Civil aviation tractor-oriented adaptive model predictive control method and system based on fusion perception

Abstract

The application discloses a civil aviation tractor-oriented adaptive model predictive control method and system based on fusion perception, and belongs to the technical field of automatic driving and vehicle motion control. The application identifies load parameters such as the number, mass and geometric size of a baggage compartment through fusion perception of a vehicle-mounted laser radar and a visual sensor, dynamically constructs an adaptive variable-structure multi-body dynamics model, adaptively adjusts the weight and constraint boundary of a model predictive controller based on the real-time model, solves an optimal control sequence by rolling and executes the optimal control sequence through a drive-by-wire chassis, and simultaneously realizes safety monitoring through closed-loop feedback of a vehicle motion state. The application can adapt to 1-10 baggage compartment dynamic load working conditions, solves problems such as mismatch of a traditional fixed model, low trajectory tracking precision, high risk of folding and collision and the like, realizes high-precision trajectory tracking and all-condition safety control, adapts to a civil aviation apron operation scene, and has high safety, strong adaptability and good engineering landing value.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving and vehicle motion control technology, specifically to an adaptive model predictive control method and system based on fusion perception for civil aviation tractors. Background Technology

[0002] With the rapid development of the civil aviation transportation industry, the demand for automation and intelligence in airport apron ground support operations is becoming increasingly urgent. As a core piece of apron support equipment, civil aviation baggage towing vehicles must tow varying numbers of baggage bins in the narrow, dynamic environment of the apron to complete cargo transfers. The accuracy of their trajectory and the stability of their operation directly affect flight support efficiency and apron operation safety. Currently, most automatic driving control schemes for civil aviation towing vehicles adopt traditional PID control or fixed-parameter model predictive control methods. The core technical problem faced in actual apron operations is that the number of baggage bins towed by the towing vehicle typically varies frequently between 1 and 10, and the mass difference between an empty and fully loaded bin can reach several tons. This causes drastic changes in the inertia, moment of inertia, and motion characteristics of the entire vehicle system. Fixed-parameter control models cannot adapt to the dynamic changes in load, resulting in serious deviations between the model and the actual physical system. This leads to a significant decrease in trajectory tracking accuracy and even control instability during sudden load changes. Furthermore, long arrays of multi-section bins present significant risks of inner wheel difference and sweeping collisions when turning, and are highly susceptible to metal breakage and instability accidents during reversing and emergency braking. Therefore, how to achieve high-precision trajectory tracking and safe and stable control of tractor fleets under operating conditions where the quantity, quality, and geometric dimensions of the load change dynamically in real time is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the purpose of this invention is to provide an adaptive model predictive control method and system based on fusion perception for civil aviation tractors, which solves problems such as mismatch of traditional fixed models, low trajectory tracking accuracy, and high risk of bending and collision, and achieves high-precision trajectory tracking and safe control under all operating conditions.

[0004] To achieve the above objectives, the embodiments of this invention provide the following technical solutions: This application provides an adaptive model predictive control method based on fusion perception for civil aviation tractors, comprising the following steps: S1, using an onboard lidar and vision sensor, and employing a three-level cascaded algorithm model of "6D attitude detection - space occupancy gridded volume - mass inversion" to identify the physical parameters of the attached baggage compartments, including the number of baggage compartments, the geometric dimensions of each compartment, the mass of a single compartment, rotational inertia, and the ground adhesion coefficient; S2, based on the physical parameters, dynamically constructing and adjusting the state-space equations of the tractor-multi-section baggage compartment multibody dynamics system, linearizing and discretizing the nonlinear dynamic equations, and outputting a real-time updated system state matrix and control matrix; S3, based on the real-time updated system state matrix and control matrix, identifying the current load condition and adaptively adjusting the weight matrix and safety constraint boundary of the model predictive controller, constructing an optimization objective function and solving it in a rolling manner, and outputting the optimal control sequence for the next multiple steps; S4, converting the first step control command in the optimal control sequence into a CAN bus command of the vehicle's drive-by-wire chassis through parameter mapping, and executing the control of the tractor's front wheel steering angle and traction torque or braking torque.

[0005] Further, S1 specifically includes: S11, 6D pose detection of the luggage compartment: extracting point cloud features based on PointNet++ and image features based on ResNet-50, and regressing 9D pose parameters and detection confidence after cross-attention fusion; S12, calculating the rasterized volume of spatial occupancy: establishing a voxelized occupancy roxet with a resolution of 0.05m within the 3D space defined by the pose in S11, counting the number of voxels that meet the occupancy determination threshold, and calculating the volume based on the formula... Calculate the cargo volume, apply the occupancy correction factor, and output the occupancy grid confidence score. For cargo volume, The number of voxels required to meet the occupancy threshold. The reference volume for a single voxel is determined by the raster resolution. , For occlusion correction coefficient; S13, Loading volume mass inversion: Obtain the cargo density prior value by querying the density prior knowledge base through the first path, and combine the loading rate correction factor based on the formula The quality of the first path was calculated, where The prior density values ​​of the goods are matched to the density prior knowledge base. The loading rate correction factor is used; the mass of the second path is calculated based on the visual suspension height change and suspension stiffness characteristics; the mass of the first path and the mass of the second path are fused according to preset weights to obtain the single bucket mass, and the moment of inertia is calculated simultaneously; the density confidence score is output based on the density prior matching degree, based on the formula... Calculate the overall confidence score, where To calculate the overall confidence score, To test the confidence level, To occupy the grid confidence level, For density confidence, These are the corresponding weight coefficients, and .

[0006] Furthermore, S2 specifically includes: S21, adaptive adjustment of model structure: automatically constructing a state vector of the corresponding dimension based on the number of identified luggage compartments, the standard expression of which is: ,in, For the overall position of the tractor unit, The heading angle of the tractor unit. For the longitudinal and lateral speeds of the tractor, The yaw rate of the tractor unit. Let be the hinge angle between the i-th luggage compartment and the vehicle in front. Let be the lateral angular velocity of the i-th baggage compartment. Let the heading angle of the i-th baggage compartment be , To control the cycle, corresponding state components are automatically added or deleted when N changes; S22, Construction of multibody dynamics equations: Based on vehicle system dynamics theory, and combined with real-time updated mass, moment of inertia, aggregate dimensions, and ground adhesion coefficient, nonlinear multibody dynamics equations are constructed for the tractor-multi-section luggage compartment to describe the articulation points and torque transmission, tire lateral slip characteristics, and road adhesion characteristics; S23, Linearization and discretization: The nonlinear dynamics equations are linearized for the current operating point, converted into linear time-varying state-space equations, and then processed according to the control cycle. Discretization is performed to obtain a discrete state-space expression, which is as follows: ;in, Let k+1 be the system state vector. Let k be the system state matrix at time k. Let k be the state vector at time k. Let k be the control matrix. To control the input vector, the control input vector includes the front wheel angle of the tractor and the traction / braking torque; S24, online correction of model parameters: based on the actual motion state collected from the vehicle's CAN bus and sensors, the predicted state is compared with the actual state. When the deviation exceeds the preset threshold, online correction of model parameters is triggered to compensate for the model error.

