A Fault Tolerance Method and Fault Tolerance System for a Five-Axis Heavy-Duty AGV

Abstract

This invention discloses a fault containment method and system for a five-axis heavy-duty AGV, relating to the field of port automation equipment. To address the shortcomings of existing five-axis heavy-duty AGVs in resisting the risk of hub motor failure, this invention acquires the operating status of each wheel's hub motor and calculates dynamic fault factors based on the acquired status data to determine the motor's fault level. It generates dynamic control commands to provide input instructions for fault response and control strategies for dynamic fault tolerance. Fault recovery and safety chain execution are implemented by reallocating the motor torque of non-faulty wheels according to the fault level and control strategy, while simultaneously feeding back the system status to the control command generation module in real time to adjust the control strategy. The invention also compares a virtual model with a physical model, providing feedback to the control command generation module, and continuously optimizing system performance by extracting historical data. This invention is primarily used for timely handling of failures in intelligent transfer robots.

Description

Technical Field

[0001] This invention relates to the field of port automation equipment, and in particular to a fault containment method and fault containment system for a five-axis heavy-duty AGV applied to a roll-on / roll-off terminal. Background Technology

[0002] Due to the harsh working environment of the heavy-duty AGVs (automated guided vehicles, i.e., intelligent transfer robots) responsible for transferring vehicles inside RoRo ships, their hub motors are at risk of failure. Existing technologies suffer from the following shortcomings that urgently need to be addressed:

[0003] First, there is a lack of multi-wheel failure control strategies and a lack of systemic fault tolerance. Existing AGV failure control technologies mainly focus on single-wheel failure scenarios, but lack effective solutions for failure of two or more drive wheels in the same area. When two wheels fail, the vehicle will fail and the yaw torque will become unbalanced, leading to loss of energy distribution control.

[0004] Secondly, the insufficient precision of mechanical coupling control has a significant impact on the torque redistribution after motor failure, which can easily lead to problems such as control model failure, energy distribution imbalance, and dynamic response lag.

[0005] Third, existing heavy-duty AGVs generally adopt motor redundancy design, but only implement physical hardware backup and do not build a dynamic redundancy allocation mechanism. The utilization rate of hardware redundancy is low, and motor failures lead to idle redundant resources, resulting in significant additional costs.

[0006] Fourth, after a motor failure, in order to maintain stability, the transfer vehicle speed will be greatly reduced, resulting in low efficiency at low speeds and insufficient stability at high speeds. It relies on manual intervention and is difficult to adapt to various scenarios.

[0007] Finally, fault diagnosis relies on manual methods. It lacks the ability to fusion and diagnose different fault causes, and relies on manual troubleshooting, resulting in low fault-finding efficiency, a high rate of missed faults, and a long time for a single manual fault-finding process.

[0008] Therefore, there is a need for a fault containment method and system for five-axis heavy-duty AGVs that can resist the risk of hub motor failure, improve the transfer efficiency of heavy-duty AGVs, and reduce manual working time. Summary of the Invention

[0009] To address the shortcomings of existing five-axis heavy-duty AGVs that cannot resist the risk of hub motor failure, this invention provides a fault-tolerant method and system for five-axis heavy-duty AGVs that can resist the risk of hub motor failure, improve the transfer efficiency of heavy-duty AGVs, and reduce manual working time.

[0010] The fault containment method for a five-axis heavy-duty AGV according to the present invention includes the following steps:

[0011] Step 1: Acquire the status data of the five-axis heavy-duty AGV in real time through the sensor group, obtain the working status data of the hub motor of each wheel, and calculate the dynamic fault factor h based on the acquired working status data;

[0012] Step 2: Calculate the working condition weighting factor w based on the current working condition, and determine the fault level of each wheel hub motor based on the dynamic fault factor h obtained in Step 1.

[0013] Step 3: Generate dynamic control commands through the control command generation module to provide input commands for fault response and control strategies for dynamic fault tolerance;

[0014] Step 4: Fault recovery and safety chain execution. Based on the fault level and control strategy, the torque of the motor in the non-faulty wheel hub is redistributed, and the system status is fed back to the control command generation module in real time to adjust the control strategy.

