A multi-agv stress self-adaptive cooperative control method for weak-rigid cargo transportation

Abstract

The application discloses a multi-AGV stress adaptive cooperative control method for weak rigid cargo carrying, and belongs to the technical field of automatic guided vehicle cooperative control. The method comprises the following steps: acquiring force and torque data of a cargo supporting point when acquiring an AGV running state in real time, introducing a trained neural network model to obtain a stress set of the cargo, determining a stress safety index based on the stress set, embedding a decision and planning layer optimization objective function into a stress safety item including the stress safety index to obtain a planning path considering stress, and then performing subsequent trajectory optimization and closed-loop control execution steps. The method establishes a technical route of indirectly sensing internal stress through external reaction force, realizes nondestructive sensing of the internal mechanical state of the cargo, and applies the sensing result to the decision and planning layer of cooperative control, so that a trajectory with low stress in nature is generated, the final control result is safer, and the mechanical safety of the cargo in the carrying process is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic guided vehicle cooperative control, and particularly relates to a multi-AGV stress adaptive cooperative control method for weak rigid cargo handling. BACKGROUND

[0002] Multi-AGV cooperative handling technology is a key link in the fields of intelligent logistics and intelligent manufacturing. The existing technology (such as the Chinese patent application for invention with the application number 202510338043.9 and the application date of 2025.03.21, and the name of a multi-AGV cooperative control method for cooperative handling of heavy objects) mainly solves the problems of formation keeping, motion synchronization and geometric obstacle avoidance on an ideal horizontal plane. The optimization objectives of this kind of technology are focused on geometric motion parameters such as path efficiency and pose accuracy. However, when the above-mentioned existing technology is applied to the handling of weak rigid high-value cargos such as aircraft composite skin, aerospace composite wallboard and large glass panels, especially under the condition of uneven road surface existing in factories and airports, there are inherent defects that the theoretical framework cannot overcome:

[0003] Mechanical safety objective definition is fuzzy: the control framework of the existing technology lacks an effective bearing mechanism for the core control objective of "cargo mechanical safety". This objective only exists as a vague and qualitative boundary condition and is not modeled as a specific state variable that can be optimized with kinematic objectives. This fuzzy definition makes it impossible to embed it into the optimization function of the controller, so that the control system cannot understand and respond to the real mechanical safety needs of the cargo in essence, and cannot quantify and track and protect the mechanical safety.

[0004] Mechanical perception is missing: the existing technology cannot perceive the stress and strain caused by internal position mismatch (such as different support heights of each AGV due to uneven ground) or dynamic impact (such as sudden stop and turning) of the cargo. It cannot ensure the mechanical safety of the cargo. Its sensor system such as encoder, IMU and laser radar can only monitor the pose of the AGV itself and the external environment, and there is a lack of perception of the core safety state of the cargo.

[0005] Control objective deviation: the control model and objective function of the existing technology only contain kinematic quantities such as position, velocity and acceleration, and never take stress as a control objective, which directly determines whether the weak rigid cargo will be irreversibly damaged. This causes a serious disconnection between the optimization direction of the system and the mechanical safety needs of the cargo in essence.

[0006] Compensation mechanism lag: the existing technology generally adopts a passive control mode of "monitoring-response", that is, compensation is carried out when the pose error is monitored. However, for weak rigid goods, when the pose error is detected, the internal plastic deformation or damage that cannot be recovered may have been generated. The system lacks the ability of active feed-forward compensation to terrain changes and the ability of forward-looking suppression to dynamic impact.

[0007] The root of these problems lies in the fact that the theoretical framework of the existing technology is limited to the kinematics and geometry level, lacking the core element of the mechanical state of the goods. The model, perception and control target are seriously disconnected from the mechanical safety target required for weak rigid goods handling. Therefore, there is an urgent need for an innovative collaborative control method that takes the mechanical safety of the goods as the core and has the ability of active stress adaptation. SUMMARY

[0008] Therefore, the purpose of the present application is to provide a multi-AGV stress self-adaptive collaborative control method for weak rigid goods handling, to solve the problem that the existing collaborative planning is only limited to the kinematics and geometry level, lacking the core element of the mechanical state of the goods, which is seriously disconnected from the mechanical safety target required for weak rigid goods handling, resulting in mechanical safety risks of the goods under non-flat road surfaces and dynamic working conditions.

[0009] To achieve the above-mentioned purpose, the method of the present application establishes the technical route of "indirectly perceiving the internal stress through external reaction force", realizes non-destructive perception of the internal state of the goods, and applies the perception results to the decision and planning level of collaborative control, so as to generate a trajectory with essentially low stress, so that the final control result is safer and ensures the mechanical safety of the goods during handling. Specifically, a six-axis force sensor is arranged at the position of each AGV for supporting the goods, the six-axis force sensor is used to acquire force and torque data of the supported point of the goods, and the control method comprises the following steps:

[0010] Step 1: When multiple AGVs collaboratively handle the goods, real-time acquisition of the chassis attitude information, platform attitude information, speed information, digital terrain model information in front of the path and data of the six-axis force sensor of each AGV; inputting the data of each six-axis force sensor into a trained neural network model to obtain a stress set of the goods, the stress set comprising stress values at each set position of the goods; selecting one of the results obtained by dividing the stress value at each set position by the pre-set stress safety threshold value at the set position as a stress safety index of the stress set according to a set selection rule;

[0011] The training of the neural network model comprises: taking data in a stress mapping database corresponding to force and torque data of each supported point of the cargo and stress values at each set position of the cargo as a data set in a training process, and training the neural network model, so that the trained neural network model can obtain stress values at each set position of the cargo according to force and torque data of each supported point of the cargo.

[0012] Step 2, according to the information obtained in real time in step 1 and the stress safety index, a preliminary space-time trajectory is obtained by using a nonlinear optimization solver, wherein an optimization objective function of the nonlinear optimization solver comprises a stress safety item of the stress safety index; a stress safety index sequence in a future period of time is predicted according to the preliminary space-time trajectory, and if each stress safety index in the stress safety index sequence is less than a set value, the preliminary space-time trajectory is taken as a high-level reference trajectory, otherwise, trajectory re-planning is performed, and the re-planned trajectory is taken as the high-level reference trajectory;

[0013] Step 3, the high-level reference trajectory is optimized by using a dynamics model to obtain an MPC reference trajectory, and an MPC controller takes the MPC reference trajectory as a tracking target, and dynamically solves an optimal control amount according to real-time state feedback of the AGV, and outputs the optimal control amount to an AGV actuator.

[0014] The method comprises the following steps: 1, acquiring force and torque data of the cargo, and mapping the force and torque data into a stress value set of key points of the cargo by using a neural network model; 2, calculating a stress safety index based on the stress value set; and 3, generating a low-stress trajectory based on the stress safety index.

[0015] Further, in step 1, the establishment of the stress mapping database comprises: establishing a finite element model of the cargo, in a finite element software, applying a force and torque combination covering the expected working range to positions in the finite element model corresponding to actual positions of each supported point of the cargo, performing statics analysis to obtain stress values of each set position of the cargo corresponding to different force and torque combinations, forming a corresponding relationship between the multiple groups of force and torque combinations and the stress values of each set position of the cargo corresponding thereto, and obtaining the stress mapping database.

[0016] The method of the application systematically establishes a quantitative mapping relationship from the "external load" to the "internal stress field" by high-fidelity reproduction of the mechanical state of the physical entity in the virtual space, so as to realize non-destructive and real-time estimation of the internal stress of the cargo by using the mapping relationship.

