Dynamic risk perception prediction stable control system and method for cold-chain logistics transport vehicle

Abstract

The invention provides a dynamic risk perception prediction stability control system and method for a cold-chain logistics transport vehicle, and belongs to the technical field of vehicle control. Comprising three core modules: a physical-data mixed residual state predictor, a quantitative stability evaluation module based on a maximum Lyapunov index, and a self-adaptive nonlinear model prediction controller fusing predetermined performance control. And finally, the control instruction is distributed to the four in-wheel motors through a multi-target torque distributor. According to the method, the prediction precision is remarkably improved, continuous risk quantification is realized, and meanwhile, the control performance and the system robustness are synchronously improved.

Description

Technical Field

[0001] This invention provides a dynamic risk perception, prediction, and stability control system and method for cold chain logistics transport vehicles, belonging to the field of vehicle control technology. Background Technology

[0002] In assessing the stability of cold chain logistics transport vehicles, the phase plane method (such as...) Classical methods such as phase plane have fundamental limitations: they rely too heavily on static thresholds and can usually only provide a binary judgment of stability or instability, failing to quantify the continuity of risk.

[0003] Existing technology one, risk-prediction-based integrated control of active front-wheel steering and direct yaw moment (H. Liu, C. Liu, L. Han, and C. Xiang, “Handling and stability integrated control of AFS and dyc for distributed drive electric vehicles based on risk assessment and prediction,” IEEE transactions on intelligent transportation systems, vol.23, no. 12, pp. 23148–23163, 2022.), proposes an integrated control framework aimed at coordinating the AFS and DYC subsystems through risk assessment and prediction to improve the handling stability of distributed drive electric vehicles under extreme conditions. This technology has the following drawbacks:

[0004] (1) Discrete risk quantification and lack of gradient perception: Its risk assessment still relies on the "inside / outside" judgment of the phase plane boundary, which is essentially a discrete binary (stable / unstable) logic. It cannot generate a continuously changing risk quantification index (such as a continuous risk factor between 0 and 1), so the controller cannot make smooth and adaptive strategy adjustments according to the subtle and continuous evolution of risk. This is particularly disadvantageous for cold chain logistics transport vehicles with high center of gravity and large inertia, because it is impossible to trigger gradual intervention in the early and low-level stages of instability risk, which may miss the best intervention time, resulting in the need to apply more drastic control to correct the posture and increase the risk of cargo shaking or even overturning. In addition, the proposal of fuzzy rules has serious empiricism and is difficult to cover all working conditions.

[0005] (2) The accuracy of state prediction methods is limited under extreme dynamic conditions: When dealing with highly nonlinear and abrupt extreme driving conditions (such as emergency obstacle avoidance on low-adhesion surfaces), the prediction errors tend to accumulate, which may lead to inaccurate or untimely predictions of future instability moments, affecting the optimal intervention time of forward control. Inaccurate predictions may prevent the system from predicting the instability risk caused by changes in vehicle state in advance, thus failing to effectively prevent cargo damage.

[0006] (3) The system lacks robustness guarantees to cope with controller failure: When the SMPC optimization solver fails under extremely complex or unmodeled conditions, the framework does not have a lightweight backup control law designed, which may lead to the interruption of control commands and the loss of stability guarantee at critical moments. For cold chain logistics transport vehicles, even a brief control interruption under high-speed or curve conditions may quickly cause vehicle instability, which not only endangers road traffic safety, but may also lead to the complete destruction of high-value temperature-sensitive goods, causing significant economic losses.

[0007] Existing technology two, a predictive vehicle stability assessment method based on three-dimensional Lyapunov exponents (R. Lian, Z. Li, W. Li, J. Ge, and L. Li, “Predictive vehicle stability assessment using lyapunov exponent under extreme conditions,” IEEE Transactions on Intelligent Transportation Systems, 2024.), proposes a framework for vehicle stability prediction and assessment that combines Physical Information Neural Networks (PINN) and three-dimensional Lyapunov exponents (3D-LEs), aiming to assess vehicle stability risks in advance under extreme conditions. However, it has the following drawbacks:

[0008] (1) Prediction module relies on high-precision model: In terms of calculating MLE, although the traditional model-based Jacobi matrix method can calculate the complete Lyapunov exponent, this method is highly dependent on the accuracy of the model. Low DOF models cannot reflect the real vehicle state under extreme conditions, resulting in poor time-varying performance; while high DOF models can generate more accurate Lyapunov exponents, the calculation process is complex and difficult to meet the real-time requirements of the control system.

[0009] (2) Lack of continuous risk quantification and adaptive control mechanism: The output of 3D-LEs is the stability judgment (positive / negative) of each dimension. Although it can identify unstable states, it cannot generate a continuous and smoothly adjustable risk index (such as a risk factor between 0 and 1). Therefore, the controller is unable to adaptively adjust the weight according to the continuous changes in risk, and cannot achieve a smooth switch of "focusing on tracking when the risk is low and focusing on stability when the risk is high". The control action may change abruptly, exacerbating cargo shaking.