[0015] Step 5: The feedback module compares the virtual model with the physical model and then provides feedback to the control command generation module. At the same time, it extracts historical data to continuously optimize system performance.

[0016] Further: In step 1, the dynamic fault factor is calculated based on the acquired state data. The specific process is as follows:

[0017] ;

[0018] in For current, Let A be the rotational speed and A be the vibration. h is temperature; i is dynamic fault factor; j is speed weight factor; k is vibration weight factor;

[0019] The sensor group includes a current sensor, an encoder, a wheel speed sensor, a vibration sensor, and / or a temperature sensor;

[0020] The status data includes motor operating current, motor speed, wheel linear speed, mechanical vibration, and motor temperature.

[0021] Further: In step 2, the specific steps for determining the fault level of the motor are as follows:

[0022] Step 21: Calculate the operating condition weighting factor :

[0023] ;

[0024] Where a is the gradient weight factor, b is the vehicle speed weight factor, c is the load weight factor, S is the gradient coefficient, V is the vehicle speed coefficient, and β is the load coefficient.

[0025] Step 22: Based on dynamic fault factors With operating condition weighting factor Obtain the fault level of the motor;

[0026] like This means that one motor of a single wheel in the front or rear section fails, which is considered a Level 1 fault.

[0027] like This means that either a single-wheel dual-motor failure occurs completely in a single area or a single-wheel dual-motor failure occurs in each of two areas, which is considered a Level 2 fault.

[0028] like This means that two wheels on opposite sides of the same area fail, or two wheels on opposite sides of each of two areas fail, which is considered a level 3 fault.

[0029] like This means that there are two motors on the same side of a certain area that have failed, or at least three motors in a certain area that have failed, or a power unit motor that has failed. This is considered a level 4 fault.

[0030] If the packet loss rate L between the front and rear chips is ≥30%, indicating sensor failure, it is considered a level 5 fault. , , , These are the judgment thresholds corresponding to fault levels 1 to 4.

[0031] Further: In step 4, the specific steps of the control strategy for level 1 faults are as follows:

[0032] First, in-wheel hardware redundancy is used to isolate the faulty motor to prevent the failure from propagating; the control command generation module includes a model prediction control module and a reinforcement learning module.

[0033] The dual in-wheel motors are switched to the healthy motor. The healthy motor is monitored by a temperature sensor. The model predictive control module adjusts the additional yaw torque to compensate for the torque shortfall of the failed motor. The reinforcement learning module balances efficiency and temperature rise, thereby limiting the temperature of the healthy motor.

[0034] The faulty motor enters "hot standby mode," maintaining only a portion of the current for status monitoring.

[0035] Furthermore: In step 4, the specific steps of the control strategy for level 2 faults are as follows:

[0036] The model prediction control module tightens the yaw rate error constraint to control vehicle stability;

[0037] The reinforcement learning module uses a "torque increase on the same side + torque decrease on the opposite side" strategy to increase the weight of "yaw stability" in the reward function and output actions to make the healthy motor on the same side fully loaded, thereby increasing the path curvature.

[0038] Furthermore: In step 4, the specific steps of the control strategy for level 3 faults are as follows:

[0039] The torque is evenly distributed to the other working wheels where the dual motors have not failed. The model prediction control module keeps the torque of all working wheels equal and minimizes it to avoid unwanted yaw torque.

[0040] At the same time, the reinforcement learning module forces the electronic differential lock to lock the normal working wheels, replans the path downgrade, and selects low curvature road sections with curvature less than the preset curvature and road sections without slopes.

