of minimizing the maximum completion time, and the grey wolf optimization algorithm is improved. Firstly, a new formula for individual position update is proposed, which reduces the guiding role of the alpha wolf on the grey wolf population. Secondly, since the linear convergence factor cannot fully exploit the performance of the grey wolf optimization algorithm, a nonlinear convergence factor is introduced to enhance its global exploration ability in the early stage and local exploitation ability in the later stage. Thirdly, in order to address the excessive guidance of optimal individuals in the grey wolf optimization algorithm, two neighborhood exploration strategies are proposed, allowing some individuals to conduct self-exploration and key individuals to undergo machine-based mutation, thereby reducing the completion time. Finally, an elite solution update mechanism combined with the Hamming distance is proposed. The experimental results show that the proposed improvement strategies are all effective, the improved grey wolf optimization algorithm can better solve the flexible job-shop scheduling problem and enhance the production efficiency of the job-shop.

Process industries, as the main sources of energy consumption and carbon emissions, have particularly significant issues related to energy efficiency optimization. To tackle the uncertainties of the time-of-use electricity price, a real-time demand response scheduling methodology for energy systems in the cement production industry is proposed, which combines the long short-term memory (LSTM) neural network with reinforcement learning. Firstly, mathematical models are established for the main energy-consuming equipment and storage bin in each cement production process, and a constrained Markov decision process is constructed to reflect the energy consumption characteristics and production demands. Then, the sequence processing capability of the LSTM neural network is utilized to handle the uncertainties of the time-of-use electricity price, thereby providing robust data support for the development of the scheduling strategy. Finally, a reinforcement learning agent is employed to sense the environment and optimize the scheduling strategy to achieve the goal of energy efficiency optimization. The simulation results verify the feasibility and reliability of the proposed method in the demand response scheduling of the energy system in the cement production industry. A new idea for achieving energy efficiency improvement and sustainable development is provided in the industries.

A model-free control framework based on iterative learning is proposed to realize the synchronous and coordinated control of a double-lift overhead cranes system for the problems of inaccurate modeling, system parameter variation and uncertainty perturbation are common in double-lift overhead cranes system. Firstly, a time-varying sliding mode surface using Sigmoid-like functions is proposed to improve the convergence speed of the system state. Then, an iterative learning law based on the time-varying sliding mode surface is introduced to compensate for the inclusion of unknown system dynamics and external disturbances, etc., to achieve model-free control. At the same time, a dynamic learning rate is designed instead of the fixed-value learning rate to improve the convergence speed of the error of the double-lift overhead cranes system as well as the steady-state performance. Secondly, an improved adaptive convergence law is proposed to reduce unnecessary chattering, improve the robustness of the double-lift overhead cranes system, and achieve finite time convergence. Finally, the stability of the controlled system is demonstrated using Lyapunov stability theory. The simulation experiments verify the effectiveness of the designed synchronous control scheme.

For the large amount of band data in hyperspectral remote sensing images and the difficulty in distinguishing and classifying conventional features, a method for hyperspectral image classification based on a semi-supervised subsidiary classifier-generative adversarial network (SSC-GAN) is proposed. Firstly, the discriminator is developed into a semi-supervised multi-classifier, and train the GAN generator in cascade with the discriminator. Then, to expand the data, the discriminator pulls higher-level features from the generated images, and the auxiliary classification labels are coordinated with the generated samples. Finally, a convolutional neural network replaces the discriminator and generator in the GAN, and the dynamic stacking optimization of the two upgraded networks is carried out. The improved algorithm model achieves the best classification accuracy compared to some classical hyperspectral remote sensing image feature extraction and classification approaches, and has more advantages in classification ability and robustness.