[0010] (3) Insufficient robustness in extreme conditions: When PINN predictions have large errors or 3D-LEs calculations fail, the system lacks a lightweight backup control strategy, which may lead to control interruption. For cold chain logistics vehicles, even a brief control interruption during high-speed driving may cause vehicle instability and cargo damage, and the overall reliability of the system needs to be improved. Summary of the Invention

[0011] To address the aforementioned technical problems, this invention provides a dynamic risk perception, prediction, and stability control system and method for cold chain logistics transport vehicles, to solve the following technical issues:

[0012] (1) Solve the problem of response delay or even loss of control caused by traditional passive control intervening only after the vehicle reaches or exceeds the stability limit, and avoid cargo damage and traffic accidents caused by this.

[0013] (2) Overcome the problem that existing prediction methods are difficult to achieve high-precision and physically consistent predictions under extreme conditions of sparse data and high uncertainty.

[0014] (3) Solve the problem that traditional geometric stability assessment methods (such as phase plane method) rely too much on static thresholds, can only give binary judgments, and cannot quantify the degree of risk continuity, and realize the gradient assessment of the instability risk of cold chain logistics transport vehicles.

[0015] (4) Solve the problems of control robustness, continuous risk perception and adaptation and solver failure response mechanism of the existing model predictive control (MPC) framework under extreme working conditions of cold chain logistics transport vehicles, and improve the overall robustness and reliability of the system.

[0016] This invention proposes a dynamic risk perception, prediction, and stability control system and method for cold chain logistics transport vehicles, specifically for dynamic risk perception, prediction, and stability control (DR-PSC) of four-wheel independently driven electric cold chain logistics transport vehicles. This system integrates state prediction, stability risk assessment, and adaptive control, and is particularly designed to proactively avoid impending instability, taking into account the high center of gravity and large cargo variations characteristic of cold chain logistics transport vehicles.

[0017] The specific technical solution is as follows:

[0018] The dynamic risk perception, prediction, and stability control system for cold chain logistics transport vehicles includes three core modules: a physical-data hybrid residual state predictor, a quantitative stability assessment module based on the maximum Lyapunov exponent, and an adaptive nonlinear model predictive controller that integrates predetermined performance control; finally, a multi-objective torque distributor distributes control commands to the four in-wheel motors.

[0019] Furthermore, the aforementioned physical-data hybrid residual state predictor, used to predict future vehicle states, employs a hybrid residual framework:

[0020] The final predicted state sequence is composed of the physical baseline prediction and the data-driven residual prediction:

[0021]

[0022] In the formula, For the final predicted state sequence, For physical baseline prediction, For data-driven residual prediction.

[0023] Physical baseline prediction Based on a simplified discrete kinematics model;

[0024] yaw rate Baseline predictions are propagated using a linear 2-DOF reference model:

[0025]

[0026] In the formula, For baseline prediction of yaw rate at time k in the future, Let yaw rate be the current angular velocity. The derivative of the current yaw rate. For the response coefficient, For time step, For the baseline prediction of the front wheel steering angle at time k in the future, This represents the front wheel steering angle at the current moment.

[0027] Data-driven residual prediction Prediction is performed using a pre-trained tab-based model, TabPFN; at each time step... The historical state residual sequence As contextual input, TabPFN leverages its zero-shot context learning capability to output the probability distribution of future residuals, taking the median as the deterministic residual prediction. .

[0028] Furthermore, the quantitative stability assessment module based on the maximum Lyapunov exponent is used to quantify the stability risk of future predicted sequences.

[0029] The two initial distances are The distance between adjacent trajectories that evolves over time Approximately:

[0030]

[0031] In the formula, The Lyapunov index, For time. By taking the natural logarithm of both sides of equation (3), The definition is expressed in the following form:

[0032]

[0033] MLE computation directly analyzes the chaotic characteristics of state sequences. For a given predicted state time series, in phase space, for a reference point... Search for it nearest neighbors Calculate after a fixed evolution time Distance divergence after; local Lyapunov index Estimate using the following formula:

[0034]

[0035] The MLE sequence is obtained by traversing the prediction sequence. ;

[0036] Time-weighted instability risk (TWIR) factor generation: By weighting over time, it focuses more on risks in the near future, which meets the need for early intervention in cold chain logistics transportation vehicles to prevent instability;

[0037] (1) Define the time-weighted vector:

[0038]

[0039] in It is the attenuation factor;

[0040] (2) Extracting the positive risk sequence:

[0041]

[0042] (3) Calculate the original TWIR factor:

[0043]

[0044] TWIR factor As a continuous risk indicator input to the controller.

[0045] Furthermore, the pre-stabilized adaptive predetermined performance nonlinear model predictive controller integrates risk adaptation, predetermined performance constraints, and a backoff mechanism.