[0007] Further, S3 specifically includes: S31, Operating condition identification and adaptive weight adjustment: Based on load parameters and vehicle status, the current operating condition type is identified, and the state weight matrix and control weight matrix in the objective function are adaptively adjusted and optimized. The expression for the objective function is: ;in, To predict the time domain, To control the time domain, This is the system state vector at time k+i based on the system state prediction at time k. Let be the reference trajectory state vector at time k+i. To control the rate of change of quantity, The state weight matrix is... To control the weight matrix, As a relaxation factor, S32. Dynamic setting of safety constraint boundaries: Setting hard constraints including the front wheel steering angle limit of the tractor, the upper and lower limits of traction torque or braking torque, the maximum safety threshold of each articulation angle, and the upper and lower limits of vehicle speed, as well as soft constraints including the collision safety distance of the whole vehicle envelope, the rate of change of control quantity, and the trajectory tracking deviation; S33. Rolling optimization solution: Based on the state vector at the current time k, the system state matrix and control matrix, and the constraint conditions, the optimization problem is transformed into a constrained quadratic programming problem and the solver is called to solve it, obtaining the optimal control sequence in the prediction time domain. The expression of the optimal control sequence is: ,in The optimal control sequence obtained at time k. The optimal control input at time k+i is obtained based on the system solution at time k. To control the time domain.

[0008] Furthermore, in S31, based on the identified heavy-load mode, multi-bucket long array mode, or reversing mode, the following weight adjustment strategies are executed respectively: In heavy-load mode, the penalty weight of the control quantity change rate is automatically increased to limit drastic changes in acceleration and deceleration and protect the towing hook; In multi-bucket long array mode, the path deviation penalty weight of the last baggage compartment at the end of the track is increased, and the penalty term of the articulation angle change rate is increased to suppress the serpentine swaying of the convoy; In reversing mode, the constraint weight and deviation penalty of the articulation angle are significantly increased to prioritize the anti-bend control and tighten the hard constraint threshold of the articulation angle from ±30° to ±20°.

[0009] Accordingly, this application also provides an adaptive model predictive control system based on fusion perception for civil aviation tractors, comprising: a fusion perception module for real-time acquisition and output of load physical parameters of the attached baggage compartments via onboard LiDAR and visual sensors; an adaptive variable structure model building module connected to the fusion perception module for dynamically building and outputting a real-time updated system state matrix and control matrix based on the load physical parameters; an adaptive model predictive control optimization solving module connected to the adaptive variable structure model building module for building a predictive model and solving for the optimal control sequence based on the system state matrix and control matrix; a control execution module connected to the adaptive model predictive control optimization solving module for converting control commands into CAN bus commands through parameter mapping and sending them to the drive-by-wire chassis; and a state feedback and safety monitoring module connected to the adaptive variable structure model building module, the adaptive model predictive control optimization solving module, and the fusion perception module, for acquiring the real-time motion state of the tractor and each baggage compartment, performing state smoothing using extended Kalman filtering, and feeding back the filtered state data to the front-end modules to trigger closed-loop rolling optimization, while monitoring safety thresholds and initiating emergency intervention in case of anomalies.

[0010] Furthermore, the fusion perception module includes: a 6D attitude detection unit, based on a multimodal feature fusion network of PointNet++ and ResNet-50, which outputs the attitude parameters and detection confidence of each cargo compartment; a space occupancy rasterization unit, connected to the 6D attitude detection unit, which is used to establish a voxelized occupancy raster in the attitude frame space, calculate the cargo volume by counting the number of occupancy voxels, generate an occupancy raster matrix, and output the occupancy raster confidence; a mass inversion estimation unit, connected to the space occupancy rasterization unit, which is used to estimate the mass and moment of inertia of a single compartment by fusion of density prior and visual suspension height, and output the density confidence; and a confidence evaluation unit, connected to the 6D attitude detection unit, the space occupancy rasterization unit, and the mass inversion estimation unit, respectively, which is used to calculate a comprehensive confidence score according to preset weights and output the confidence level.

[0011] Furthermore, the adaptive structural model construction module includes: a model structure adaptive adjustment unit, used to automatically construct or adjust the dimension of the state vector based on the number of luggage compartments; a multibody dynamics equation construction unit, connected to the model structure adaptive adjustment unit, used to construct nonlinear multibody dynamics equations to describe the force and torque transmission at the hinge point, tire lateral slip characteristics, and road surface adhesion characteristics; and a linearization and discretization unit, connected to the multibody dynamics equation construction unit, used to output a discretized state-space expression according to a control period T=0.1s. ;in, Let k+1 be the system state vector. Let k be the system state matrix at time k. Let k be the state vector at time k. Let k be the control matrix. To control the input vector, the control input vector includes the front wheel angle of the tractor and the traction / braking torque; the parameter verification and online correction unit is used to compare the predicted state with the actual state collected by the state feedback and safety monitoring module. When the deviation exceeds the preset threshold, the online correction of the model parameters is triggered. If the correction still does not meet the standard, the effective model of the previous cycle is used and the fusion perception module is triggered to re-collect data.

[0012] Furthermore, the status feedback and safety monitoring module further includes: a multi-source data acquisition and time synchronization unit, used to acquire the motion status of the tractor and each luggage compartment through the CAN bus, attitude sensor, and articulation angle sensor, and perform time synchronization to control the synchronization error to be ≤ a preset threshold; a data rationality verification and filtering fusion unit, connected to the multi-source data acquisition and time synchronization unit, used to remove abnormal data that exceeds the physical reasonable range, and use an extended Kalman filter algorithm to output smooth motion state parameters, including tractor position, vehicle speed, yaw rate, and articulation angle of each compartment; and a safety threshold judgment and process control unit, connected to the data rationality verification and filtering fusion unit, used to monitor the trajectory tracking deviation, articulation angle of each compartment, and collision risk coefficient in real time; when the articulation angle exceeds the hard constraint threshold or the collision risk coefficient is ≥ a preset value, an emergency braking command is immediately triggered and the autonomous driving process is terminated; when a consecutive preset frame of CAN data is lost, data re-acquisition is triggered and a slow braking command is sent.