[0017] Further, in step 1, the set selection rule is that, in the results of dividing the stress values at each set position by a pre-set stress safety threshold value at the set position, a result with the largest value in the results is selected as a stress safety index of the stress set; and the determination of the pre-set stress safety threshold value at the set position comprises: obtaining an allowable stress value at the set position of the cargo according to a design specification of the cargo, and setting the pre-set stress safety threshold value at the set position as less than the allowable stress value at the set position of the cargo.

[0018] By selecting the maximum value, it is ensured that the attention of the system is always focused on the most dangerous key point and outputs a dimensionless scalar value (i.e., the stress safety index), which maps a high-dimensional stress vector to a one-dimensional system state signal with extremely clear physical meaning, greatly simplifying the complexity of subsequent decision-making and control.

[0019] Further, in step 2, the stress safety term has the highest priority compared to other terms in the optimization objective function.

[0020] By setting the highest priority, the planner actively avoids aggressive movements that will cause stress to exceed the limit, thereby outputting a mechanically essentially safe spatiotemporal trajectory.

[0021] Further, in step 2, the prediction of the stress safety index sequence in the future period of time comprises: predicting the platform height and attitude information in the future period of time according to the preliminary spatiotemporal trajectory, and estimating the forces and torques required for maintaining the static balance of the cargo by solving force and torque balance equations based on the known mass, center of gravity position and geometric dimensions of the cargo, inputting the forces and torques into the trained neural network model to obtain the stress safety index sequence in the future period of time.

[0022] This invention constructs a time-series prediction sequence of stress safety indicators, enabling the system to perceive future mechanical potentials. This represents a fundamental leap from real-time perception of only the "current physical state" to predictive perception of the "future state evolution." This capability provides crucial forward-looking information for core decision-making, shifting the intervention point of the control system from "after stress exceeds the limit" to "before stress exceeds the limit," laying an irreplaceable perceptual foundation for achieving true proactive immunity.

[0023] Furthermore, in step 2, trajectory replanning includes: taking the current state as the starting point, taking the time after a set time interval as the virtual stop line, and using the virtual stop line as a constraint to generate a smooth deceleration trajectory leading to the virtual stop line.

[0024] When the risk of stress exceeding the limit is predicted at a certain point in the future, decision-making intervention is initiated and a forward-looking replanning mechanism is triggered to plan a smooth deceleration trajectory towards the "virtual stop line" instead of emergency braking, thereby keeping the acceleration low and proactively suppressing dynamic shocks from a dynamic perspective.

[0025] Furthermore, in step 3, dynamically solving for the optimal control quantity includes: the MPC controller predicts the stress safety index sequence over a future period for each candidate control sequence within it; the cost function of the MPC controller includes a trigger-based penalty term based on the predicted stress safety index; and the optimal control quantity is obtained by using the MPC controller containing the cost function of the MPC controller.

[0026] This invention also considers the influence of stress factors in the trajectory optimization and closed-loop control execution steps, dynamically solves the optimal control command, and achieves accurate, compliant, and safe final execution.

[0027] Furthermore, in step 3, the triggering penalty term based on the predicted stress safety index is: when all predicted stress safety indices are less than 1, the triggering penalty term is zero; when there is a predicted stress safety index that is not less than 1, the triggering penalty term is much larger than other terms in the cost function of the MPC controller.

[0028] The method of this invention ensures that this term only takes effect and penalizes the excess stress when the predicted stress exceeds the limit. Within the safe zone, the term is zero and does not affect tracking performance, thus achieving intelligent dual-modal control. Furthermore, this trigger-based penalty term is much larger than other terms in the cost function of MPC, ensuring that once a risk is predicted, the stress reduction objective unconditionally and overwhelmingly dominates the optimization process, forcing the controller to perform risk avoidance operations.

[0029] Further, in step 2, while planning the trajectory, according to the pre-look information of the digital terrain model and the time and space mapping relationship of the planned trajectory, the platform height feedforward instruction required to keep the goods level is reversely calculated, and the timing of the platform height feedforward instruction is ahead of the forward pre-look time dynamically calculated by the AGV chassis reaching the corresponding terrain point; the dynamically calculated forward pre-look time is the sum of the inherent response delay of the platform actuator, the travel time required for the platform to move from the current position to the target height, and the inherent delay of system data processing and communication.

[0030] The dynamic calculation of the forward pre-look time by the above method ensures that the completion time of the platform leveling action is accurately synchronized with the time when the AGV chassis reaches the corresponding terrain point, thereby fundamentally avoiding static configuration mismatch. The core value of this mechanism is that by pre-looking at the future terrain and issuing precisely synchronized leveling instructions in advance, the traditional lag control is changed to lead compensation, thereby actively eliminating stress before the physical conditions of static stress occur.

[0031] Further, in step 3, while obtaining the MPC reference trajectory by optimization, the feedforward control amount for tracking the MPC reference trajectory is also solved, and the feedforward control amount is output as part of the MPC reference trajectory attributes.

[0032] The method of the present application makes the finally output MPC reference trajectory not only smoother and more accurate than the high-level reference trajectory in the state sequence, but more importantly, it embeds the feedforward control instruction, which makes the bottom-level MPC controller not obtain a "target" that needs to be completely overcome through feedback, but a "target-instruction" joint reference. On this basis, the MPC can focus on compensating for model mismatch and instantaneous disturbances, greatly improving tracking accuracy and control efficiency.

[0033] The above description is only a summary of the technical solutions of the present application, in order to more clearly understand the technical means of the present application, the contents of the specification can be implemented, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are as follows. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is the system total architecture and full-link stress self-adaptive closed-loop information flowchart of the method of the present application;

[0035] Figure 2 is the stress real-time sensing flowchart based on digital twinning of the present application;

[0036] Figure 3 is the stress self-adaptive decision and planning layer structure diagram of the present application;

[0037] Figure 4The trajectory rolling optimization and feedforward instruction generation flowchart of the application is as follows:

[0038] Figure 5 The model predictive control structure based on stress prediction of the application is as follows. DETAILED DESCRIPTION

[0039] The technical solutions of the application will be described clearly and completely in combination with specific implementation solutions, but the person skilled in the art should understand that the implementation solutions described below are only used for illustrating the application and should not be regarded as limiting the scope of the application. Based on the implementation solutions in the application, all other implementation solutions obtained by the person skilled in the art without making creative efforts belong to the scope of protection of the application.

[0040] Multi-AGV stress self-adaptive cooperative control method for weak-rigidity cargo handling

[0041] To solve the core technical problem that the existing multi-AGV cooperative control method cannot guarantee the mechanical safety of weak-rigidity cargo under non-flat road surface and dynamic working conditions, the application proposes a multi-AGV cooperative control method based on digital twinning and stress safety index driving full-link stress self-adaptive closed loop. The core idea of the method is to build an intelligent control system that actively maintains the mechanical safety of cargo with "stress safety index ζ" as the unified language and driving core, through the closed-loop linkage of four core technologies of stress real-time sensing, stress safety index construction, stress forward-looking prediction and prediction driving control, the system has the ability of self-stress sensing, self-safety evaluation and self-decision optimization, and finally realizes the fundamental paradigm shift from "geometric cooperation" to "mechanical safety cooperation".

[0042] The method of the application will be introduced from the system architecture and information flow of the method, the core technology of the method and the full-link stress self-adaptive closed loop formed by the method of the application as follows:

[0043] (1) System architecture and information flow: five-layer intelligent control architecture.