[0046] Vehicle control model: A 2-DOF bicycle model is used as the internal prediction model; state vector With control input vector Defined as:

[0047]

[0048] In the formula, The sideslip angle is the angle of the centroid. The yaw rate is angular velocity. Add front wheel steering angle to the output. To add a yaw moment to the output. The continuous-time state-space equation is as follows:

[0049]

[0050] in The driver's steering input is considered a measurable disturbance; assuming a small sideslip angle and linear tire characteristics in the prediction time domain, the dynamics are described by the following equation:

[0051]

[0052] in For vehicle quality, For longitudinal velocity, This is the yaw moment of inertia; and These are the distances from the center of gravity to the front and rear axles, respectively. and Indicates the steering stiffness of the tire;

[0053] Constraints: To ensure the physical feasibility of the control action, the actuator must satisfy the saturation constraint condition:

[0054]

[0055] Risk perception adaptive weight scheduling: When the risk is low, the controller focuses on tracking the driver's intentions; when the risk is high, it decisively strengthens stability control, smoothly adapting to the needs of different driving stages of cold chain logistics transport vehicles; TWIR factor By mapping with the Sigmoid function, the tracking error weight matrix in the optimization problem is dynamically adjusted. and control weight matrix ,in and Defined by equations (13) and (14);

[0056]

[0057] In the formula, , These are the upper and lower bounds of the tracking error weights, respectively. , These are the upper and lower bounds of the control weights, respectively. The center value of the TWIR factor sequence, , These are the sensitivity coefficients.

[0058] Pre-defined performance control for PPC conversion and embedding:

[0059] (1) Reference generation and constraints: ideal reference value and The calculations are performed using the linear 2-DOF model, i.e., Equations 17 and 18, and are constrained by the road surface friction coefficient. Obtain the final reference value and That is, Equations 19 and 20:

[0060]

[0061] (2) Performance boundary and error transformation: defining tracking error That is, Equation 21; Design the exponentially decaying performance function. That is, Equation 23, requires the error to satisfy That is, Equation 22;

[0062]

[0063] In the formula, and These are the tracking errors of the sideslip angle and yaw rate, respectively. This is the initial value for the performance function. This represents the steady-state value of the performance function. The convergence rate is denoted by a strictly increasing logarithmic transformation function. Normalization error That is, Equation 24, which maps to an unbounded transformation state. :

[0064]

[0065] Its inverse transformation is:

[0066]

[0067] Therefore, the original state constraint problem is transformed into a problem of... Unconstrained positive definite problems;

[0068] (3) Adaptive PPC-NMPC optimization problem: at each control time step Solve the following finite-time optimization problem:

[0069]

[0070] in It is the predicted value of the transformed state. It is the control input sequence to be optimized. and It is an adaptive weight matrix;

[0071] (4) PPC-NMPC prediction model and constraints: Since vehicle dynamics are determined by physical states The description is as shown in Equation 10), while NMPC optimizes the state transition. Therefore, it is necessary to establish a bridge between the two in the prediction process; the "inverse prediction-forward recursive update" strategy is adopted: in each prediction step, the current transformed state is first mapped back to the physical state through inverse transformation, and then the state derivative is calculated using the vehicle dynamics model. Then, the derivative of the state transformation is derived using the chain rule. :

[0072]

[0073] The future state transformation is propagated forward through Euler integral; at the same time, the control input must satisfy the physical saturation constraint of the actuator, as shown in equation (12);

[0074] Lightweight backoff controller: This mechanism ensures basic stability for cold chain logistics transport vehicles, even under extreme or unmodeled conditions, preventing loss of control; it is activated when the NMPC solver cannot find a feasible solution; this controller uses a high-gain proportional control law to directly apply to the changing state. :

[0075]

[0076] In the formula, and These represent the controller's revenue.

[0077] Stability analysis of the lightweight rollback controller:

[0078] Consider the quadratic Lyapunov candidate function used for state transformation as Equation (30), and its time derivative as Equation (31).

[0079]

[0080] By selecting negative fixed control gain The controller forms a negative feedback loop; if ,but This leads to negative consequences. and , making ,therefore The system state is driven back to the origin. ;

[0081] By proof The convergence of this property can theoretically guarantee the safety of the vehicle system in rollback mode;

[0082] The final execution logic adopts a hybrid switching strategy:

[0083] .

[0084] The dynamic risk perception, prediction, and stability control method for cold chain logistics transport vehicles of the present invention employs the aforementioned dynamic risk perception, prediction, and stability control system for cold chain logistics transport vehicles; the method includes the following steps:

[0085] (1) State prediction: At each control moment, the PDHR state predictor receives the current and historical vehicle states and outputs the future state. A precise state prediction sequence for each step;

[0086] (2) Risk assessment: The MLE-QSA module receives the predicted state sequence, calculates its maximum Lyapunov exponent MLE sequence, and aggregates it into a continuous time-weighted instability risk (TWIR) factor; this factor can continuously quantify the risk level of future instability phenomena such as tilting or skidding of cold chain logistics transport vehicles, rather than a simple binary judgment.