[0013] Furthermore, it also includes a lightweight alternative implementation architecture: the fusion perception module uses multi-view vision sensors instead of LiDAR to measure the geometric parameters and loading capacity of the luggage compartment through binocular stereo vision; the adaptive variable structure model construction module uses a simplified bicycle model and articulated angle kinematic model instead of a complex multibody dynamics model; the adaptive model predictive control optimization solution module uses an explicit model predictive control scheme, pre-calculates control laws offline, and calls tables online to adapt to low-computing-power edge computing platforms.

[0014] The beneficial effects of this invention are as follows: by sensing load parameters in real time, dynamically adjusting the prediction model, and adaptively optimizing the control strategy, it achieves sub-meter level high-precision trajectory tracking under dynamic loads, while completely eliminating the risks of iron bending instability and sweeping collisions. It balances control accuracy, driving stability, and operational safety, with fast control response speed and strong adaptability to all working conditions, perfectly matching the narrow space and high safety requirements of civil aviation aprons. Attached Figure Description

[0015] Figure 1This is a flowchart illustrating an adaptive model predictive control method based on fusion perception for civil aviation tractors, provided as an embodiment of this application.

[0016] Figure 2 This is a schematic diagram of the structure of an adaptive model predictive control system based on fusion perception for civil aviation tractors, provided in an embodiment of this application. Detailed Implementation

[0017] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0018] In this invention, the terms "system" and "network" are used interchangeably. "Multiple" refers to two or more; therefore, in this invention, "multiple" can also be understood as "at least two." "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, it should be understood that in the description of this invention, terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0019] Example 1:

[0020] like Figure 1 As shown in the embodiment of this application, an adaptive model predictive control method based on fusion perception for civil aviation tractors is provided, including the following steps: S1, using an onboard lidar and vision sensor, a three-level cascaded algorithm model of "6D attitude detection - space occupancy gridded volume - mass inversion" is adopted to identify the physical parameters of the attached luggage compartments. The physical parameters include the number of luggage compartments, the geometric dimensions of each compartment, the mass of a single compartment, rotational inertia, and the ground adhesion coefficient; S2, based on the physical parameters, the state space equation of the tractor-multi-section luggage compartment multibody dynamics system is dynamically constructed and adjusted, the nonlinear dynamic equation is linearized and discretized, and the real-time updated system state matrix and control matrix are output; S3, based on the real-time updated system state matrix and control matrix, the current load condition is identified and the weight matrix and safety constraint boundary of the model predictive controller are adaptively adjusted, an optimization objective function is constructed and solved in a rolling manner, and the optimal control sequence for the next multiple steps is output; S4, the first step control command in the optimal control sequence is converted into a CAN bus command of the vehicle's drive-by-wire chassis through parameter mapping, and the control of the tractor's front wheel steering angle and traction torque or braking torque is executed.

[0021] In another possible embodiment, firstly, multi-source data of the tractor and its attached luggage compartments are simultaneously collected using an onboard LiDAR and vision sensors. A three-level cascaded algorithm model of "6D attitude detection - spatial occupancy rasterized volume - mass inversion" is then employed to accurately identify the number of luggage compartments on the tractor, the geometric dimensions of each compartment, the mass of a single compartment, the moment of inertia, and the ground adhesion coefficient. Standardized load physical parameters are output to provide accurate input for subsequent modeling. Next, based on the load physical parameters identified by S1, the state-space equations of the multibody dynamics system composed of the tractor and multiple luggage compartments are dynamically constructed. The nonlinear dynamic equations are linearized and discretized, and the system state is updated and output in real time. The matrix and control matrix ensure that the prediction model always maintains a high degree of matching with the actual physical system. Then, the real-time updated system state matrix and control matrix output by S2 are used to build an accurate prediction model, identify the current load condition of the vehicle, adaptively adjust the weight matrix and safety constraint boundary of the model predictive controller, construct the optimization objective function and perform common hole optimization to solve it, and output the optimal control sequence for the next multiple steps. Finally, the first step control command in the optimal control sequence output by S3 is extracted, and the abstract control command is converted into a CAN bus command that can be recognized by the tractor's drive-by-wire chassis through parameter mapping, so as to accurately control the front wheel angle and traction torque or braking torque of the tractor and complete the precise execution of vehicle actions.

[0022] By sensing load parameters in real time, dynamically adjusting the prediction model, and adaptively optimizing the control strategy, sub-meter-level high-precision trajectory tracking is achieved under dynamic loads. At the same time, the risks of bending instability and sweeping collisions are completely eliminated. It takes into account control accuracy, driving stability, and operational safety. The control response speed is fast and the adaptability to all working conditions is strong, which is fully matched to the narrow space and high safety requirements of civil aviation aprons.

[0023] Traditional load parameter estimation methods rely solely on a single visual sensor or static weighbridge weighing, lacking a multimodal fusion detection mechanism. In complex apron environments such as rain, snow, heavy fog, and low light, the accuracy of parameter recognition can plummet by more than 50%. Furthermore, without a three-level confidence assessment mechanism, erroneous load parameters can be directly injected into the dynamic model, causing control failure. Mass estimation uses only a single algorithm and cannot be verified by both volume and suspension deformation. The low accuracy and poor reliability of load parameter recognition make it impossible to provide stable and accurate input data for adaptive modeling.

[0024] In this embodiment, S1 specifically includes: S11, 6D pose detection of the luggage compartment: extracting point cloud features based on PointNet++ and image features based on ResNet-50, and regressing 9D pose parameters and detection confidence after cross-attention fusion; S12, calculating the rasterized volume of spatial occupancy: establishing a voxelized occupancy roxet with a resolution of 0.05m within the 3D space defined by the pose in S11, counting the number of voxels that meet the occupancy determination threshold, and calculating the volume based on the formula... Calculate the cargo volume, apply the occupancy correction factor, and output the occupancy grid confidence score. For cargo volume, The number of voxels required to meet the occupancy threshold. The reference volume for a single voxel is determined by the raster resolution. , For occlusion correction coefficient; S13, Loading volume mass inversion: Obtain the cargo density prior value by querying the density prior knowledge base through the first path, and combine the loading rate correction factor based on the formula The quality of the first path was calculated, where The prior density values ​​of the goods are matched to the density prior knowledge base. The loading rate correction factor is used; the mass of the second path is calculated based on the visual suspension height change and suspension stiffness characteristics; the mass of the first path and the mass of the second path are fused according to preset weights to obtain the single bucket mass, and the moment of inertia is calculated simultaneously; the density confidence score is output based on the density prior matching degree, based on the formula... Calculate the overall confidence score, where To calculate the overall confidence score, To test the confidence level, To occupy the grid confidence level, For density confidence, These are the corresponding weight coefficients, and .