Currently, indoor mobile robots are used in industries such as intelligent storage and security systems, in which high-precision positioning technology plays an important role. To this end, a vision servo-based indoor 3D localization technique is designed in order to achieve high accuracy of localization error while getting rid of the limitations of camera placement and field of view in visual image processing localization techniques. Firstly, the framework and application scenarios of the positioning algorithm are introduced. Then, the 3D positioning algorithm based on geometric relations is designed and its least-squares problem is solved, and the tracking and aiming problem of the visual pan tilt is solved by using image processing and feedback control techniques. Finally, the experimental results of positioning the target robot show that the proposed indoor 3D positioning algorithm has high accuracy and stability with a positioning accuracy within 4 cm in the laboratory scenario.

For the four-motor synchronous drive system with backlash, a high-performance Partial loss of effectiveness (PLOE) fault-tolerance control technique is proposed. Firstly, the nonlinear conditions on the motor side and load side are approximated separately using the fuzzy logic system to compensate during the control design process to reduce the tracking error and ensure the tracking performance. In the designed control method, only one adaptive parameter needs to be estimated, reducing the design difficulty. Secondly, a current observer is used to monitor the motor current in real time and estimate the failure factor to obtain the current system parameter information. Finally, a four-motor synchronization architecture based on cross-coupled synchronization control is designed to synchronize the control signals and ensure synchronization performance among the four motors. To verify the effectiveness of the method, the proposed method is compared with dynamic surface fault-tolerant control through simulation. The results show that the proposed method can effectively reduce the tracking error of the servo system.

For the multi-quadrotor unmanned aerial vehicle (QUAV) system with dynamic uncertainty and external disturbances, a predefined-time formation control algorithm based on the leader-follower method is proposed. Firstly, a predefined-time command filter is introduced to solve the “complexity explosion” problem caused by repeated derivation of virtual signals, and a non-smooth error compensation mechanism is constructed to eliminate the effect of filtering error on the system. Secondly, Lyapunov stability theory is used to prove that the predefined-time formation controller can make the closed-loop system reach a stable state within the predefined time, all the signals in the closed-loop system are bounded within the predefined time, and the formation tracking errors of the multi-QUAV system converge to a neighborhood near the origin within the predefined time. Finally, the proposed algorithm is tested through simulation using a multi-QUAV system consisting of 1 leader and 4 followers. The simulation results demonstrate the effectiveness of the proposed algorithm.

Plate crown is the most important quality index to evaluate the cross-section profile of hot-rolled plate, and good plate crown is the guarantee of normal production, so accurate prediction of bad crown is of crucial significance to ensure production. In the actual production of hot rolling, the number of qualified crown samples is much higher than the undesirable crown, and a large number of nonlinear parameters and strong coupling between parameters make plate crown prediction a very complex imbalanced classification problem. According to the complex data characteristics of hot-rolled plate crown, considering the powerful nonlinear fitting ability of deep learning, combined with cost-sensitive learning to improve the misclassification cost of bad crown, a cost-sensitive deep belief network (CS-DBN) model is proposed. Evaluation indicators such as Macro-F1, Micro-F1, G-Mean and Aacc are used as evaluation indicators of the model. By adjusting the model hyperparameters and optimizer to determine the optimal cost-sensitive deep belief network, and comparing and analyzing it with traditional machine learning algorithms ANN, SVC, KNN, DBN, LR, the results show that CS-DBN is better than traditional machine learning models in all evaluation indicators, and the plate crown prediction results are good.

The existing research results have problems such as large gap between the scheduling model and the actual problem, insufficient research on the essential characteristics of the problem, and insufficient systematic research on the optimization algorithm, and there is no literature on modeling or optimization for the prototype vehicle testing problem. There is no literature on modeling or optimization of the test problems in the experimental stage, which makes it difficult to effectively serve production management. This paper starts from the production reality, investigates the test scheduling problem in depth, establishes mathematical models, systematically discusses its essential characteristics and optimization methods, proposes a hybrid intelligent scheduling method for test vehicles in the development and testing stage of new models, and carries out simulations and industrial applications. The results show that the method can enrich and deepen the existing optimal scheduling theory and method; and can directly serve the automotive industry and promote its production management level and market competitiveness.