[0087] (3) Adaptive control decision: The PS-PMPC controller receives the current vehicle state and TWIR factor; first, it dynamically adjusts the weight matrix of its internal optimization problem according to the TWIR factor to achieve risk adaptation; second, it embeds the predetermined performance control PPC into the optimization problem, transforms the hard constraint of state tracking error into a positive definite problem for unconstrained state transformation, and solves it to obtain the optimal front wheel steering angle correction. and additional yaw moment The control objective is to minimize cargo swaying by controlling the vehicle as smoothly as possible while ensuring vehicle stability. If the NMPC solver fails, a lightweight backoff controller is activated to ensure system feasibility.

[0088] (4) Torque distribution and execution: The multi-target torque distributor receives the required additional yaw moment. Under the premise of satisfying the physical constraints of each motor, the independent drive / braking torque of the motors in the four wheels is calculated and executed by optimizing the allocation strategy; the allocation strategy can take into account both tire load uniformity and energy efficiency, and adapt to the long-distance operation needs of cold chain logistics transport vehicles.

[0089] The technical effects of the present invention are as follows:

[0090] (1) Significantly improved prediction accuracy: The proposed PDHR predictor performs excellently under extreme dynamic conditions (such as the fishhook test), with prediction accuracy significantly higher than that of pure physical models or pure data-driven models. It can more accurately predict the future attitude of cold chain logistics transport vehicles under changes in cargo loading status, thus providing reliable prior information for the control system. Based on this, the PS-PMPC controller can dynamically adjust the control strategy according to the TWIR factor to achieve early proactive intervention—taking countermeasures about 0.2-0.3 seconds before vehicle instability occurs, changing passive response to proactive prevention, and significantly reducing the risk of cargo overturning or damage caused by sudden instability. The combination of the two constitutes a complete closed-loop solution from high-precision prediction to forward-looking control.

[0091] (2) Achieve continuous risk quantification: The MLE-QSA module can output continuous TWIR factors, transforming stability assessment from discrete binary judgments into continuous risk gradients, enabling the system to perceive the accumulation process of instability risk of cold chain logistics transport vehicles earlier and more delicately.

[0092] (3) Simultaneous improvement of control performance and system robustness: The proposed PS-PMPC controller not only effectively maintains vehicle stability, but also significantly reduces the violent shaking of the vehicle body and cargo by generating smooth control actions, thereby improving the quality of cargo transportation; at the same time, the system ensures that the tracking error is always within the preset safety boundary through PPC hard constraints, and is equipped with a lightweight backoff mechanism, which can maintain the basic stability of the system in the event of unexpected failure of the main controller. This design comprehensively enhances the control performance and safety assurance of the cold chain transportation system in complex environments. Attached Figure Description

[0093] Figure 1 This is the framework for the dynamic risk perception, prediction, and stability control model of the present invention.

[0094] Among them, (a) a physical-data hybrid residual state predictor; (b) a quantitative stability assessment based on MLE; and (c) pre-stabilized adaptive NMPC specified performance. Detailed Implementation

[0095] The overall framework of the dynamic risk perception, prediction, and stability control system for cold chain logistics transport vehicles is as follows: Figure 1As shown, it mainly includes three core modules: a physical-data hybrid residual (PDHR) state predictor, a quantitative stability assessment (MLE-QSA) module based on the maximum Lyapunov exponent (MLE), and an adaptive nonlinear model predictive controller that integrates predetermined performance control. Finally, a multi-objective torque distributor distributes control commands to the four in-wheel motors.

[0096] 2.2.2 Core Module Technical Solution

[0097] 1. Physical-Data Hybrid Residual (PDHR) State Predictor

[0098] This predictor is designed to predict future vehicle states, and its structure is as follows: Figure 1 As shown in (a), a hybrid residual framework is used:

[0099] The final predicted state sequence is composed of the physical baseline prediction and the data-driven residual prediction:

[0100]

[0101] In the formula, For the final predicted state sequence, For physical baseline prediction, For data-driven residual prediction.

[0102] Physical baseline prediction Based on a simplified discrete kinematics model. For example, yaw rate. Baseline predictions are propagated using a linear 2-DOF reference model:

[0103]

[0104] In the formula, For baseline prediction of yaw rate at time k in the future, Let yaw rate be the current angular velocity. The derivative of the current yaw rate. For the response coefficient, For time step, For the baseline prediction of the front wheel steering angle at time k in the future, This represents the front wheel steering angle at the current moment.

[0105] Data-driven residual prediction Prediction is performed using a pre-trained tab-based model, TabPFN. At each time step... The historical state residual sequence As contextual input, TabPFN leverages its zero-shot context learning capability to output the probability distribution of future residuals, taking the median as the deterministic residual prediction. .