[0041] Furthermore: In step 4, the specific steps of the control strategy for level 4 and level 5 faults are as follows:

[0042] A three-tiered safety chain of "communication emergency response - combined braking - mechanical locking" is adopted to ensure equipment safety under extreme failure conditions;

[0043] Emergency Communication: Switch to the backup channel, which is a LoRa channel, used to broadcast fault codes to the central control system;

[0044] Combined braking: Redundancy activation of electric motor braking combined with hydraulic braking;

[0045] Mechanical locking: Dual electromagnetic locks lock synchronously, and the optimal locking point is predicted by the feedback module to prevent slippage.

[0046] Furthermore: In step 5, the specific steps of the feedback module comparing the virtual model with the physical model are as follows:

[0047] Step 51: The feedback module includes a digital twin calibration module and a transfer learning module; the digital twin calibration module optimizes and enhances the model prediction control module based on the virtual model, compares the output of the virtual model with that of the physical system, and adjusts the weight of the model prediction control module in real time based on the error value obtained from the comparison.

[0048] Step 52: The transfer learning module aligns the distribution of historical faults with the current scene features based on the domain adversarial network; it extracts similar fault policy skeletons from the historical database, thereby enabling fine-tuning of the reinforcement learning module network.

[0049] The fault containment system for implementing the fault containment method of a five-axis heavy-duty AGV according to the present invention includes a sensor group, a fault level judgment module, a control command generation module and a feedback module.

[0050] The sensor group is used to acquire the status data of the five-axis heavy-duty AGV in real time.

[0051] The fault level judgment module is used to calculate dynamic fault factors based on status data, and to determine the fault level of the motor based on the dynamic fault factors and the operating condition weight factor.

[0052] The control command generation module is used to generate dynamic control commands according to a preset control strategy.

[0053] The feedback module is used to compare the simulation results with the actual situation, and then feed the comparison results back to the control command generation module, thereby optimizing the dynamic control commands.

[0054] Furthermore: the five-axis heavy-duty AGV includes a front section, a rear section, and a power unit. The front section includes a first axis and a second axis, the rear section includes a third axis and a fourth axis, and the power unit includes a fifth axis; thus forming a five-axis, ten-wheel structure, and each wheel uses a hub motor with independently controlled torque; the front section and the rear section exchange data via wireless communication.

[0055] The beneficial effects of this invention are:

[0056] This invention addresses brake fault containment by providing a fault containment method and system for diagnosing and controlling five-axis heavy-duty AGVs under different motor failure conditions. It employs different strategies based on the vehicle's current fault level and switches fault containment modes according to efficiency and safety considerations. Utilizing different control strategies, it achieves the required functions and needs of vehicle operation under each fault level. Furthermore, this method reduces the risk of motor failure, saves human resources, and aligns with the green and intelligent requirements of my country's port transportation.

[0057] Compared with the prior art, the technical effects and advantages of the present invention are:

[0058] 1. Dynamic adaptability: Adapts to complex terminal environments through operating condition weighting factors and hybrid control strategies;

[0059] 2. High reliability: The three-tier safety chain design ensures zero-risk downtime under extreme failures;

[0060] 3. Self-evolution capability: Digital twins and transfer learning enable continuous system optimization; using digital twin systems to collect scenario data can improve the response speed when faults occur in similar scenarios; it improves the utilization rate of redundant hardware in AGVs, and the construction of MPC+RL joint control strategy improves the utilization rate of remaining healthy motor torque when motors fail;

[0061] 4. Layered and progressive response: Four-level fault diagnosis from sensor level to vehicle level, with differentiated processing strategies matched to each level; different control strategies are made according to the fault level, and different control strategies are used to realize the functions and requirements of vehicle operation under each fault level; through multi-sensor fusion and dynamic hierarchical strategy, millisecond-level fault response is achieved, and MPC+RL collaborative control compresses the stability error to within 5%, improves fault diagnosis efficiency, and reduces fault location time.

[0062] 5. Scene Adaptive Control: Independent fault-tolerant logic for low-speed / high-speed scenes, balancing efficiency and safety.