[0044] The systematic innovation of the application is first reflected in the top-level architecture design. The system is logically divided into three functional domains of perception and cognition, decision and planning, and execution, and is specifically implemented through the following five-layer intelligent control architecture, ensuring that the stress safety target can run through the whole process of perception, decision, planning and execution from the source of system design. The core information flow between each level is as shown in Figure 1 , ensuring the closed-loop transmission of stress safety index ζ.

[0045] 1. Perception and cognition domain: responsible for acquiring data from the physical world and deeply understanding the mechanical state of the cargo. It includes:

[0046] Layer 1 - Multi-source information perception layer: As a data base, it is responsible for real-time collection of four types of key information: DTM information is obtained through laser radar and pre-stored map; AGV chassis attitude and speed information are obtained through GPS, IMU and encoder; platform attitude information [φ_i, ψ_i] (i=1, 2, 3, 4) is obtained through an inclination sensor, where φ_i represents the i-th AGV platform pitch angle and ψ_i represents the i-th AGV platform roll angle; vertical force and bending moment information of each support point, i.e. force / torque vector [Fz_i, Mx_i, My_i] (i=1, 2, 3, 4), is obtained through a six-dimensional force / torque sensor installed at each AGV single support point, where Fz_i represents the vertical support reaction force at the i-th AGV support point, Mx_i represents the bending moment around the X-axis at the i-th AGV support point, and My_i represents the bending moment around the Y-axis at the i-th AGV support point.

[0047] Layer 2 - Stress state estimation layer (digital twin layer): This is the core perception innovation of the present application. This layer receives force / torque data from the perception layer, and through an embedded lightweight cargo mechanics model (such as a neural network model based on offline finite element simulation calibration), it maps the full-dimensional force / torque information of the support points to a set of stress estimation values σ_estimated of the key points of the cargo in real time, and calculates the global stress safety index ζ accordingly, thereby realizing non-destructive perception of the internal state of the cargo.

[0048] 2. Decision and planning domain: responsible for forward-looking strategic decision-making and real-time tactical optimization based on stress safety index. It includes:

[0049] Layer 3 - High-dimensional prediction and planning layer (strategic decision-making layer): This is the core decision-making innovation of the present application. This layer receives stress safety index ζ and upper-level instructions, and performs two key tasks: first, in the high-dimensional state space that integrates planar motion [x, y, θ] (where x, y, θ represent AGV horizontal coordinate, vertical coordinate and yaw angle) and spatial attitude [z, φ, ψ] (z, φ, ψ represent support point height coordinate, platform pitch angle and platform roll angle), it performs stress-embedded joint trajectory planning, with the optimization objective function explicitly containing stress safety items, thereby generating a reference trajectory that is essentially safe; second, its internal dynamics model and stress prediction mechanism predict future stress trends, implement forward-looking decisions (such as triggering a "virtual stop line"), and determine the macro behavior strategy of the system.

[0050] Layer 4 - Trajectory Rolling Optimization Layer (Tactical Decision Layer): This layer receives the rough trajectory from the planning layer, uses a high-fidelity dynamic model to perform rolling optimization in a local prediction time domain (1.0-3.0 seconds), and generates a high-precision, executable trajectory containing both state references and feedforward control quantities. It is responsible for making local optimization decisions to eliminate deviations between ideal models and actual systems, and converting strategic planning into precise tactical actions.

[0051] 3. Execution Domain: Responsible for accurately and reliably executing decision instructions. It includes:

[0052] Layer 5 - High-Precision Closed-Loop Control Layer: As an execution unit, it is composed of a model predictive controller (MPC). It receives the optimized trajectory and real-time state feedback of the system (including stress safety indicators ζ), solves the optimal control problem at each control period, and introduces a penalty term based on stress prediction values into the objective function (i.e., the cost function of MPC), directly issuing drive instructions to AGV actuators, forming the last barrier for the system to resist disturbances.

[0053] The fundamental innovation of this architecture is that by using the core variable of stress safety indicators ζ, the three domains and five levels are tightly coupled to form a complete stress self-adaptive closed loop, realizing the leap from "passive reaction" to "active adaptation".

[0054] (II) Core Technology 1: Real-time Perception of Cargo Stress Based on Digital Twin

[0055] To solve the problem of "lack of mechanical perception", the invention constructs a stress digital twin of the cargo. The method flow is shown in Figure 2 , which high-fidelity reproduces the mechanical state of physical entities in virtual space through "offline simulation calibration and online real-time mapping", thereby achieving non-destructive and real-time estimation of internal stress of the cargo.

[0056] 1. Offline modeling and calibration: High-fidelity stress mapping dataset construction.

[0057] This step is the precision cornerstone of stress perception, aiming to systematically establish the quantitative mapping relationship from "external load" to "internal stress field".

[0058] a) Physical quantity definition and sensor mapping of key parameters:

[0059] Mechanical input vector Input: Precisely defined as [Fz_1, Mx_1, My_1, Fz_2, Mx_2, My_2, Fz_3, Mx_3, My_3, Fz_4, Mx_4, My_4]. [Fz_1, Fz_2, Fz_3, Fz_4] corresponds to the vertical support reaction forces at the four AGV support points, and [Mx_1, My_1,..., Mx_4, My_4] corresponds to the bending moments around the X and Y axes at the four support points. This variable directly corresponds to the measured values of the six-axis force / torque sensor installed on the AGV cargo platform.

[0060] Stress output vector Output: Precisely defined as [σ_P1, σ_P2,...,σ_Pi,..., σ_Pn], where σ_Pi is the equivalent stress of the i-th key monitoring point in the cargo finite element model. These key points are pre-selected according to the CAD model of the cargo, material properties and offline statics analysis results, ensuring coverage of all vulnerable areas.

[0061] b) Finite element simulation data generation covering all working conditions:

[0062] In the finite element software, a parameterized high-fidelity model consistent with the physical cargo is established. By applying tens of thousands of force and torque combinations covering the expected working range on its four support points, batch statics simulation is performed, and the corresponding Input and Output are recorded. Finally, the corresponding relationship data between multiple mechanical input vectors and stress output vectors is constructed into a large-scale, high-dimensional stress mapping database D = {Input_j→Output_j}, j = 1, 2,..., N, where Input_j represents the j-th corresponding mechanical input vector, and Output_j represents the stress output vector corresponding to the j-th corresponding mechanical input vector.

[0063] 2. Real-time mapping model training: Condensation of high-precision lightweight proxy model.

[0064] This step aims to condense the high-fidelity but computationally expensive FEA database into a lightweight proxy model suitable for real-time control, with high precision and high speed.

[0065] a) Data preprocessing and robustness division:

[0066] Import the stress mapping database D into the computing environment, and perform Z-score standardization processing on Input_j and Output_j in D respectively. Then, the standardized data is randomly divided into training set, validation set and test set according to the ratio of 7:2:1.

[0067] b) Neural network model structure and training strategy:

[0068] Model Architecture: A deep feedforward neural network (DNN) is adopted as the real-time mapping model M. The number of neurons in the input and output layers strictly corresponds to the dimensions of the mechanical input vector and stress output vector. The hidden layers use a three-layer decreasing structure of [128, 64, 32], with ReLU activation functions in the hidden layers and a linear activation function in the output layer.

[0069] Optimization Training and Regularization: The Adam optimizer (initial learning rate 0.001) is used, with the root mean square error (RMSE) as the loss function, to train the model on the training set. During training, early stopping is achieved by monitoring the validation set loss, and L2 regularization and Dropout techniques are introduced to effectively suppress overfitting.

[0070] Performance Evaluation: After training, the model performance is evaluated on an independent test set. It is considered qualified if the R² coefficient of determination for stress prediction is >0.98 and the RMSE is less than 5% of the material allowable stress.