[0106] 2. Quantitative Stability Assessment Module Based on Maximum Lyapunov Index (MLE) (MLE-QSA)

[0107] This module is used to quantify the stability risk of future predicted sequences, and the process is as follows: Figure 1 As shown in (b).

[0108] The Lyapunov exponent is a key indicator characterizing the average divergence or convergence rate of a system in phase space with respect to neighboring orbits. For nonlinear dynamic systems, the sign of the Lyapunov exponent represents its stability: a negative value indicates that the system is stable and neighboring orbits converge exponentially to a stable fixed point; a positive value indicates that the system is unstable and neighboring orbits diverge exponentially

[31] . Consider two initial distances of... The distance between adjacent trajectories that evolves over time Approximately:

[0109]

[0110] In the formula, The Lyapunov index, For time. By taking the natural logarithm of both sides of equation (3), The definition is expressed in the following form:

[0111]

[0112] MLE calculation (based on an improved Wolf algorithm): This method directly analyzes the chaotic characteristics of state sequences and has better evaluation capabilities for systems with significant nonlinearity, such as cold chain logistics transport vehicles. For a given predicted state time series, in phase space, for a reference point... Search for it nearest neighbors Calculate after a fixed evolution time The distance diverges afterward. Local Lyapunov exponent. Estimate using the following formula:

[0113]

[0114] The MLE sequence is obtained by traversing the prediction sequence. .

[0115] Time-Weighted Instability Risk (TWIR) Factor Generation: By weighting over time, it focuses more on risks in the near future, which meets the need for early intervention in cold chain logistics transportation vehicles to prevent instability.

[0116] (1) Define the time-weighted vector:

[0117]

[0118] in This is the attenuation factor.

[0119] (2) Extracting the positive risk sequence:

[0120]

[0121] (3) Calculate the original TWIR factor:

[0122]

[0123] TWIR factor As a continuous risk indicator input to the controller.

[0124] 3. Pre-stabilized adaptive predetermined performance nonlinear model predictive controller (PS-PMPC)

[0125] This controller is the core of the system's execution, and its structure is as follows: Figure 1 As shown in (c), it integrates risk adaptation, predetermined performance constraints, and a fallback mechanism.

[0126] Vehicle control model: A 2-DOF bicycle model is used as the internal prediction model. State vector. With control input vector Defined as:

[0127]

[0128] In the formula, The sideslip angle is the angle of the centroid. The yaw rate is angular velocity. Add front wheel steering angle to the output. To add a yaw moment to the output. The continuous-time state-space equation is as follows:

[0129]

[0130] in The driver's steering input is considered a measurable disturbance. Assuming a small slip angle and linear tire characteristics in the prediction time domain, the dynamics are described by the following equation:

[0131]

[0132] in For vehicle quality, For longitudinal velocity, This is the yaw moment of inertia. and These are the distances from the center of gravity to the front and rear axles, respectively. and This indicates the steering stiffness of the tire.

[0133] Constraints: To ensure the physical feasibility of the control action, the actuator must satisfy the saturation constraint condition:

[0134]

[0135] Risk-aware adaptive weighted scheduling: When the risk is low, the controller focuses on tracking the driver's intentions; when the risk is high, it decisively strengthens stability control, smoothly adapting to the needs of different driving stages of cold chain logistics transport vehicles. TWIR factor By mapping with the Sigmoid function, the tracking error weight matrix in the optimization problem is dynamically adjusted. and control weight matrix ,in and Defined by equations (13) and (14).

[0136]

[0137] In the formula, , These are the upper and lower bounds of the tracking error weights, respectively. , These are the upper and lower bounds of the control weights, respectively. The center value of the TWIR factor sequence, , These are the sensitivity coefficients.

[0138] Predefined Performance Control (PPC) Conversion and Embedding

[0139] (1) Reference generation and constraints: ideal reference value and Calculated using a linear 2-DOF model (Equations 17 and 18), and constrained by the road surface friction coefficient. Obtain the final reference value and (Equations 19 and 20) This constraint ensures that the reference value is always within a physically accessible safe range, setting a safe tracking target for cold chain logistics transport vehicles.

[0140]

[0141] (2) Performance boundary and error transformation: defining tracking error (Equation 21). Design the exponentially decaying performance function. (Equation 23) requires the error to satisfy (Equation 22).

[0142]

[0143] In the formula, and These are the tracking errors of the sideslip angle and yaw rate, respectively. This is the initial value for the performance function. This represents the steady-state value of the performance function. The convergence rate is denoted by a strictly increasing logarithmic transformation function. Normalization error (Equation 24) is mapped to an unbounded transformation state. :

[0144]

[0145] Its inverse transformation is:

[0146]

[0147] Therefore, the original state constraint problem is transformed into a problem of... The unconstrained positive definite problem.

[0148] (4) Adaptive PPC-NMPC optimization problem: at each control time step Solve the following finite-time optimization problem:

[0149]

[0150] in It is the predicted value of the transformed state. It is the control input sequence to be optimized. and It is an adaptive weight matrix.