[0025] In another possible embodiment, S11, 6D pose detection of the luggage compartment, is performed first. 6D pose detection refers to detecting the pose information of the luggage compartment in six dimensions: three-dimensional spatial position and three-dimensional pose. PointNet++ algorithm is used to extract LiDAR point cloud features, and ResNet-50 algorithm is used to extract visual sensor image features. The two types of features are then weighted and fused using a cross-attention fusion method, and the nine-dimensional pose parameters of the luggage compartment and the detection confidence are obtained through regression. Next, S12, spatial occupancy raster volume calculation, is performed. Voxel occupancy raster refers to discretizing the three-dimensional space into small cubic grids with a side length of 0.05m. Spatial modeling is achieved by determining whether the grid is occupied. A voxel occupancy raster with a resolution of 0.05m is established within the three-dimensional space defined by the pose in S11. The number of voxels that meet the occupancy threshold is counted, and the result is calculated according to the formula... The cargo volume is calculated, and an occupancy correction coefficient is applied to compensate for volume calculation errors caused by viewpoint occupancy. Simultaneously, the occupancy grid confidence score is output. Finally, S13 loading mass inversion is performed. The first path queries the density prior knowledge base to obtain the cargo density prior value, and combines it with the loading rate correction factor to estimate the cargo mass. The second path estimates the mass based on visual recognition of suspension height changes. The two mass results are fused with a weight of 0.7:0.3 to obtain the single-bucket mass, and the rotational inertia of the luggage compartment is simultaneously calculated. Based on the density prior matching degree, the density confidence score is output, and then the formula is applied... Calculate the overall confidence score to accurately identify and verify the load parameters.

[0026] By employing a fusion of LiDAR and visual multi-sensor systems with a three-level cascaded algorithm, the system's environmental robustness is improved, enabling it to adapt to various complex weather and lighting conditions on the apron. This reduces mass estimation and volume estimation errors. Furthermore, through a three-level confidence weighted scoring system based on detection, grid, and density, the reliability of parameters is accurately determined, preventing control failures caused by erroneous parameter injection into the model. The dual-path mass fusion estimation method further enhances the accuracy of mass identification, providing reliable load parameter support for subsequent adaptive modeling and control.

[0027] Traditional dynamic models are static models with fixed dimensions and parameters, which cannot automatically adjust the model structure according to changes in the number of cargo boxes. They do not perform linearization adaptation on the nonlinear dynamic characteristics of the tractor and multi-section cargo box, resulting in increased model prediction deviation. At the same time, they lack an online correction mechanism for model parameters. When the deviation between the predicted state and the actual state exceeds the standard, the error cannot be automatically compensated. The accumulation of deviation will eventually lead to the instability of the control system.

[0028] In this embodiment of the application, S2 specifically includes: S21, adaptive adjustment of model structure: automatically constructing a state vector of the corresponding dimension based on the number of recognized luggage compartments, the standard expression of which is: ,in, For the overall position of the tractor unit, The heading angle of the tractor unit. For the longitudinal and lateral speeds of the tractor, The yaw rate of the tractor unit. Let be the hinge angle between the i-th luggage compartment and the vehicle in front. Let be the lateral angular velocity of the i-th baggage compartment. Let the heading angle of the i-th baggage compartment be , To control the cycle, corresponding state components are automatically added or deleted when N changes; S22, Construction of multibody dynamics equations: Based on vehicle system dynamics theory, and combined with real-time updated mass, moment of inertia, aggregate dimensions, and ground adhesion coefficient, nonlinear multibody dynamics equations are constructed for the tractor-multi-section luggage compartment to describe the articulation points and torque transmission, tire lateral slip characteristics, and road adhesion characteristics; S23, Linearization and discretization: The nonlinear dynamics equations are linearized for the current operating point, converted into linear time-varying state-space equations, and then processed according to the control cycle. Discretization is performed to obtain a discrete state-space expression, which is as follows: ;in, Let k+1 be the system state vector. Let k be the system state matrix at time k. Let k be the state vector at time k. Let k be the control matrix. To control the input vector, the control input vector includes the front wheel angle of the tractor and the traction / braking torque; S24, online correction of model parameters: based on the actual motion state collected from the vehicle's CAN bus and sensors, the predicted state is compared with the actual state. When the deviation exceeds the preset threshold, online correction of model parameters is triggered to compensate for the model error.

[0029] In another possible embodiment, S21 model structure adaptive adjustment is first performed, automatically constructing a state vector of the corresponding dimension based on the identified number of luggage compartments. The state vector refers to the set of core parameters describing the vehicle's motion state, and its standard expression is: ,in, For the overall position of the tractor unit, The heading angle of the tractor unit. For the longitudinal and lateral speeds of the tractor, The yaw rate of the tractor unit. Let be the hinge angle between the i-th luggage compartment and the vehicle in front. Let be the lateral angular velocity of the i-th baggage compartment. Let the heading angle of the i-th baggage compartment be , To control the cycle, when the number of luggage compartments changes, the corresponding state components are automatically added or deleted to adaptively adjust the model dimensions. Next, the S22 multibody dynamics equations are constructed. Based on the fundamental theory of vehicle system dynamics, and combined with real-time updated parameters such as mass, moment of inertia, aggregate dimensions, and ground adhesion coefficient, a nonlinear multibody dynamics equation is constructed for the tractor and multiple luggage compartments. This equation accurately describes the force and torque transmission characteristics at the articulation points, tire lateral slip characteristics, and road adhesion characteristics. Subsequently, the S23 linearization and discretization are performed. For the vehicle's current operating point, the nonlinear dynamics equations are linearized into linear time-varying state-space equations. The continuous equations are then discretized according to a control cycle of 0.1s to obtain the discrete state-space expression. ,in, Let k+1 be the system state vector. Let k be the system state matrix at time k. Let k be the state vector at time k. Let k be the control matrix. To control the input vector, the control input vector includes the front wheel angle of the tractor and the traction / braking torque; finally, the S24 model parameter online correction is executed. The actual motion state of the vehicle is collected through the vehicle CAN bus and on-board sensors, and the model prediction state is compared with the actual running state. When the deviation between the two exceeds 5%, the model parameter online correction is triggered to automatically compensate for the model error and ensure the accuracy of the model prediction.

[0030] By linearizing and discretizing the nonlinear equations, the model is adapted to the real-time solution requirements of model predictive control. The online correction mechanism of model parameters can automatically compensate for errors when the prediction deviation exceeds 5%, improving the matching degree between the prediction model and the actual physical system by more than 90%, maintaining high-precision prediction capability, and providing a model foundation that best fits the actual working conditions for subsequent control solutions.