[0063] This invention enables intelligent transfer vehicles to maintain controllable operation or safe shutdown in various scenarios, from single motor failure to multi-area faults, through refined hierarchical and cross-layer collaborative control, thus meeting the stringent reliability requirements of port automation equipment. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the hardware configuration structure of a five-axis heavy-duty AGV;

[0065] Figure 2 This is a diagram of the intelligent transfer vehicle system architecture;

[0066] Figure 3 This is a flowchart of dynamic fault classification and control mode switching;

[0067] Figure 4 This is a schematic diagram of a collaborative framework between Model Predictive Control (MPC) and Reinforcement Learning (RL).

[0068] Figure 5 This is a three-level security chain execution logic diagram;

[0069] Figure 6 This is a schematic diagram of digital twin calibration and transfer learning;

[0070] In the diagram, 1 is the first axis, 2 is the second axis, 3 is the third axis, 4 is the fourth axis, 5 is the fifth axis, 6 is the controller, 7 is the steering motor, 8 is the rear wheel, and 9 is the clamping mechanism. Detailed Implementation

[0071] The following are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. The embodiments described below are only for explaining the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention should be determined by the scope of the claims. The embodiments of the present invention are described in detail below. In order to facilitate the description of the present invention and simplify the description, the technical terms used in the specification of the present invention should be interpreted broadly, including but not limited to conventional alternatives not mentioned in this application, and including both direct and indirect implementation methods.

[0072] Example 1

[0073] Combination Figure 1 and Figure 6 This embodiment describes a fault containment method for a five-axis heavy-duty AGV, comprising the following steps:

[0074] Step 1: Real-time status data of the five-axis heavy-duty AGV is acquired through a sensor group (sampling rate 1kHz). The sensor group includes current sensors, encoders, wheel speed sensors, vibration sensors, temperature sensors, etc., to acquire the working status data of the wheel hub motors, and calculate the dynamic fault factor h based on the acquired working status data to determine the fault status of the motor.

[0075] like Figure 3 As shown, the specific process of calculating the dynamic fault factor h based on the acquired state data is as follows:

[0076] ;

[0077] in For current, For speed, For vibration, h is temperature; i is dynamic fault factor; j is speed weight factor; k is vibration weight factor;

[0078] The operating status data includes motor operating current, motor speed, wheel linear speed, mechanical vibration, and motor temperature.

[0079] Step 2: Calculate the operating condition weighting factor w and determine the motor's fault level based on the dynamic fault factor h obtained in Step 1.

[0080] In step 2, the specific logical steps for classifying faults of different degrees are as follows:

[0081] Step 21: Calculate the operating condition weighting factor :

[0082] ;

[0083] Where a is the gradient weight factor, b is the vehicle speed weight factor, c is the load weight factor, S is the gradient coefficient, V is the vehicle speed coefficient, and β is the load coefficient.

[0084] Step 22: Based on dynamic fault factors With operating condition weighting factor Obtain the fault level of the motor;

[0085] The threshold for judging the failure of one motor in a single wheel of the front or rear axle is obtained based on the dynamic failure factor h and the operating condition weight factor w. In this case, it is considered a Level 1 fault;

[0086] Based on the dynamic fault factor h and the operating condition weight factor w, the threshold for determining whether a single-wheel dual-motor system in a single area is completely failed or whether each of the two areas has a single-wheel dual-motor system failing is as follows: In this case, it is considered a level 2 fault;

[0087] Based on the dynamic fault factor h and the operating condition weight factor w, the threshold for determining whether two wheels on opposite sides of the same area fail or two wheels on opposite sides of each of two areas fail is: In this case, it is considered a level 3 fault;

[0088] Based on the dynamic fault factor h and the operating condition weight factor w, the threshold for determining whether there is a failure of two motors on the same side in a certain area, or a failure of multiple motors (≥3) in a certain area, or a failure of the power unit motor (fifth axis steering drive motor) is: This is considered a level 4 fault.

[0089] If the packet loss rate L between the front and rear chips is ≥30%, indicating sensor failure, it is considered a level 5 fault. , , , These are the judgment thresholds corresponding to fault levels 1 to 4.