[0151] (4) PPC-NMPC prediction model and constraints: Since vehicle dynamics are determined by physical states The description (Equation 10) is provided, while NMPC optimizes the state transition. Therefore, it is necessary to establish a bridge between the two during the prediction process. This paper adopts an "inverse prediction-forward recursive update" strategy: in each prediction step, the current transformed state is first mapped back to the physical state through inverse transformation, and then the state derivative is calculated using the vehicle dynamics model. Then, the derivative of the state transformation is derived using the chain rule. :

[0152]

[0153] The future state transformation is propagated forward via Euler integral. Meanwhile, the control input must satisfy the physical saturation constraint of the actuator, as shown in equation (12).

[0154] Lightweight backoff controller: This mechanism ensures basic stability for cold chain logistics transport vehicles, even under extreme or unmodeled conditions, preventing loss of control. It is activated when the NMPC solver cannot find a feasible solution. This controller uses a high-gain proportional control law to directly apply to the changing state. :

[0155]

[0156] In the formula, and These represent the controller's revenue.

[0157] Stability analysis of the lightweight rollback controller:

[0158] The stability of the backoff controller can be rigorously proven using the Lyapunov method. Consider the quadratic Lyapunov candidate function used for the state transition as Equation (30), and its time derivative as Equation (31).

[0159]

[0160] By selecting negative fixed control gain The controller forms a negative feedback loop. Taking a yaw channel as an example: if... ,but This leads to negative consequences. and , making ,therefore The system state is driven back to the origin. .

[0161] Due to the change of state Boundedness It is the original tracking error Strictly within the preset performance envelope The necessary and sufficient condition for this is thus proved. The convergence of the property can theoretically guarantee the safety of the vehicle system in rollback mode.

[0162] The final execution logic adopts a hybrid switching strategy:

[0163] .

[0164] The system workflow is as follows:

[0165] (1) State prediction: At each control moment, the PDHR state predictor receives the current and historical vehicle states (such as sideslip angle β, yaw rate γ) and outputs the future state prediction. The precise state prediction sequence of the step (prediction time domain).

[0166] (2) Risk assessment: The MLE-QSA module receives the predicted state sequence, calculates its maximum Lyapunov exponent (MLE) sequence, and aggregates it into a continuous time-weighted instability risk (TWIR) factor. This factor can continuously quantify the risk level of future instability phenomena such as tilting or skidding of cold chain logistics transport vehicles, rather than a simple binary judgment.

[0167] (3) Adaptive Control Decision: The PS-PMPC controller receives the current vehicle state and the TWIR factor. First, it dynamically adjusts the weight matrix of its internal optimization problem according to the TWIR factor (Equations 13-15) to achieve risk adaptation. Second, it embeds predetermined performance control (PPC) into the optimization problem, transforming the hard constraint of state tracking error into a positive definite problem of unconstrained state transformation (Equations 21-26), and solves it to obtain the optimal front wheel steering angle correction. and additional yaw moment The control objective is to minimize cargo sway by controlling the vehicle as smoothly as possible while ensuring vehicle stability. If the NMPC solver fails, a lightweight backoff controller (Equations 29 and 32) is activated to ensure system feasibility.

[0168] (4) Torque distribution and execution: The multi-target torque distributor receives the required additional yaw moment. Under the premise of satisfying the physical constraints of each motor, the independent drive / braking torque of the motors in the four wheels is calculated and executed through an optimized allocation strategy. The allocation strategy can take into account both tire load uniformity and energy efficiency, adapting to the long-distance operation needs of cold chain logistics transport vehicles.

[0169] This invention employs a physical-data hybrid residual prediction mechanism: vehicle state prediction is decomposed into two parts: physical baseline prediction based on a simplified kinematic model and data-driven residual prediction based on a tabular basic model (TabPFN), and the two are superimposed. The key lies not in TabPFN itself, but in the hybrid architecture of "physical baseline ensuring basic rationality and consistency, and data-driven residual online correction of model errors and unmodeled dynamics," enabling it to achieve high-precision, physically consistent multi-step prediction even under extreme conditions of small sample size and time-varying loads and operating conditions.

[0170] This invention employs continuous stability risk quantification based on the Maximum Lyapunov exponent: using an improved Wolf algorithm, it directly calculates the Maximum Lyapunov exponent (MLE) sequence from the predicted future state time series, and further generates a single, continuously changing Time-Weighted Instability Risk (TWIR) factor through time-weighted aggregation. The key is the transformation of stability assessment from discrete, binary threshold judgments to continuous, quantifiable risk gradients. This provides a direct basis for gradient-based and refined early warning and decision-making for cold chain logistics transport vehicles before instability.