[0031] If the weight matrix and constraint boundaries of the model predictive controller are fixed, they cannot be adaptively adjusted according to different working conditions such as heavy load, long array, and reversing. The optimization objective function does not have soft constraint relaxation terms, which can easily lead to problems where the optimization solution is not feasible. At the same time, the lack of a rolling optimization mechanism means that the control commands cannot adapt to the dynamic changes in working conditions in real time, and the trajectory tracking accuracy and safety control capabilities cannot be balanced.

[0032] In this embodiment, S3 specifically includes: S31, Operating condition identification and adaptive weight adjustment: The current operating condition type is identified based on load parameters and vehicle status, and the state weight matrix and control weight matrix in the objective function are adaptively adjusted and optimized. The expression of the objective function is: ;in, To predict the time domain, To control the time domain, This is the system state vector at time k+i based on the system state prediction at time k. Let be the reference trajectory state vector at time k+i. To control the rate of change of quantity, The state weight matrix is... To control the weight matrix, As a relaxation factor, S32. Dynamic setting of safety constraint boundaries: Setting hard constraints including the front wheel steering angle limit of the tractor, the upper and lower limits of traction torque or braking torque, the maximum safety threshold of each articulation angle, and the upper and lower limits of vehicle speed, as well as soft constraints including the collision safety distance of the whole vehicle envelope, the rate of change of control quantity, and the trajectory tracking deviation; S33. Rolling optimization solution: Based on the state vector at the current time k, the system state matrix and control matrix, and the constraint conditions, the optimization problem is transformed into a constrained quadratic programming problem and the solver is called to solve it, obtaining the optimal control sequence in the prediction time domain. The expression of the optimal control sequence is: ,in The optimal control sequence obtained at time k. The optimal control input at time k+i is obtained based on the system solution at time k. To control the time domain.

[0033] In another possible embodiment, S31 condition identification and weight adaptive adjustment are first performed. Based on the load parameters and the real-time vehicle motion status, the current working condition type is identified, and the objective function is adaptively adjusted and optimized. The state weight matrix and control weight matrix in the diagram. To predict the time domain, To control the time domain, This is the system state vector at time k+i based on the system state prediction at time k. Let be the reference trajectory state vector at time k+i. To control the rate of change of quantity, The state weight matrix is... To control the weight matrix, As a relaxation factor, The penalty coefficient for the relaxation factor is defined. Next, S32, dynamic setting of safety constraint boundaries, is executed, setting insurmountable hard constraints, including the front wheel steering angle limit of the tractor, upper and lower limits of traction torque or braking torque, maximum safe thresholds for each articulation angle, and upper and lower limits of vehicle speed. Simultaneously, soft constraints for optimization penalties are set, including the vehicle envelope collision safety distance, rate of change of control variables, and trajectory tracking deviation, balancing driving safety and optimization feasibility. Finally, S33, rolling optimization solution, is executed. Based on the current state vector at time k, the system state matrix, the control matrix, and the hard and soft constraints, the optimization problem is transformed into a constrained quadratic programming problem. A specialized solver is then called to complete the solution, obtaining the optimal control sequence in the prediction time domain. ,in The optimal control sequence obtained at time k. The optimal control input at time k+i is obtained based on the system solution at time k. To control the time domain.

[0034] By adaptively adjusting the control weights according to real-time operating conditions, the accuracy of trajectory tracking and driving stability are precisely balanced. Hard constraints are used to ensure driving safety and soft constraints to ensure the feasibility of optimization solutions, thus completely solving the problem of infeasible solutions. Through rolling optimization solutions, the control commands are adapted to changes in operating conditions in real time. The trajectory tracking error in the whole scene is controlled within ±0.2m. At the same time, safety issues such as metal breakage and collisions are eliminated from the algorithm level.

[0035] If the model predictive control does not have a dedicated weight strategy for the three core operating conditions of heavy load, multi-bucket long array, and reversing, the drastic changes in control quantity under heavy load conditions may easily break the towing hook; under multi-bucket long array conditions, the convoy may easily sway in a serpentine manner; and under reversing conditions, the excessive articulation angle may easily cause a metal breakage accident. It is impossible to perform targeted safety optimization for special operating conditions.

[0036] In this embodiment of the application, in step S31, the following weight adjustment strategies are executed according to the identified heavy-load mode, multi-bucket long array mode, or reversing mode: In the heavy-load mode, the penalty weight of the control quantity change rate is automatically increased to limit drastic changes in acceleration and deceleration and protect the towing hook; in the multi-bucket long array mode, the path deviation penalty weight of the last baggage compartment at the end of the track is increased, and the penalty term of the articulation angle change rate is increased to suppress the serpentine swaying of the convoy; in the reversing mode, the constraint weight and deviation penalty of the articulation angle are significantly increased to prioritize the anti-bend control and tighten the hard constraint threshold of the articulation angle from ±30° to ±20°.

[0037] In another possible embodiment, when a heavy-load mode is identified, the penalty weight of the rate of change of the control quantity is automatically increased to strictly limit drastic changes in acceleration and deceleration, and to prevent damage to the towing hook due to overload. When a multi-bucket long array mode is identified, the penalty weight of the path deviation of the last baggage bucket at the end of the track is increased, and the penalty term of the rate of change of the articulation angle is also increased to effectively suppress the serpentine swaying problem of a convoy composed of multi-bucket trucks. When a reversing mode is identified, the constraint weight and deviation penalty of the articulation angle are significantly increased to prioritize the anti-brick-breaking control effect, while the hard constraint threshold of the articulation angle is tightened from ±30° to ±20° to prevent the brick-breaking caused by the articulation angle exceeding the limit from the constraint level.

[0038] Example 2:

[0039] Reference Figure 2 This application also provides an adaptive model predictive control system based on fusion perception for civil aviation tractors, comprising: a fusion perception module for real-time acquisition and output of load physical parameters of the attached luggage compartments via onboard LiDAR and visual sensors; an adaptive variable structure model construction module connected to the fusion perception module for dynamically constructing and outputting a real-time updated system state matrix and control matrix based on the load physical parameters; an adaptive model predictive control optimization solution module connected to the adaptive variable structure model construction module for constructing a predictive model and solving for the optimal control sequence based on the system state matrix and control matrix; a control execution module connected to the adaptive model predictive control optimization solution module for converting control commands into CAN bus commands through parameter mapping and sending them to the drive-by-wire chassis; and a state feedback and safety monitoring module connected to the adaptive variable structure model construction module, the adaptive model predictive control optimization solution module, and the fusion perception module, for acquiring the real-time motion state of the tractor and each luggage compartment, performing state smoothing using extended Kalman filtering, and feeding back the filtered state data to the front-end modules to trigger closed-loop rolling optimization, while monitoring safety thresholds and initiating emergency intervention in case of anomalies.