[0071] 3. Online Estimation and Feedback:

[0072] When the system is running online, the following processes are executed synchronously within each control period (e.g., 10 ms):

[0073] a) Data Collection and Fusion: Real-time collection of six-dimensional force / torque data [Fz_i, Mx_i, My_i] from the four support points.

[0074] b) Forward Calculation and Stress Estimation: The normalized six-dimensional force / torque data from the four support points is input into the deployed lightweight neural network model M. The model completes forward propagation within milliseconds, outputting the stress estimation value set σ_estimated for all key points of the cargo at the current time.

[0075] c) Feedback and Decision: σ_estimated is immediately sent to the subsequent module for calculating the stress safety index ζ.

[0076] (Three) Core Technology Two: Construction of a Unified Stress Safety Index

[0077] This is a control theory innovation based on "stress perception" to solve the problems of "control target deviation" and "system decision complexity." The invention does not stop at obtaining stress data, but creatively constructs a unified stress safety index ζ. The transmission process of the index ζ in the system architecture is shown in Figure 1 , which serves as the "unified language" and "decision cornerstone" for the entire collaborative control system to understand and ensure cargo safety. The innovation of this technology is reflected in the following aspects:

[0078] 1. As a global unified quantitative evaluation index.

[0079] The present application discards the cumbersome mode of isolated monitoring of stress values in traditional methods, and realizes unified and quantitative evaluation of the overall safety state of the system by defining a scalar index ζ. The calculation formula of the index is:

[0080] ζ=max(σ_estimated[i] / σ_safe[i] ) where i is the index of all monitored key points of the goods, σ_safe[i] is the safety stress threshold of each point, and σ_estimated[i] is the stress value of each point included in the stress estimation set σ_estimated.

[0081] The core of the formula is the max() operation, which ensures that the attention of the system is always focused on the most dangerous key point at present and outputs a dimensionless scalar value ζ, which maps the high-dimensional σ_estimated to a one-dimensional system state signal with extremely clear physical meaning, greatly simplifying the complexity of subsequent decision and control.

[0082] 2. As a deterministic criterion for system safety.

[0083] The index ζ is designed as a clear and unambiguous safety criterion, providing the system with the ability to make instantaneous decisions.

[0084] When ζ<1, it indicates that the stress of all key points is below their respective safety thresholds, and the system is in an absolutely safe zone. In this state, the control system can focus on path tracking accuracy, energy efficiency optimization and other performance goals.

[0085] When ζ≥1, it indicates that the stress of at least one key point has reached or exceeded its carrying limit, and the system enters the risk zone or the over-limit zone. This state will unconditionally trigger the priority intervention mechanism of the system.

[0086] Value: This deterministic to quantitative transformation makes "safety" no longer a vague concept, but an explicit goal that can be measured and strictly guarded in real time.

[0087] 3. As a bridge connecting the physical world and information decision.

[0088] The stress safety index ζ plays a crucial role as an "information hub" in the system. Downlink: it encapsulates complex physical field information (stress distribution) from the digital twin. Uplink: it exposes to the planning and control layer a simple objective function or constraint condition that is naturally compatible with the control system optimization framework. Role: this "bridge" solves the "semantic gap" problem faced by control theory in dealing with complex physical objects, enabling advanced algorithms such as model-based predictive control to directly model and optimize the mechanical safety of goods.

[0089] 4. The intelligent switch that triggers the mode switching.

[0090] Based on the explicitness of the indicator ζ, the invention realizes a dual-mode, intelligent switching control strategy.

[0091] Mode 1 (performance priority mode): when ζ < 1, the system runs in a high-performance tracking mode, and the controller strives to make the AGV cluster accurately, smoothly, and efficiently track the planned trajectory.

[0092] Mode 2 (safety priority mode): when ζ ≥ 1, the system immediately switches to the safety braking and recovery mode. At this time, the core goal of the controller changes to maximize the reduction of the indicator ζ, even if it temporarily sacrifices tracking accuracy.

[0093] Advantages: This mode switching based on accurate measurement is more adaptive and intelligent than fixed strategies based on time or experience, fundamentally avoiding the efficiency loss caused by conservative control or the safety risk caused by aggressive control.

[0094] In summary, the construction of the stress safety indicator ζ is not a simple mathematical formula, but a core methodology created by the invention to achieve "mechanical safety coordination". It unifies the safety concept of the system, simplifies the complexity of decision-making, and ultimately makes the entire multi-AGV system an intelligent organism that can self-perceive risks, self-assess situations, and self-decide actions.

[0095] (Four) Core Technology Three: Stress Anticipation Estimation Mechanism Based on Model Prediction

[0096] The key innovation of the invention is that the system not only perceives the current stress, but also predicts the stress safety state at future time, laying the foundation for forward-looking control. This mechanism is independent of specific control actions and provides a unified, forward-looking safety situation assessment for all subsequent decisions.

[0097] 1. Architecture and process of stress prediction:

[0098] The stress prediction mechanism is an independent service module, with the system's planned trajectory or control sequence as input and a sequence of predicted stress safety indicators over a future period as output. To accurately describe this time series prediction process, the invention uses the following representation method:

[0099] t represents the current time for which the prediction calculation is performed; k represents the prediction step index (k = 0, 1, 2,..., N_p-1), where k = 0 corresponds to the current time t, k = 1 corresponds to the next sampling time t+ΔT, and so on; (k|t) represents "prediction k steps ahead at time t" to clearly distinguish between the predicted value and the actual measured value of a variable.

[0100] Based on this, the core of the prediction mechanism is a series of prediction architecture, the process is as follows:

[0101] a) System state prediction: Based on the high-fidelity dynamic model of AGV and leveling platform, according to the given future control sequence or the currently tracked planning trajectory, a series of state sequences of the system in the future prediction time domain N_p is predicted , which includes the position, velocity, acceleration of AGV and the predicted height of the platform. Here, , which represents the predicted motion state at time t.

[0102] b) Support reaction force and torque estimation: input the predicted platform height and attitude information into an embedded statics calculation module. Based on the known mass, center of gravity position and geometric size of the goods, the required support point force and torque at each future time to maintain the static balance of the goods (That is, the vertical support reaction force of each support point, the bending moment around the X axis and the bending moment around the Y axis required to maintain the static balance of the goods).

[0103] c) Stress safety index mapping: the estimated future force / torque vector is input into the lightweight digital twin model (i.e. the neural network model M in core technology one). The process is as shown in Figure 2 , which quickly outputs the future predicted stress distribution sequence and calculates the future predicted stress safety index:

[0104] .

[0105] The physical meaning is that, at the current time t, based on all available information, the predicted stress safety index sequence at time t+k·ΔT in the future is obtained.

[0106] 2. Application interface of stress prediction mechanism:

[0107] This stress prediction mechanism serves as a general service and is called by multiple upper-level modules in the system:

[0108] a) Called by trajectory planning layer: used to evaluate the stress safety of candidate trajectories on the entire time axis.

[0109] b) Called by model predictive controller (MPC): used to evaluate the future stress risk corresponding to different control sequences, so as to avoid in advance in optimization.

[0110] (Five) Core technology four: full-link adaptive control based on stress prediction.

[0111] This is an innovative implementation of stress self-adaptive coordination at the execution level, which utilizes the predictive stress information provided by Core Technology Three, and through three control mechanisms of feedforward compensation, model prediction, and real-time feedback, respectively solves the stress control problem under static, dynamic, and transient disturbance.

[0112] 1. Static stress elimination: terrain interactive preview and active feedforward compensation.