[0090] For example, the determination of the fault level is shown in Table 1:

[0091] Table 1 Fault Level Determination Criteria

[0092]

[0093] Level 1-3 faults: Activate hybrid control (MPC+RL);

[0094] Level 4-5 fault: Triggers a three-level safety chain (communication-braking-locking).

[0095] Step 3: Generate dynamic control commands through model predictive control (MPC) and reinforcement learning (RL) to provide key inputs for fault response and control strategies for dynamic fault tolerance;

[0096] like Figure 4 As shown, MPC is a feedforward-feedback control method based on a dynamic model. In intelligent transport vehicles, the goal of MPC is to maintain yaw stability and path tracking accuracy, while considering motor torque limitations.

[0097] State-space model calculation:

[0098] Ideal yaw rate for:

[0099] ;

[0100] ;

[0101] in: Here, is the ideal yaw rate, u is the longitudinal speed, L is the wheelbase, K is the stability factor, and m is the total vehicle mass. It is the distance from the center of gravity to the front axle. It is the distance from the center of mass to the intermediate axis. It is the distance from the center of gravity to the rear axle. It refers to the lateral stiffness of the front tires. It refers to the lateral stiffness of the rear tire. It's the steering angle.

[0102] Subject to road surface adhesion coefficient Limiting maximum yaw rate satisfy:

[0103] ;

[0104] In the picture, is the road surface adhesion coefficient, and g is the acceleration due to gravity.

[0105] In summary, the ideal yaw rate of the vehicle is:

[0106] ;

[0107] In the picture, This is the theoretical ideal yaw rate considering road surface adhesion limitations. () is a sign function that returns the sign of the expression within the parentheses.

[0108] Vehicle dynamics is simplified to a three-degree-of-freedom model (longitudinal, lateral, and yaw), with the following state equations:

[0109] ;

[0110] in: It is the centroid sideslip angle. It is the actual yaw rate. It's the rear wheel steering angle. It is an additional yaw moment. It is the total mass of the vehicle. It is the moment of inertia about the Z-axis. It refers to the lateral stiffness of the front tires. It refers to the lateral stiffness of the rear tires.

[0111] Optimize the objective function:

[0112] ;

[0113] in, The ideal yaw rate (calculated from the reference model). It is the weight of yaw rate tracking accuracy. It is a matrix control energy consumption weight matrix that balances tracking accuracy and control energy consumption.

[0114] Constraints:

[0115] Torque limit: ,in This is the motor torque. This refers to the rated torque of the motor.

[0116] Yaw stability: .

[0117] MPC has two functions in fault tolerance: the first is prediction and compensation. MPC uses predictive models to identify potential instability trends in advance (such as an increase in yaw rate deviation) and dynamically adjusts the additional yaw moment. The first is to offset the torque loss of the faulty motor. The second is constraint protection: when a motor overload (such as a level 1 fault) is detected, MPC automatically tightens the torque constraint to prevent the healthy motor from overloading.

[0118] Reinforcement Learning (RL):

[0119] Reinforcement Algorithm (RL) learns the optimal policy through the interaction between the agent and the environment to maximize cumulative reward. In this scheme, RL is used to dynamically optimize the parameters of the Multiprocessor Control Program (MPC), such as the weight matrix. , It also incorporates torque distribution strategies to adapt to complex fault scenarios.

[0120] State space s (State):

[0121] ;

[0122] It is the longitudinal speed of the intelligent transfer vehicle. It's the slope angle. These are the temperatures of each motor.

[0123] Action space a (Action):

[0124] ;

[0125] Torque distribution coefficient That is, the torque amplification ratio of a healthy motor;

[0126] Path curvature adjustment amount That is, to avoid losing control on sharp bends.

[0127] Reward function R(Reward):

[0128] ;

[0129] in, For the target vehicle speed, Temperature threshold This indicates a deviation from the path.

[0130] Bonus items: vehicle speed tracking (40%), yaw stability (30%);

[0131] Penalties: Motor temperature rise (20%), path deviation (10%).