[0040] In another possible embodiment, the fusion perception module first collects data in real time through onboard LiDAR and visual sensors, and outputs standardized load physical parameters of the luggage compartment. The adaptive variable structure model building module is connected to the fusion perception module, receives the load physical parameters, dynamically builds the model, and outputs the real-time updated system state matrix and control matrix. The adaptive model predictive control optimization solution module is connected to the adaptive variable structure model building module, and solves the optimal control sequence based on the real-time model and operating parameters. The control execution module is connected to the adaptive model predictive control optimization solution module, converts the optimal control command into CAN bus commands through parameter mapping, and sends them to the tractor's drive-by-wire chassis to complete the action execution. The state feedback and safety monitoring module is connected to the adaptive variable structure model building module, the adaptive model predictive control optimization solution module, and the fusion perception module, respectively. It collects the real-time motion state of the tractor and each luggage compartment through onboard sensors, uses the extended Kalman filter algorithm to smooth the state data, and feeds the filtered accurate state data back to the front-end modules to initiate closed-loop rolling optimization. At the same time, it monitors the safety threshold in real time and immediately triggers emergency intervention operations in abnormal operating conditions.

[0041] In this embodiment, the fusion perception module includes: a 6D attitude detection unit, based on a multimodal feature fusion network of PointNet++ and ResNet-50, which outputs the attitude parameters and detection confidence of each luggage compartment; a space occupancy rasterization unit, connected to the 6D attitude detection unit, which establishes a voxelized occupancy raster in the attitude frame space, calculates the cargo volume by counting the number of occupancy voxels, generates an occupancy raster matrix, and outputs the occupancy raster confidence; a mass inversion estimation unit, connected to the space occupancy rasterization unit, which calculates the mass and moment of inertia of a single compartment by fusion of density prior and visual suspension height, and outputs the density confidence; and a confidence evaluation unit, connected to the 6D attitude detection unit, the space occupancy rasterization unit, and the mass inversion estimation unit, respectively, which calculates a comprehensive confidence score according to preset weights and outputs the confidence level.

[0042] In another possible embodiment, the 6D attitude detection unit is based on a multimodal feature fusion network composed of PointNet++ and ResNet-50, outputting the attitude parameters and detection confidence of each bag compartment; the space occupancy rasterization unit is connected to the 6D attitude detection unit, establishing a voxelized occupancy raster within the three-dimensional space defined by the attitude, calculating the cargo volume by counting the number of occupancy voxels, generating an occupancy raster matrix and outputting the occupancy raster confidence; the mass inversion estimation unit is connected to the space occupancy rasterization unit, estimating the mass and moment of inertia of a single bag by fusing density prior queries and visual suspension height changes, and outputting the density confidence; the confidence evaluation unit is connected to the 6D attitude detection unit, the space occupancy rasterization unit, and the mass inversion estimation unit respectively, calculating the comprehensive confidence score according to weights of 0.4, 0.35, and 0.25, and outputting the corresponding confidence level, completing the full-process quality verification of the load parameters.

[0043] In this embodiment, the adaptive structural model construction module includes: a model structure adaptive adjustment unit, used to automatically construct or adjust the dimension of the state vector based on the number of luggage compartments; a multibody dynamics equation construction unit, connected to the model structure adaptive adjustment unit, used to construct nonlinear multibody dynamics equations to describe the force and torque transmission at the hinge point, tire lateral slip characteristics, and road surface adhesion characteristics; and a linearization and discretization unit, connected to the multibody dynamics equation construction unit, used to output a discretized state-space expression according to a control period T=0.1s. ;in, Let k+1 be the system state vector. Let k be the system state matrix at time k. Let k be the state vector at time k. Let k be the control matrix. To control the input vector, the control input vector includes the front wheel angle of the tractor and the traction / braking torque; the parameter verification and online correction unit is used to compare the predicted state with the actual state collected by the state feedback and safety monitoring module. When the deviation exceeds the preset threshold, the online correction of the model parameters is triggered. If the correction still does not meet the standard, the effective model of the previous cycle is used and the fusion perception module is triggered to re-collect data.

[0044] In another possible embodiment, the model structure adaptive adjustment unit automatically constructs or adjusts the dimension of the state vector based on the identified number of luggage compartments to adapt to the number of compartments in different formations; the multibody dynamics equation construction unit is connected to the model structure adaptive adjustment unit, constructing nonlinear multibody dynamics equations based on dynamics theory to accurately describe the force and torque transmission at the articulation points, tire lateral slip characteristics, and road surface adhesion characteristics; the linearization and discretization unit is connected to the multibody dynamics equation construction unit, outputting a discretized state-space expression according to a control cycle of 0.1s. The parameter verification and online correction unit compares the model's predicted state with the actual motion state collected by the state feedback and safety monitoring modules in real time. When the deviation exceeds 5%, it triggers online correction of the model parameters. If the correction still fails to meet the standard, it automatically adopts the effective model from the previous cycle and triggers the fusion sensing module to re-collect load parameters.

[0045] In this embodiment, the state feedback and safety monitoring module further includes: a multi-source data acquisition and time synchronization unit, used to acquire the motion state of the tractor and each luggage compartment through the CAN bus, attitude sensor, and articulation angle sensor, and perform time synchronization to control the synchronization error to be ≤ a preset threshold; a data rationality verification and filtering fusion unit, connected to the multi-source data acquisition and time synchronization unit, used to remove abnormal data that exceeds the physical reasonable range, and use an extended Kalman filter algorithm to output smooth motion state parameters, including tractor position, vehicle speed, yaw rate, and articulation angle of each compartment; and a safety threshold judgment and process control unit, connected to the data rationality verification and filtering fusion unit, used to monitor the trajectory tracking deviation, articulation angle of each compartment, and collision risk coefficient in real time; when the articulation angle exceeds the hard constraint threshold or the collision risk coefficient is ≥ a preset value, an emergency braking command is immediately triggered and the autonomous driving process is terminated; when a consecutive preset frame of CAN data is lost, data re-acquisition is triggered and a slow braking command is sent.

[0046] In another possible embodiment, the multi-source data acquisition and time synchronization unit collects motion state data of the tractor and each luggage compartment via CAN bus, attitude sensor, and articulation angle sensor, and performs high-precision time synchronization on the multi-source data, controlling the synchronization error within 5ms. The data rationality verification and filtering fusion unit is connected to the multi-source data acquisition and time synchronization unit. It first removes abnormal dirty data that exceeds the physical reasonable range, and then uses the extended Kalman filter algorithm to smooth the effective data, outputting accurate and smooth motion state parameters, including tractor position, vehicle speed, yaw rate, and articulation angle of each compartment. The safety threshold judgment and process control unit is connected to the data rationality verification and filtering fusion unit, and monitors the trajectory tracking deviation, articulation angle of each compartment, and collision risk coefficient in real time. When the articulation angle exceeds the default hard constraint threshold of ±30° (±20° in reversing condition) or the collision risk coefficient is ≥0.9, an emergency braking command is immediately triggered and the autonomous driving process is terminated. When 5 consecutive frames of CAN bus data are lost, a data re-acquisition operation is automatically triggered and a slow braking command is sent to the vehicle.