[0113] Control objective: Eliminate the quasi-static bending stress caused by the mismatch of support point shape and position due to uneven road surface from the physical root cause.

[0114] To solve the "timing mismatch" problem of the chassis and platform motion, the static stress caused by the mismatch of motion and leveling instructions is avoided from the root cause. This invention introduces a terrain perception feedforward control mechanism:

[0115] (1) Terrain interactive preview:

[0116] This step aims to give the system a forward-looking perception ability to obtain accurate future terrain disturbance information.

[0117] a) Preview information fusion: Based on the planned path of AGV and high-precision digital terrain model (DTM), the system performs spatial mapping and query to obtain elevation information z_terrain(s) in front of the path in real time.

[0118] b) Space-time conversion prediction: Combined with the current state of AGV and the reference speed curve v(t), through the established space-time mapping relationship s(t), the elevation information z_terrain(s) on the spatial path is converted into a predicted disturbance sequence z_terrain(t) that changes with time, thus accurately predicting the terrain changes that the AGV chassis will encounter at a specific future time point.

[0119] (2) Active feedforward compensation:

[0120] a) Instruction inverse calculation: With the terrain disturbance information z_terrain(t) obtained by looking ahead and the expected horizontal attitude of the goods as input, through the kinematics-dynamics model, the leveling platform height target instruction z_platform(t) required to completely compensate for the disturbance is calculated.

[0121] b) Timing synchronization design: To achieve precise synchronization of compensation action and terrain interference, the system precisely controls the timing of the instruction issuance. A dynamically calculated forward preview time t_forward is introduced to comprehensively compensate for the mechanical response delay and movement stroke of the platform execution mechanism. Wherein, t_forward = t_response + t_travel + t_processing, t_response is the inherent response delay of the platform actuator (such as electric cylinder) which is determined by the equipment specification, t_travel is the stroke time required for the platform to move from the current position to the target height, t_processing is the inherent delay of system data processing and communication. Through the above method of dynamically calculating t_forward, it ensures that the completion time of the platform leveling action is accurately synchronized with the time when the AGV chassis arrives at the corresponding terrain point, thereby avoiding the static shape mismatch from the root.

[0122] The core value of this mechanism is that by previewing the future terrain and issuing precise and synchronized leveling instructions in advance, it changes the traditional lag control into advance compensation, thereby actively eliminating stress before the physical conditions of static stress occur.

[0123] 2. Dynamic stress suppression: Stress-embedded planning and forward-looking decision-making.

[0124] Control objective: Active suppression of inertial force and dynamic impact stress generated by sudden changes in the system's own motion state (such as acceleration, deceleration, and turning).

[0125] a) Stress-embedded high-dimensional space-time joint planning: Unlike the traditional serial mode of first planning a plane path and then planning the height, the invention performs a one-time joint optimization in a six-dimensional state space composed of [x, y, θ, z, φ, ψ]. Its optimization objective function is:

[0126] ;

[0127] Wherein the definitions and roles of key parameters are as follows:

[0128] J plan represents the space-time joint planning objective function, t0 represents the planning initial time, t f represents the planning terminal time.

[0129] J track (Trajectory tracking term): Punish the deviation of system state X(t) (pose, velocity, etc.) from the expected reference trajectory X _plan (t). The weight matrix Q determines the tracking accuracy of each state component.

[0130] J control(Control cost term): penalize excessive control effort U(t) (e.g. motor torque, hydraulic thrust), regulated by weight matrix R, related to system energy consumption and actuator lifetime.

[0131] J smooth (Motion smoothness term): penalize the rate of change of control effort (Jerk), regulated by weight matrix S. This term is crucial for suppressing dynamic impact stress, which directly encourages the planner to generate a smooth acceleration profile.

[0132] J stress (Stress safety term): the core innovation of the invention. Directly use the square of stress safety indicator ζ(t) as the penalty term. The weight wσ must be set as a dominant value (e.g. 10 3 ~10 5 times of Q's element) to ensure that the stress safety objective has the highest priority over tracking accuracy, energy consumption and smoothness objectives in the optimization process. This makes the planner actively avoid aggressive motions that will cause stress overrun, thus outputting a mechanically intrinsically safe trajectory in time and space.

[0133] The optimization objective explicitly includes the stress safety term, making the planner prioritize the stress safety indicator ζ as one of the primary optimization objectives when coordinating all degrees of freedom. This generates an intrinsically low-stress trajectory X_plan(t) that is not only geometrically feasible, but also mechanically safe.

[0134] b) Model prediction-based forward-looking decision: the planning layer invokes a stress prediction mechanism as shown in Figure 3 , continuously predicting the trend of the system's stress safety indicator ζ in the future (5-30 seconds) to evaluate the safety of the generated trajectory X_plan(t). When it predicts that there is a risk of stress overrun at a certain time in the future (i.e. ≥1) (e.g. due to an imminent sharp turn of the AGV), it initiates decision intervention and triggers the forward-looking re-planning mechanism. For example, it plans a smooth deceleration trajectory to a "virtual stop line" instead of emergency braking, thereby controlling the jerk to a low level and proactively suppressing dynamic impact from a dynamic perspective. The modified trajectory is output as the high-level reference trajectory X_ref_high(t).

[0135] 3. Trajectory refinement and command generation: differential smoothing and reference trajectory reconstruction based on high-fidelity models.

[0136] Control objective: solve the problem of high-frequency jitter and dynamic infeasibility of the reference trajectory X_ref_high(t) output by the strategic planning layer due to model simplification and discretization, and generate a bottom-layer execution command that has both high-order smoothness and high dynamic tracking accuracy.

[0137] The flow of the trajectory rolling optimization layer is shown in Figure 4 After receiving the high-level reference trajectory X_ref_high(t), the layer performs a rolling optimization based on a high-fidelity system dynamics model within its local prediction horizon H_tactic(1.0-3.0 seconds). The core innovation lies in the synergistic optimization of differential smoothing trajectory generation and feedforward command embedding:

[0138] a) Differential smoothness guarantee:

[0139] The optimization process explicitly imposes constraints on the high-order derivatives of the control variables (such as jerk), thereby ensuring that the generated trajectory X_ref_mpc(t) is at least twice continuously differentiable at the dynamics level. This fundamentally suppresses the inertial impact stress caused by sudden acceleration changes, ensuring the essential smoothness of motion.

[0140] b) Synergistic solution of feedforward command:

[0141] Unlike traditional methods that only optimize state trajectories, this layer simultaneously optimizes X_ref_mpc(t) and, based on high-fidelity model forward simulation, synergistically solves the ideal feedforward control variable U_ff(t) required to accurately track this trajectory, and outputs it as part of the trajectory attribute.

[0142] c) Reference trajectory reconstruction:

[0143] The final output MPC reference trajectory X_ref_mpc(t) is a reconstructed and enhanced reference signal. It is not only smoother and more accurate than X_ref_high(t) in the state sequence, but more importantly, it embeds the feedforward control command. This allows the bottom-layer MPC controller to obtain not a "target" that needs to be completely overcome through feedback, but a "target-command" joint reference, on which the MPC can focus on compensating for model mismatch and transient disturbances, greatly improving tracking accuracy and control efficiency.

[0144] 4. Adaptive closed-loop control: model predictive control based on stress safety indicators.

[0145] Control objective: To build the last barrier against unknown disturbances and model mismatches, dynamically optimizing the trade-off between time safety and performance.