[0171] This invention employs a risk-adaptive predetermined performance model predictive control integration framework, specifically including:

[0172] (1) Risk-adaptive weight scheduling: The TWIR factor is mapped through the Sigmoid function to dynamically and smoothly adjust the state tracking weight and control quantity weight in the NMPC optimization problem (Equations 13-15), so as to realize the control strategy continuously adapts with risk. This enables the control system to focus on trajectory tracking efficiency when the vehicle is running at low risk, and to focus on stability assurance when the vehicle is running at high risk, which is especially suitable for cold chain logistics transport vehicles that require both stability and safety.

[0173] (2) Embedding of pre-defined performance control hard constraints: The hard performance boundary constraints of state tracking error are transformed into a positive definite problem for unbounded variables through the error transformation function (Equation 25), and this transformation is integrated into the prediction model and optimization objective of NMPC (Equations 27-28) so as to strictly ensure that the tracking error is always within the preset safe envelope at the optimization level.

[0174] (3) Lightweight rollback mechanism: A backup controller (Equation 29) based on transformed state proportional feedback is designed, independent of the main NMPC solver, and is seamlessly activated through hybrid switching logic (Equation 32) when the main solver fails, ensuring the feasibility of the system under any circumstances. This mechanism significantly improves the control robustness and operational fault tolerance of cold chain logistics transport vehicles under extreme or unpredictable operating conditions, avoiding serious accidents caused by controller failure.

Claims

1. A dynamic risk perception, prediction, and stability control system for cold chain logistics transport vehicles, characterized in that: It includes three core modules: a physical-data hybrid residual state predictor, a quantitative stability assessment module based on the maximum Lyapunov exponent, and an adaptive nonlinear model predictive controller that integrates predetermined performance control; finally, the control commands are distributed to the four in-wheel motors through a multi-objective torque distributor.

2. The dynamic risk perception, prediction, and stability control system for cold chain logistics transport vehicles according to claim 1, characterized in that, The described physical-data hybrid residual state predictor, used to predict future vehicle states, employs a hybrid residual framework: The final predicted state sequence is composed of the physical baseline prediction and the data-driven residual prediction: ； In the formula, For the final predicted state sequence, For physical baseline prediction, For data-driven residual prediction; Physical baseline prediction Based on a simplified discrete kinematics model; yaw rate Baseline predictions are propagated using a linear 2-DOF reference model: ； In the formula, For baseline prediction of yaw rate at time k in the future, Let yaw rate be the current angular velocity. The derivative of the current yaw rate. For the response coefficient, For time step, For the baseline prediction of the front wheel steering angle at time k in the future, The current steering angle of the front wheels; Data-driven residual prediction Prediction is performed using a pre-trained tab-based model, TabPFN; at each time step... The historical state residual sequence As contextual input, TabPFN leverages its zero-shot context learning capability to output the probability distribution of future residuals, taking the median as the deterministic residual prediction. .

3. The dynamic risk perception, prediction, and stability control system for cold chain logistics transport vehicles according to claim 1, characterized in that, The quantitative stability assessment module based on the maximum Lyapunov exponent is used to quantify the stability risk of future predicted sequences. The two initial distances are The distance between adjacent trajectories that evolves over time Approximately: ； In the formula, The Lyapunov index, For time; by taking the natural logarithm of both sides of equation (3), The definition is expressed in the following form: ； MLE computation directly analyzes the chaotic characteristics of state sequences. For a given predicted state time series, in phase space, for a reference point... Search for it nearest neighbors Calculate after a fixed evolution time Distance divergence after; local Lyapunov index Estimate using the following formula: ； The MLE sequence is obtained by traversing the prediction sequence. ; Time-weighted instability risk (TWIR) factor generation: By weighting over time, it focuses more on risks in the near future, which meets the need for early intervention in cold chain logistics transportation vehicles to prevent instability; (1) Define the time-weighted vector: ； in It is the attenuation factor; (2) Extracting the positive risk sequence: ； (3) Calculate the original TWIR factor: ； TWIR factor As a continuous risk indicator input to the controller.