[0047] In this embodiment, a lightweight alternative implementation architecture is also included: the fusion perception module uses a multi-view vision sensor instead of LiDAR to measure the geometric parameters and loading capacity of the luggage compartment through binocular stereo vision; the adaptive variable structure model construction module uses a simplified bicycle model and articulated angle kinematic model instead of a complex multibody dynamics model; the adaptive model predictive control optimization solution module uses an explicit model predictive control scheme, pre-calculates control laws offline, and calls tables online to adapt to low-computing-power edge computing platforms.

[0048] In another possible embodiment, the fusion perception module uses a multi-view vision sensor instead of LiDAR, and uses binocular stereo vision technology to measure the geometric parameters and load capacity of the luggage compartment, reducing hardware modification costs; the adaptive variable structure model building module uses a simplified bicycle model and articulated angle kinematic model to replace the complex multibody dynamics model, significantly reducing the computational load of the model; the adaptive model predictive control optimization solution module uses an explicit model predictive control scheme, pre-calculates the control laws for all operating conditions offline, and directly calls the control instructions by looking up tables during online operation, without the need for real-time solving, effectively adapting to the hardware conditions of low-computing-power edge computing platforms.

[0049] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0050] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0051] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. An adaptive model predictive control method based on fused perception for civil aviation tractors, characterized in that, Includes the following steps: S1. Using vehicle-mounted LiDAR and vision sensors, a three-level cascaded algorithm model of "6D attitude detection - space occupation gridded volume - mass inversion" is adopted to identify the physical parameters of the attached luggage compartment. The physical parameters include the number of luggage, the geometric dimensions of each luggage compartment, the mass of a single compartment, rotational inertia and ground adhesion coefficient. S2. Based on the physical parameters, dynamically construct and adjust the state space equation of the tractor-multi-section dump truck multibody dynamics system, linearize and discretize the nonlinear dynamic equation, and output the real-time updated system state matrix and control matrix. S3. Based on the real-time updated system state matrix and control matrix, identify the current load condition and adaptively adjust the weight matrix and safety constraint boundary of the model predictive controller, construct the optimization objective function and solve it in a rolling manner, and output the optimal control sequence for the next multiple steps. S4. The first step control command in the optimal control sequence is converted into a CAN bus command of the vehicle's drive-by-wire chassis through parameter mapping, and the control of the front wheel angle and traction torque or braking torque of the tractor is executed.

2. The adaptive model predictive control method based on fusion perception for civil aviation tractors according to claim 1, characterized in that, S1 specifically includes: S11, Luggage compartment 6D pose detection: Point cloud features are extracted based on PointNet++, image features are extracted based on ResNet-50, and 9-dimensional pose parameters and detection confidence are regressed after cross-attention fusion; S12. Calculation of Rasterized Occupancy Volume: Within the 3D space defined by the pose in S11, a voxelized occupancy grid with a resolution of 0.05m is established. The number of voxels that meet the occupancy threshold is counted, and the calculation is based on the formula... Calculate the cargo volume, apply the occupancy correction factor, and output the occupancy grid confidence score. For cargo volume, The number of voxels required to meet the occupancy threshold. The reference volume for a single voxel is determined by the raster resolution. , This is the occlusion correction factor; S13. Loading Volume Quality Inversion: Obtain the prior value of cargo density by querying the density prior knowledge base through the first path, and combine the loading rate correction factor based on the formula. The quality of the first path was calculated, where The prior density values ​​of the goods are matched to the density prior knowledge base. The loading rate correction factor is used; the mass of the second path is calculated based on the visual suspension height change and suspension stiffness characteristics; the mass of the first path and the mass of the second path are fused according to preset weights to obtain the single bucket mass, and the moment of inertia is calculated simultaneously; the density confidence score is output based on the density prior matching degree, based on the formula... Calculate the overall confidence score, where To calculate the overall confidence score, To test the confidence level, To occupy the grid confidence level, For density confidence, These are the corresponding weight coefficients, and .

3. The adaptive model predictive control method based on fusion perception for civil aviation tractors according to claim 1, characterized in that, S2 specifically includes: S21. Adaptive Adjustment of Model Structure: Based on the number of recognized luggage compartments, a state vector of the corresponding dimension is automatically constructed, the standard expression of which is: ,in, For the overall position of the tractor unit, The heading angle of the tractor unit. For the longitudinal and lateral speeds of the tractor, The yaw rate of the tractor unit. Let be the hinge angle between the i-th luggage compartment and the vehicle in front. Let be the lateral angular velocity of the i-th baggage compartment. Let the heading angle of the i-th baggage compartment be , To control the cycle, the corresponding state components are automatically added or deleted when N changes; S22. Construction of multibody dynamics equations: Based on vehicle system dynamics theory, and combined with real-time updated mass, moment of inertia, aggregate size and ground adhesion coefficient, nonlinear multibody dynamics equations of tractor-multi-section luggage compartment are constructed to describe articulation points and torque transmission, tire lateral slip characteristics and road adhesion characteristics. S23. Linearization and Discretization: For the current operating point, the nonlinear dynamic equations are linearized, transforming them into linear time-varying state-space equations, and then processed according to the control period. Discretization is performed to obtain a discrete state-space expression, which is as follows: ;in, Let k+1 be the system state vector. Let k be the system state matrix at time k. Let k be the state vector at time k. Let k be the control matrix. The control input vector includes the front wheel steering angle of the tractor and the traction / braking torque. S24. Online correction of model parameters: Based on the actual motion state collected from the vehicle's CAN bus and sensors, the predicted state is compared with the actual state. When the deviation exceeds a preset threshold, online correction of model parameters is triggered to compensate for model errors.