[0146] a) Mechanical safety indicator control:

[0147] The MPC closed-loop control structure is shown in Figure 5As shown, the MPC controller receives the optimized trajectory X_ref_mpc(t) and system real-time state feedback to solve the following finite horizon optimization problem at each control period. The optimization problem adopts the ζ-based smart design to add a stress optimization control term in its objective function. During the solving process, the MPC invokes the stress prediction mechanism described in Core Technology Three to calculate the corresponding future stress safety index sequence for each candidate control sequence inside it .

[0148]

[0149] Where the key parameters are defined and their roles are:

[0150] J mpc represents the cost function of the MPC controller at each control period, N P represents the prediction horizon, N P -1 represents the last step of the prediction horizon, N C represents the control horizon, N C -1 represents the last step of the control horizon.

[0151] J track_mpc (MPC tracking term): penalizes the deviation of the predicted state from the MPC internal reference trajectory X_ref_mpc(k|t) is the discretized form of X_ref_mpc(t), Q mpc is the state weight matrix of the MPC.

[0152] J control_mpc (MPC control term): penalizes the control increment to ensure the smoothness of the control. R mpc is the control increment weight matrix of the MPC.

[0153] J stress_mpc (MPC stress penalty term): the core executive layer innovation of the present application. This nonlinear design realizes a "triggered" penalty. Only when the predicted stress exceeds the limit ≥1), does this term take effect and square penalize the exceeding part; within the safety zone, this term is zero and does not affect the tracking performance. This realizes the intelligent dual-mode control. Dominant weight w_penalty: this weight is set to a very high value (e.g., 10 3 ~10 5 times the elements of Q mpc , ensuring that once the risk is predicted, the stress reduction goal can unconditionally and overwhelmingly dominate the optimization process, forcing the controller to perform risk-averse operations.

[0154] b) Control target smart trade-off:

[0155] High performance tracking mode <1, safety zone): stress penalty term J stress_mpc is zero, the controller prioritizes trajectory tracking accuracy and motion smoothness;

[0156] Safety priority mode ≥1, risk zone or overrun zone): stress penalty term J increases sharply and dominates the optimization objective, the controller will automatically switch behavior mode, prioritizing stress reduction operations (such as active deceleration, adjusting attitude), until Return to safety zone.

[0157] The direct benefit of this technology is the creation of an intelligent collaborative system driven by quantitative safety indicators, with complete "feedforward-lookahead-feedback" capabilities, achieving a leap from "passive reaction" to "active adaptation".

[0158] (VI) System integration: full-link stress adaptive closed loop.

[0159] The present invention is not a simple stack of the above technologies, but rather, for the core variable of "stress", the system architecture integrates stress perception, safety index construction, stress prediction, and prediction-driven control into an organic intelligent agent.

[0160] Stress safety index ζ as the "information link" and "driving core" throughout, forms a clear closed-loop information flow:

[0161] 1. Target adaptation (3rd level planning): feedback to the planning layer, making it plan a low-essence stress planning trajectory X_plan(t).

[0162] 2. Behavior adaptation (3rd level decision): the predicted value of ζ drives the lookahead decision mechanism, modifying X_plan(t) to a safer high-level reference trajectory X_ref_high(t).

[0163] 3. Instruction adaptation (4th level optimization): X_ref_high(t) is received by the tactical optimization layer, refined into an executable MPC reference trajectory X_ref_mpc(t) containing feedforward instructions.

[0164] 4. Execution adaptation (5th level control): the bottom layer MPC takes X_ref_mpc(t) as the tracking target, and simultaneously receives real-time stress feedback ζ and prediction , dynamically solving the optimal control instructions to achieve precise, compliant, and safe final execution.

[0165] Finally, through the step-by-step refinement and stress adaptive optimization of X_plan(t)→X_ref_high(t)→X_ref_mpc(t), the entire multi-AGV system becomes a stress adaptive collaborative system capable of self-stress perception, self-safety assessment, self-behavior decision and optimization.

[0166] The method described in the application will be described in detail below in conjunction with a preferred embodiment. This embodiment takes the collaborative transportation of a large aircraft composite skin (a typical weak rigid cargo) by four AGVs in an aircraft assembly workshop as an example. The core implementation step process is clear, the technical details are clear, and it strictly follows the five-layer architecture of the application:

[0167] Step one: system initialization and digital twin deployment.

[0168] This step aims to build the hardware foundation of the system and complete the offline creation and deployment of the core software model, the cargo digital twin. For details, see Figure 2 , corresponding to the initialization of the first layer (multi-source information perception layer) and the second layer (stress state estimation layer) in the Figure 1

[0169] 1. Hardware system construction (corresponding to the first layer):

[0170] a) Deploy four AGVs with independent driving, steering, and lifting leveling platforms.

[0171] b) Install a six-axis force / torque sensor on the cargo platform of each AGV as its only cargo support and measurement unit. The lower end of the sensor is rigidly connected to the lifting leveling mechanism of the AGV, and the upper end is connected to the cargo support pad through a spherical hinge joint. This installation location is the physical interaction point (support point) between the AGV and the cargo. The sensor is a compact flange six-axis force sensor with a force (Fz) measurement accuracy of not less than 0.5%FS, torque (Mx, My) measurement accuracy of not less than 1.0%FS, and coupling error <5%FS, and outputs data through CAN bus.

[0172] c) Equip the AGV chassis with a GPS / IMU combined navigation unit and a laser radar for positioning and environment perception.

[0173] d) Install a dual-axis inclination sensor on the AGV cargo platform to obtain the platform attitude [φ, ψ].

[0174] 2. Digital twin offline creation and deployment (corresponding to the second layer):

[0175] ​a) Model building: In ANSYS software, according to the CAD model of the aircraft skin and the material certificate (such as elastic modulus, Poisson's ratio), a high-fidelity finite element model is established.

[0176] b) Data simulation: In the finite element software, thousands of groups of force and torque combinations [Fz_i, Mx_i, My_i] covering the expected working range are applied to the four support points corresponding to the physical AGV support point positions, and static analysis is performed, and the stress values σ_FEA of the preset key points such as the center and edge of the skin in each simulation are recorded.

[0177] c) Model calibration and simplification: The simulation data (input: [Fz_1, Mx_1, My_1, Fz_2, Mx_2, My_2, Fz_3, Mx_3, My_3, Fz_4, Mx_4, My_4], output: σ_FEA) is imported into the Python environment, and a three-layer feedforward neural network is trained as a mapping model using the Scikit-learn library. The predicted output of the model is the estimated stress (i.e. σ_estimated).

[0178] d) Model embedding: The trained neural network model parameter file is embedded into the AGV cluster host computer.

[0179] 3. Stress safety threshold configuration:

[0180] a) Static threshold setting: According to the design specification of the composite skin, obtain the allowable stress set σ_allowable of each key point (i.e. including σ_allowable[i]). The system configures the static safety threshold σ_safe_static[i]=0.8*σ_allowable[i] for each key point as the basic safety red line.

[0181] b) Threshold loading: When the system starts, load the above threshold and parameters. In the initial state, the final effective safety stress threshold σ_safe[i] of each key point is σ_safe_static[i].

[0182] Step two: Real-time perception and stress estimation (process within a control cycle).

[0183] This step is the periodic process of system operation like Figure 2 , which shows how the perception and cognition domains work together.