[0132] Policy Network (Actor-Critic Architecture):

[0133] Actor Network: Output Action Policy gradient update. Critic network: evaluates the state value V(s) and guides the Actor's optimization direction.

[0134] The role of RL in fault tolerance is mainly twofold: firstly, dynamic parameter adjustment, for example, when a level 2 fault occurs (failure of two motors in a single area), RL is increased. Prioritize compensating for the torque of the healthy motor on the same side, while adjusting... Reduce path curvature to lower yaw moment requirements. Second, explore and utilize balance. In Level 3 faults (cross-regional failures), RL can quickly generate emergency control strategies (such as differential lock activation threshold optimization) by exploring historical data (such as similar fault cases).

[0135] Step 4, Fault Recovery and Safety Chain Execution: Combining the fault level obtained in Step 2 with the control strategy in Step 3, the motor torque of the non-faulty wheels is redistributed, and the system status is fed back to MPC and RL in real time to adjust the control strategy.

[0136] In step 4, the specific steps of the control strategy for Level 1 faults are as follows:

[0137] First, in-wheel hardware redundancy is used to quickly isolate faulty motors and prevent failure propagation; the control command generation module includes a model predictive control module (MPC) and a reinforcement learning module (RL).

[0138] The dual in-wheel motors are switched to the healthy motor. The healthy motor is monitored by a temperature sensor. The model predictive control module adjusts the additional yaw torque to compensate for the torque shortfall of the failed motor. The reinforcement learning module balances efficiency and temperature rise, thereby limiting the temperature of the healthy motor.

[0139] The faulty motor enters "hot standby mode," maintaining only a portion of the current for status monitoring.

[0140] The specific steps of the control strategy for Level 2 faults are as follows:

[0141] The model prediction control module tightens the yaw rate error constraint to control vehicle stability;

[0142] The reinforcement learning module uses a "torque increase on the same side + torque decrease on the opposite side" strategy to increase the weight of "yaw stability" in the reward function and output actions to make the healthy motor on the same side fully loaded, thereby increasing the path curvature.

[0143] The specific steps of the control strategy for Level 3 faults are as follows:

[0144] The torque is evenly distributed to the other working wheels where the dual motors have not failed. The model prediction control module keeps the torque of all working wheels equal and minimizes it to avoid unwanted yaw torque.

[0145] At the same time, the reinforcement learning module forces the electronic differential lock to lock the normal working wheels, replans the path downgrade, and selects low curvature road sections with curvature less than the preset curvature and road sections without slopes.

[0146] The specific steps of the control strategy for level 4 and level 5 faults are as follows:

[0147] like Figure 5 As shown, a three-level safety chain of "communication emergency rescue - combined braking - mechanical locking" is adopted to ensure equipment safety under extreme failure conditions;

[0148] The three-tiered safety chain consists of: Tier 1: "Communication Emergency Rescue" (switching to the backup channel and broadcasting a fault code). Tier 2: "Combined Braking" (motor braking → hydraulic braking, with a response time of 50ms → 100ms). Tier 3: "Mechanical Locking" (dual electromagnetic locks activated, with a locking force ≥ 5000N).

[0149] The Level 3 safety chain is triggered in the event of a Level 4 or Level 5 fault. A Level 4 fault (red warning box) directly triggers the Level 3 safety chain; a Level 5 fault (yellow warning box) prioritizes communication recovery, and if that fails, triggers braking. If the Level 3 safety chain fails to be triggered during the execution of a Level 4 or Level 5 fault, a red "emergency alarm" signal is displayed.

[0150] Emergency Communication: Switch to the backup channel, which is a LoRa channel, used to broadcast fault codes to the central control system;

[0151] Combined braking: Redundancy activation of electric motor braking combined with hydraulic braking;

[0152] Mechanical locking: Dual electromagnetic locks lock synchronously, and the optimal locking point is predicted by the feedback module to prevent slippage.