4. The adaptive model predictive control method based on fusion perception for civil aviation tractors according to claim 1, characterized in that, S3 specifically includes: S31. Operating Condition Identification and Adaptive Weight Adjustment: Based on load parameters and vehicle status, the current operating condition type is identified, and the state weight matrix and control weight matrix in the objective function are adaptively adjusted and optimized. The expression of the objective function is: ;in, To predict the time domain, To control the time domain, This is the system state vector at time k+i based on the system state prediction at time k. Let be the reference trajectory state vector at time k+i. To control the rate of change of quantity, The state weight matrix is... To control the weight matrix, As a relaxation factor, , where is the penalty coefficient for the relaxation factor; S32. Dynamic setting of safety constraint boundaries: Set hard constraints including front wheel turning angle limit of tractor, upper and lower limits of traction torque or braking torque, maximum safety threshold of each articulation angle, and upper and lower limits of vehicle speed, as well as soft constraints including the collision safety distance of the whole vehicle envelope, the rate of change of control quantity, and trajectory tracking deviation. S33. Rolling Optimization Solution: Based on the state vector at the current time k, the system state matrix and control matrix, and the constraints, the optimization problem is transformed into a constrained quadratic programming problem and solved using a solver to obtain the optimal control sequence in the prediction time domain. The expression for the optimal control sequence is: ,in The optimal control sequence obtained at time k. The optimal control input at time k+i is obtained based on the system solution at time k. To control the time domain.

5. The adaptive model predictive control method based on fusion perception for civil aviation tractors according to claim 4, characterized in that, In step S31, the following weight adjustment strategies are executed according to the identified heavy-load mode, multi-bucket long array mode, or reversing mode: In heavy load mode, the penalty weight of the rate of change of control quantity is automatically increased to limit drastic changes in acceleration and deceleration and protect the towing hook; In the multi-bucket long array mode, the path deviation penalty weight of the last baggage bin at the end of the track is increased, and the penalty term for the rate of change of the articulation angle is increased to suppress the serpentine swaying of the convoy. In reverse mode, the constraint weight and deviation penalty of the articulation angle are significantly increased, prioritizing the protection against bending of the iron, and the hard constraint threshold of the articulation angle is tightened from ±30° to ±20°.

6. An adaptive model predictive control system based on fusion perception for civil aviation tractors, used to execute the adaptive model predictive control method based on fusion perception for civil aviation tractors as described in any one of claims 1-5, characterized in that, include: The fusion perception module is used to collect and output the physical parameters of the load on the luggage compartment in real time through vehicle-mounted LiDAR and vision sensors. An adaptive variable structure model construction module, connected to the fusion perception module, is used to dynamically construct and output a real-time updated system state matrix and control matrix based on the load physical parameters; An adaptive model predictive control optimization solution module is connected to the adaptive variable structure model construction module, and is used to construct a predictive model and solve for the optimal control sequence based on the system state matrix and control matrix. The control execution module, connected to the adaptive model predictive control optimization solution module, is used to convert control commands into CAN bus commands through parameter mapping and send them to the drive-by-wire chassis. The status feedback and safety monitoring module is connected to the adaptive variable structure model construction module, the adaptive model predictive control optimization solution module, and the fusion perception module, respectively. It is used to collect the real-time motion status of the tractor and each baggage compartment, use extended Kalman filtering for status smoothing, and feed back the filtered status data to the front-end modules to trigger closed-loop rolling optimization. At the same time, it monitors safety thresholds and initiates emergency intervention when abnormalities occur.

7. The adaptive model predictive control system based on fusion perception for civil aviation tractors according to claim 6, characterized in that, The fusion sensing module includes: The 6D attitude detection unit, based on a multimodal feature fusion network of PointNet++ and ResNet-50, outputs the attitude parameters and detection confidence of each compartment of the luggage compartment; The space occupancy rasterization unit is connected to the 6D attitude detection unit and is used to establish a voxelized occupancy raster in the attitude frame space, calculate the cargo volume by counting the number of occupancy voxels, generate an occupancy raster matrix, and output the occupancy raster confidence score. The mass inversion estimation unit, connected to the space occupancy rasterization unit, is used to calculate the mass and moment of inertia of a single bucket by combining density prior with visual suspension height fusion, and output the density confidence level. The confidence assessment unit is connected to the 6D attitude detection unit, the space occupancy grid unit and the quality inversion estimation unit, respectively, and is used to calculate the comprehensive confidence score according to the preset weights and output the confidence level.

8. The adaptive model predictive control system based on fusion perception for civil aviation tractors according to claim 6, characterized in that, The adaptive structural model construction module includes: The model structure adaptive adjustment unit is used to automatically construct or adjust the dimension of the state vector based on the number of luggage compartments; The multibody dynamics equation building unit is connected to the model structure adaptive adjustment unit and is used to build nonlinear multibody dynamics equations to describe the force and torque transmission at the articulation point, tire lateral slip characteristics and road surface adhesion characteristics. The linearization and discretization unit, connected to the multibody dynamics equation construction unit, is used to output the discretized state-space expression according to the control period T=0.1s. ;in, Let k+1 be the system state vector. Let k be the system state matrix at time k. Let k be the state vector at time k. Let k be the control matrix. The control input vector includes the front wheel steering angle of the tractor and the traction / braking torque. The parameter verification and online correction unit is used to compare the predicted state with the actual state collected by the state feedback and safety monitoring module. When the deviation exceeds the preset threshold, the online correction of the model parameters is triggered. If the correction still fails to meet the standard, the effective model of the previous cycle is adopted and the fusion sensing module is triggered to re-collect data.

9. The adaptive model predictive control system based on fusion perception for civil aviation tractors according to claim 6, characterized in that, The status feedback and security monitoring module further includes: The multi-source data acquisition and time synchronization unit is used to acquire the motion status of the tractor and each baggage compartment through the CAN bus, attitude sensor and articulation angle sensor, and perform time synchronization to control the synchronization error to be less than or equal to a preset threshold. The data rationality verification and filtering fusion unit is connected to the multi-source data acquisition and time synchronization unit. It is used to remove abnormal data that exceeds the physical reasonable range and to output smooth motion state parameters, including tractor position, vehicle speed, yaw rate and articulation angle of each section, using the extended Kalman filter algorithm. The safety threshold judgment and process control unit is connected to the data rationality verification and filtering fusion unit. It is used to monitor the trajectory tracking deviation, the hinge angle of each section, and the collision risk coefficient in real time. When the hinge angle exceeds the hard constraint threshold or the collision risk coefficient is greater than or equal to the preset value, an emergency braking command is immediately triggered and the autonomous driving process is terminated. When a preset frame of CAN data is lost, data re-acquisition is triggered and a slow braking command is sent.

10. The adaptive model predictive control system based on fusion perception for civil aviation tractors according to claim 6, characterized in that, It also includes lightweight alternative implementation architectures: The fusion perception module uses a multi-view vision sensor to replace LiDAR, and uses binocular stereo vision to measure the geometric parameters and loading capacity of the luggage compartment. The adaptive variable structure model construction module uses a simplified bicycle model and hinge angle kinematic model to replace the complex multibody dynamics model; The adaptive model predictive control optimization solution module adopts an explicit model predictive control scheme, pre-calculates control laws offline, and calls them online by looking up tables to adapt to low-computing-power edge computing platforms.

Images