To solve the problem that the existing traffic accident risk prediction models lack the extraction of regional spatial correlation and dynamic spatiotemporal features, a traffic accident risk prediction model is constructed based on the spatiotemporal convolutional network with regional similarity. Firstly, a spatial-channel attention multi-graph convolutional network is constructed based on the graph convolutional network, in order to comprehensively capture local geospatial similarity and global semantic attributes. Secondly, spatiotemporal attention is introduced to learn the dynamic representation of the accident features adaptively. Finally, spatial dependencies are captured through multi-head graph attention networks, and temporal correlation of long sequence is modeled by using gated units with bidirectional temporal convolution. The proposed model is tested on two real traffic accident datasets. The experimental results show that the prediction performance of the proposed model for traffic accident risk is superior to that of benchmark models such as long short-term memory neural network.

To address the challenges of excessive computational resource consumption caused by trust evaluation and updating in heterogeneous vehicle platoons under a zero-trust framework, as well as the consistent control of vehicle platoons, an innovative zero-trust hybrid event-triggering control strategy is proposed. Firstly, the update frequency of trust between vehicles is improved, different trust update frequencies are set according to the importance of vehicles in the platoon, and the trust of vehicles is introduced into the spacing control strategy of the vehicle platoon. Secondly, a dynamic hybrid event-triggering condition is designed in combination with the update frequency of trust, and the controller is designed considering the communication delay. Finally, the proposed control strategy is verified by simulation. The simulation results show that the proposed control strategy realizes the consistent control of the vehicle platoon under the zero-trust framework.

For the problems such as large prediction errors and few data features of traditional models in the study of aquaculture water quality, a pH prediction model of aquaculture water quality based on feature con-struction is proposed in this paper. The main part of the model is mixed double convolution layer gated circula-tion unit neural network (MDconv-GRU). Firstly, the original 3 effective features are increased to 6 effective features by correlation calculation and feature construction. Then, the data after the feature construction is input into the MDconv-GRU model for training prediction. The results showed that the prediction accuracy of the model is 92.26%, the root-mean-square error is 0.083 8, the average absolute error is 0.063 5, and mean abso-lute percentage error is 0.830 2. The evaluation criteria are better than other models. This model can accurately predict the pH value of water quality of Stichopus japonicus aquaculture, and lay a foundation for realizing pH warning and increasing the yield of Stichopus Japonicus.

To improve the accuracy of unmanned ship pose estimation under complex sea conditions, a unmanned ship pose estimation algorithm based on non-Gaussian feature recognition and Gaussian sum cubature particle filter (GSCPF) is proposed. Firstly, based on the idea of normality test, the largest vertical difference between the sample distribution function and the standard Gaussian cumulative distribution function, and the sample skewness-kurtosis are combined to analyze the distribution feature of the data from several angles. Then, if the data shows non-Gaussian distribution feature, an unmanned ship pose estimation method based on Gaussian sum filtering and particle cubature particle filtering is adopted, otherwise, the cubature Kalman filter (CKF) is directly adopted to estimate the unmanned ship pose. Finally, the results of two kinds of simulation experiments show that the proposed algorithm can significantly improve the accuracy of parameters such as position, velocity and course of unmanned ship.

For the low recognition accuracy and several parameters of the current human activity recognition model, a lightweight convolutional neural network (CNN) human activity recognition model is proposed. Firstly, the sensor data is preprocessed. Then, the processed data is inputted into the CNN model to identify the specificity of the human body activities. Finally, the squeeze-and-excitation (SE) attention module is embedded in the feature extraction backbone network, and different weights are assigned to each convolution channel to strengthen key features and improve model accuracy. The performance of the model is evaluated on three public datasets of UCI-HAR, WISDM and OPPORTUNITY. The 1F of the model on the UCI-HAR dataset is 97.54%, with 17 198 parameters. 1F on the WISDM dataset is 97.66%, with 16 622 parameters; 1F on the OPPORTUNITY dataset is 82.38%, with 27 545 parameters. Compared with the existing advanced human activity recognition models, the recognition accuracy is higher, the model parameters are fewer, and the model generalization ability is enhanced.