4. The dynamic risk perception, prediction, and stability control system for cold chain logistics transport vehicles according to claim 1, characterized in that, The aforementioned pre-stabilized adaptive predetermined performance nonlinear model predictive controller integrates risk adaptation, predetermined performance constraints, and a backoff mechanism. Vehicle control model: A 2-DOF bicycle model is used as the internal prediction model; state vector With control input vector Defined as: ； In the formula, The sideslip angle is the angle of the centroid. The yaw rate is angular velocity. Add front wheel steering angle to the output. To add yaw moment to the output; The continuous-time state-space equation is shown below: ； In the formula, The driver's steering input is considered a measurable disturbance; assuming a small sideslip angle and linear tire characteristics in the prediction time domain, the dynamics are described by the following equation: ； in For vehicle quality, For longitudinal velocity, This is the yaw moment of inertia; and These are the distances from the center of gravity to the front and rear axles, respectively. and Indicates the steering stiffness of the tire; Constraints: To ensure the physical feasibility of the control action, the actuator must satisfy the saturation constraint condition: ； Risk perception adaptive weight scheduling: When the risk is low, the controller focuses on tracking the driver's intentions; when the risk is high, it decisively strengthens stability control, smoothly adapting to the needs of different driving stages of cold chain logistics transport vehicles; TWIR factor By mapping with the Sigmoid function, the tracking error weight matrix in the optimization problem is dynamically adjusted. and control weight matrix ,in and Defined by equations (13) and (14); ； ； ； ； In the formula, , These are the upper and lower bounds of the tracking error weights, respectively. , These are the upper and lower bounds of the control weights, respectively. The center value of the TWIR factor sequence, , These are the sensitivity coefficients; Pre-defined performance control PPC conversion and embedding: (1) Reference generation and constraints: ideal reference value and The calculations are performed using the linear 2-DOF model, i.e., Equations 17 and 18, and are constrained by the road surface friction coefficient. Obtain the final reference value and That is, Equations 19 and 20: ； ； ； ； (2) Performance boundary and error transformation: defining tracking error That is, Equation 21; Design the exponentially decaying performance function That is, Equation 23, requires the error to satisfy That is, Equation 22; ； ； In the formula, and These are the tracking errors of the sideslip angle and yaw rate, respectively. This is the initial value for the performance function. This represents the steady-state value of the performance function. The convergence rate is determined by a strictly increasing logarithmic transformation function. Normalization error That is, Equation 24, which maps to an unbounded transformation state. : ； Its inverse transformation is: ； Therefore, the original state constraint problem is transformed into a problem of... Unconstrained positive definite problems; (3) Adaptive PPC-NMPC optimization problem: at each control time step Solve the following finite-time optimization problem: ； in It is the predicted value of the transformed state. It is the control input sequence to be optimized. and It is an adaptive weight matrix; (4) PPC-NMPC prediction model and constraints: Since vehicle dynamics are determined by physical states The description is as shown in Equation 10), while NMPC optimizes the state transition. Therefore, it is necessary to establish a bridge between the two in the prediction process; the "inverse prediction-forward recursive update" strategy is adopted: in each prediction step, the current transformed state is first mapped back to the physical state through inverse transformation, and then the state derivative is calculated using the vehicle dynamics model. Then, the derivative of the state transformation is derived using the chain rule. : ； The future state transformation is propagated forward through Euler integral; at the same time, the control input must satisfy the physical saturation constraint of the actuator, as shown in equation (12); Lightweight backoff controller: This mechanism ensures basic stability for cold chain logistics transport vehicles, even under extreme or unmodeled conditions, preventing loss of control; it is activated when the NMPC solver cannot find a feasible solution; this controller uses a high-gain proportional control law to directly apply to the changing state. : ； In the formula, and These are the controller's revenues; Stability analysis of the lightweight rollback controller: Consider the quadratic Lyapunov candidate function used for state transformation as Equation (30), and its time derivative as Equation (31). ； ； By selecting negative fixed control gain The controller forms a negative feedback loop; if ,but This leads to negative consequences. and , making ,therefore The system state is driven back to the origin. ; By proof The convergence of this property can theoretically guarantee the safety of the vehicle system in rollback mode; The final execution logic adopts a hybrid switching strategy: 。 5. A dynamic risk perception, prediction, and stability control method for cold chain logistics transport vehicles, characterized in that... The method employs the dynamic risk perception, prediction, and stability control system for cold chain logistics transport vehicles as described in any one of claims 1 to 4; the method includes the following steps: (1) State prediction: At each control moment, the PDHR state predictor receives the current and historical vehicle states and outputs the future state. A precise state prediction sequence for each step; (2) Risk assessment: The MLE-QSA module receives the predicted state sequence, calculates its maximum Lyapunov exponent MLE sequence, and aggregates it into a continuous time-weighted instability risk (TWIR) factor; this factor can continuously quantify the risk level of future instability phenomena such as tilting or skidding of cold chain logistics transport vehicles, rather than a simple binary judgment. (3) Adaptive control decision: The PS-PMPC controller receives the current vehicle state and TWIR factor; first, it dynamically adjusts the weight matrix of its internal optimization problem according to the TWIR factor to achieve risk adaptation; second, it embeds the predetermined performance control PPC into the optimization problem, transforms the hard constraint of state tracking error into a positive definite problem for unconstrained state transformation, and solves it to obtain the optimal front wheel steering angle correction. and additional yaw moment The control objective is to minimize cargo swaying by controlling the vehicle as smoothly as possible while ensuring vehicle stability. If the NMPC solver fails, a lightweight backoff controller is activated to ensure system feasibility. (4) Torque distribution and execution: The multi-target torque distributor receives the required additional yaw moment. Under the premise of satisfying the physical constraints of each motor, the independent drive / braking torque of the motors in the four wheels is calculated and executed by optimizing the allocation strategy; the allocation strategy can take into account both tire load uniformity and energy efficiency, and adapt to the long-distance operation needs of cold chain logistics transport vehicles.

Images