[0184] 1. Multi-source data acquisition (Layer 1 execution):

[0185] The host system synchronously collects all sensor data in each control cycle (such as 10ms):

[0186] a) Obtain AGV's current position and elevation information z_terrain in front of the path from laser SLAM system and pre-stored DTM;

[0187] b) Obtain AGV chassis pose [x, y, θ] and velocity [v, ω] (where v represents AGV linear speed, ω represents AGV angular velocity) from GPS / IMU and encoder;

[0188] c) Obtain platform attitude [φ, ψ] (where φ represents AGV cargo platform pitch angle, ψ represents AGV cargo platform roll angle) from each AGV's tilt sensor;

[0189] d) Obtain force / torque data [Fz_1, Mx_1, My_1, Fz_2, Mx_2, My_2, Fz_3, Mx_3, My_3, Fz_4, Mx_4, My_4] of support points from each AGV's six-axis force / torque sensor through CAN bus.

[0190] 2. Stress real-time solving (Layer 2 executes):

[0191] Input the collected force / torque data into the neural network model embedded in Step 1. The model completes forward propagation calculation within milliseconds, outputting the stress estimation value set σ_estimated of the key points of the goods at the current time.

[0192] 3. Stress safety index calculation (Layer 2 executes):

[0193] The system then calculates the unified "stress safety index ζ" in real time according to each σ_estimated[i] in σ_estimated and σ_safe[i]: ζ = max(σ_estimated[i] / σ_safe[i]). This ζ value is immediately sent to the subsequent decision and control layer as the core decision variable in the current control period.

[0194] Step 3: Stress adaptive decision and planning.

[0195] This step corresponds to the work of the decision and planning domain, Figure 3 It shows how the system makes intelligent decisions based on ζ.

[0196] 1. High-dimensional trajectory generation (Layer 3 - strategic decision):

[0197] The planner calls the nonlinear optimization solver in the six-dimensional state space of [x, y, θ, z, φ, ψ] to generate a preliminary space-time trajectory X_plan(t) with efficiency, smoothness, and ζ as the objective function. The core feature is to embed the stress safety target in the optimization function of the planning, and the objective function explicitly introduces the stress safety item, which forces the planner to prioritize stress optimization when coordinating all degrees of freedom, thereby automatically generating a high-dimensional space-time trajectory with inherently low stress.

[0198] 2. Prospective decision and re-planning (Level 3 - strategic decision):

[0199] The planning layer calls the stress prediction mechanism to perform safety assessment on the trajectory X_plan(t) and obtain the future stress prediction curve.

[0200] Scenario A (stress safety): If the prediction is always below the safety threshold 1 (i.e.,: <1), then the trajectory X_plan(t) is output as a feasible solution. At this time, the high-level reference trajectory X_ref_high(t) = X_plan(t).

[0201] Scenario B (stress overrun warning): If the prediction model shows that at t=3s, due to the AGV's imminent sharp turn, the stress will not be below 1 (i.e.,: ≥1). Then the decision mechanism immediately marks t=3s as a "virtual stop line" and triggers trajectory re-planning. The re-planner generates a new trajectory that starts to smoothly decelerate and turn with a more gentle curvature before t=3s as the high-level reference trajectory X_ref_high(t), to ensure that the stress is always safe.

[0202] 3. Terrain feedforward compensation instruction generation (Level 3 - strategic decision):

[0203] While planning the trajectory, according to the time-space mapping relationship between the DTM preview information and the planned trajectory, the platform height feedforward instruction z_platform(t) required to maintain the level of the goods is calculated in reverse, and the timing of the instruction is ensured to be sent t_forward earlier than the AGV chassis reaches the corresponding terrain point, to achieve active elimination of static stress.

[0204] Step four: trajectory optimization and closed-loop control execution.

[0205] This step corresponds to the functions of Level 4 (tactical decision) and Level 5 (execution) in the architecture, which refers to the final landing of instructions.

[0206] 1. Rolling optimization (Level 4 - tactical decision):

[0207] Trajectory optimization layer flow as shown in Figure 4 receives the trajectory X_ref_high(t) from planning layer and the feedforward command z_platform(t). It performs a rolling optimization in a shorter prediction horizon (1.0~3.0s) with accurate AGV and platform dynamics model to generate a more accurate reference trajectory X_ref_mpc(t) containing feedforward control amount to further eliminate the effects of model mismatch and un-modeled dynamics.

[0208] 2. Model predictive control (MPC) execution (Layer 5 - Execution):

[0209] The bottom layer MPC controller receives X_ref_mpc(t) as tracking target. The flow as shown in Figure 5 , at each control cycle (e.g. 2ms), MPC solves an optimal control problem in finite horizon with real-time state feedback of AGV (including the latest stress safety indicator ζ).

[0210] a) Candidate control sequence evaluation: MPC invokes stress prediction mechanism for each candidate control sequence in its internal, to calculate its corresponding future stress safety indicator sequence .

[0211] b) Optimization solving: solve the optimization problem with objective function containing a prediction based trigger penalty term (i.e. MPC stress penalty term J stress_mpc ).

[0212] When <1 (safe zone): this penalty term is zero, the controller prioritizes trajectory tracking accuracy and motion smoothness;

[0213] When ≥1 (risk zone): this penalty term sharply increases, dominates the optimization objective, forces the controller to automatically switch behavior, prioritizes AGV motion adjustment (e.g. active deceleration, adjust attitude) to reduce stress until it returns to the safe zone.

[0214] c) Command issuance: issue the first optimal control amount (drive motor speed / torque command, leveling actuator speed command) calculated to each AGV actuator.

[0215] 3. Full-link adaptive closed loop:

[0216] During the whole moving process, the stress safety index ζ and its predicted value as the core feedback continuously act on the whole process from the third layer to the fifth layer. This design based on the penalty term and model prediction makes the response of the system to stress disturbance smooth, gradual and adaptive. When the actual stress deviates from the expected due to uneven road surface, model mismatch and other reasons, the MPC can adjust the control command in real time and gently, dynamically correct the motion state of the AGV, and actively pull the stress back to the safety zone, rather than triggering a sharp emergency stop after serious over-limit. This forward and flexible control characteristic ensures that the stress of the goods is firmly controlled within the safety range, which is crucial for preventing dynamic damage to weak rigid goods.

[0217] The present specific embodiment fully demonstrates the whole process of the present application from system initialization, real-time perception, intelligent decision-making to precise execution through four clear steps, and fully embodies how the five-layer architecture and the three major functional domains work together to ultimately achieve the core goal of safe moving of weak rigid goods.

[0218] Compared with the prior art, the present application has made the following breakthroughs through the proposed five core innovations:

[0219] 1. A unified safety index is invented, laying the foundation for intelligent control:

[0220] By constructing the stress safety index, the present application converts the abstract and complex mechanical state of the goods into a clear, measurable and controllable system-level variable. This solves the fundamental problem of "how to achieve the mechanical safety goal" in multi-AGV collaborative control and is the core prerequisite for realizing intelligent closed-loop control.

[0221] 2. Paradigm leap from traditional collaboration to mechanical safety collaboration:

[0222] Through the five-layer intelligent control architecture and the digital twin stress estimation technology, the internal stress of the goods is first taken as a controllable state variable and runs through the whole control process, realizing a fundamental change from traditional geometric collaboration to essential mechanical safety collaboration, and fundamentally establishing the core position of mechanical safety in collaborative control.

[0223] 3. Sensing capability leap from "real-time perception" to "predictive perception":

[0224] The unique stress prediction mechanism constructs the time series prediction sequence of the stress safety index, enabling the system to have the perception ability of the future mechanical state. This realizes a fundamental leap from real-time perception that can only perceive the "current physical state" to predictive perception that can foresee the "future state evolution". This capability provides key forward-looking information for the core decision-making, and moves the intervention point of the control system from "after stress over-limit" to "before stress over-limit", laying an irreplaceable perception foundation for achieving true active immunity.

[0225] 4. Equipped with complete active safety capability of "feedforward-lookahead-feedback":

[0226] The terrain active feedforward compensation mechanism prevents the static stress mismatch by synchronizing the platform and chassis actions in space-time; the lookahead decision mechanism based on stress model prediction plans and executes a smooth jerk trajectory before the dynamic impact occurs, thus physically avoiding the generation of impact force and solving the dynamic impact suppression problem; the MPC adaptive closed loop can adjust the control command in real time and gently, and correct the AGV motion state when the actual stress deviates from the expected due to uneven road surface, model mismatch, etc. Through terrain feedforward compensation, model prediction lookahead decision and MPC adaptive closed loop, a complete active safety protection system is built to prevent, suppress and correct stress over-limit risks from the root, trend and instant.

[0227] 5. Improved global intelligent decision-making capability of the system based on model prediction:

[0228] The high-dimensional space-time joint planning embedded in stress enables the system to synchronize and coordinate all degrees of freedom in a unified framework to seek the global optimal solution, and the digital twin gives the system the ability to "understand" the mechanical response of the goods. Together, they form the "intelligent decision-making core" of the system, which, combined with the full-link stress adaptive closed loop, enables the system to make optimal decisions that balance safety, stability and efficiency in complex working conditions such as uneven road surfaces and sudden disturbances, demonstrating excellent environmental adaptability and system robustness.

[0229] 6. Laid the foundation for a general architecture for widespread application:

[0230] The decoupled five-layer architecture and the core methodology of stress control enable the invention to be quickly expanded and applied to various high-value weak-rigid goods handling scenarios in the fields of aviation, aerospace, precision manufacturing, etc. by adjusting the digital twin model and optimization target weight.

[0231] Although the invention has been described in detail above with general description and specific embodiments, some modifications or improvements can be made on the basis of the invention, which is obvious to those skilled in the art. Therefore, these modifications or improvements made on the basis of not deviating from the spirit of the invention are within the scope of the invention claimed.

Claims

1. A multi-AGV stress self-adaptive cooperative control method for weakly rigid cargo handling, characterized in that, Each AGV is provided with a six-dimensional force sensor at a position for supporting a cargo, the six-dimensional force sensor being configured to acquire force and torque data of a supporting point of the cargo, and the control method comprises the following steps: In step 1, the chassis attitude information, platform attitude information, speed information, digital terrain model information in front of a path, and data of the six-dimensional force sensor of each AGV are acquired in real time when the multiple AGVs cooperatively carry the cargo; the data of each six-dimensional force sensor is input into a trained neural network model to obtain a stress set of the cargo, the stress set comprising stress values at each set position of the cargo; and one of results obtained by dividing the stress value at each set position by a preset stress safety threshold value at the set position is selected as a stress safety index of the stress set according to a set selection rule. The training of the neural network model comprises: using data in a stress mapping database comprising a corresponding relationship between force and torque data of each supporting point of the cargo and stress values at each set position of the cargo as a data set in a training process, and training the neural network model, so that the trained neural network model can obtain the stress values at each set position of the cargo according to the force and torque data of each supporting point of the cargo. In step 2, a preliminary space-time trajectory is obtained by using a nonlinear optimization solver according to the information acquired in real time in step 1 and the stress safety index, an optimization objective function of the nonlinear optimization solver comprising a stress safety item of the stress safety index; a stress safety index sequence in a future period of time is predicted according to the preliminary space-time trajectory, and if each stress safety index in the stress safety index sequence is less than a set value, the preliminary space-time trajectory is taken as a high-level reference trajectory, otherwise, a trajectory is re-planned, and the re-planned trajectory is taken as the high-level reference trajectory. In step 3, the high-level reference trajectory is optimized by using a dynamics model to obtain an MPC reference trajectory, and an MPC controller takes the MPC reference trajectory as a tracking target and dynamically solves an optimal control amount according to a real-time state feedback of the AGV, and the optimal control amount is output to an AGV actuator.

2. The stress self-adaptive cooperative control method for multiple AGVs according to claim 1, characterized in that, In step 1, the establishment of the stress mapping database comprises: establishing a finite element model of the cargo, applying a force and torque combination covering an expected working range to positions in the finite element model corresponding to actual positions of each supporting point of the cargo in a finite element software, performing statics analysis to obtain stress values at each set position of the cargo corresponding to different force and torque combinations, forming a corresponding relationship between multiple force and torque combinations and the stress values at each set position of the cargo corresponding to the force and torque combinations, and obtaining the stress mapping database.

3. The stress self-adaptive cooperative control method for multiple AGVs according to claim 1, characterized in that, In step 1, the set selection rule is that one of results obtained by dividing the stress value at each set position by a preset stress safety threshold value at the set position is selected as the stress safety index of the stress set according to a set selection rule. The determination of the preset stress safety threshold value at the set position comprises: acquiring an allowable stress value at the set position of the cargo according to a design specification of the cargo, and setting the preset stress safety threshold value at the set position to be less than the allowable stress value at the set position of the cargo.

4. The stress self-adaptive cooperative control method for multiple AGVs according to claim 1, characterized in that, In step 2, the stress safety item has the highest priority compared to other items in the optimization objective function.

5. The stress self-adaptive cooperative control method for multiple AGVs according to claim 1, characterized in that, In step 2, predicting the stress safety index sequence in the future period of time comprises: predicting the platform height and attitude information in the future period of time according to the preliminary space-time trajectory, and estimating the forces and moments required for maintaining the static balance of the goods by solving the force and moment balance equations based on the known mass, center of gravity position and geometric size of the goods, inputting the forces and moments into the trained neural network model to obtain the stress safety index sequence in the future period of time.

6. The stress self-adaptive cooperative control method for multiple AGVs according to claim 1, characterized in that, In step 2, the trajectory re-planning comprises: taking the current state as the starting point, taking the time after the set time interval as the virtual stop line, and generating a smooth deceleration trajectory to the virtual stop line with the virtual stop line as the constraint.

7. The stress self-adaptive cooperative control method for multiple AGVs according to claim 1, characterized in that, In step 3, the dynamic solution of the optimal control quantity comprises: the MPC controller predicts the stress safety index sequence in the future period of time for each candidate control sequence in the MPC controller, the cost function of the MPC controller includes a trigger penalty term based on the predicted stress safety index, and the optimal control quantity is obtained by using the MPC controller including the cost function of the MPC controller.

8. The stress self-adaptive cooperative control method for multiple AGVs according to claim 7, characterized in that, In step 3, the trigger penalty term based on the predicted stress safety index is: when all the predicted stress safety indexes are less than 1, the trigger penalty term is zero, and when there is a predicted stress safety index not less than 1, the trigger penalty term is much larger than other items in the cost function of the MPC controller.

9. The stress self-adaptive cooperative control method for multiple AGVs according to claim 1, characterized in that, In step 2, while planning the trajectory, the platform height feedforward instruction required to keep the goods horizontal is calculated in reverse according to the preview information of the digital terrain model and the time and space mapping relationship of the planned trajectory, and the timing of the platform height feedforward instruction is ahead of the forward preview time calculated dynamically when the AGV chassis reaches the corresponding terrain point; the dynamically calculated forward preview time is the sum of the inherent response delay of the platform actuator, the travel time required for the platform to move from the current position to the target height, and the inherent delay of system data processing and communication.

10. The stress self-adaptive cooperative control method for multiple AGVs according to claim 9, characterized in that, In step 3, while obtaining the MPC reference trajectory by optimization, the feedforward control quantity for tracking the MPC reference trajectory is also solved, and the feedforward control quantity is output as part of the MPC reference trajectory attributes.

Images