[0153] Step 5: Digital twin calibration and transfer learning. By comparing the virtual model with the real situation, feedback is given to the MPC. At the same time, historical data is extracted to continuously optimize system performance. The feedback module compares the virtual model with the physical model and then gives feedback to the control command generation module. At the same time, historical data is extracted to continuously optimize system performance.

[0154] In step 5, the feedback module compares the virtual model with the physical model using the following specific steps:

[0155] Step 51: The feedback module includes a digital twin calibration module and a transfer learning module; the digital twin calibration module optimizes and enhances the model predictive control module based on the virtual model, compares the output of the virtual model with that of the physical system, and adjusts the weights of the model predictive control module in real time based on the error value obtained from the comparison; the virtual model is the model on which the "digital twin calibration module" relies, namely, a "high-precision dynamic model" based on a three-degree-of-freedom vehicle model; the physical model refers to the actual transport robot entity, including hardware components such as hub motors, sensors, and controllers. The output parameters of the virtual model and the physical model include motion state, motor state, and dynamic adjustment effect after a fault occurs. The comparison between the physical model and the virtual model is performed by the "digital twin calibration module" of the feedback module in the patent, and the specific process is as follows:

[0156] The virtual model outputs predicted results (such as ideal yaw rate and motor torque distribution scheme) based on input parameters (such as fault level and operating conditions); the physical system (the physical transport robot) outputs actual operating results (such as actual yaw rate and actual motor torque); the digital twin calibration module calculates the error between the two and adjusts the weight matrix of the model predictive control module (MPC) in real time according to the error (optimizing the Q and R weights in the objective function), so that the prediction of the virtual model is closer to the actual performance of the physical system, and ultimately improves the accuracy of the control strategy.

[0157] The virtual model belongs to the digital twin calibration module within the feedback module; the model it relies on is the high-precision dynamic model. The virtual model is not the digital twin calibration module itself, but rather a tool used by this module to optimize the control strategy (the digital twin calibration module adjusts the MPC by comparing the virtual model with the physical system).

[0158] The physical system, or physical model, refers to the actual physical entity of the transfer robot, including hardware components such as hub motors, sensors, and controllers. The physical model is not the Model Predictive Control Module (MPC). The MPC is part of the control instruction generation module, which is used to calculate control strategies based on the model. The physical model is the physical device being controlled.

[0159] The output parameters of the virtual model (high-precision dynamic model) and the physical model (transfer robot entity) include:

[0160] Motion status: vehicle speed, yaw rate, center of gravity sideslip angle, path tracking accuracy, etc. (corresponding to the core indicators of MPC optimization in the patent);

[0161] Motor status: output torque, temperature, vibration, etc. (status data collected by the corresponding sensor group);

[0162] Fault response: The dynamic adjustment effect after a fault occurs (such as stability after torque redistribution, braking distance, etc.);

[0163] The feedback module's digital twin calibration module executes the following process:

[0164] The virtual model outputs prediction results (such as ideal yaw rate and motor torque distribution scheme) based on input parameters (such as fault level and operating conditions).

[0165] The physical system (the physical transfer robot) outputs the actual operating results (such as the actual yaw rate and the actual torque of the motor).

[0166] The digital twin calibration module calculates the error between the two and adjusts the weight matrix of the model predictive control module (MPC) in real time based on the error (optimizing the Q and R weights in the objective function), so that the prediction of the virtual model is closer to the actual performance of the physical system, and ultimately improves the accuracy of the control strategy.

[0167] Step 52: The transfer learning module aligns the feature distributions of historical faults and the current scene based on a domain adversarial network (DANN); it extracts similar fault policy skeletons from the historical database, thereby fine-tuning the reinforcement learning module network. The determination of the similarity between historical faults and the current scene is primarily achieved through aligning feature distributions using a DANN. The "features" to be aligned include key parameters of historical faults and the current scene, such as fault level, operating parameters, motor status, and motion state. The DANN learns the feature distribution patterns of "historical fault data (source domain)" and "current scene data (target domain)." When the feature distributions of the two types of data tend to be consistent (the distribution difference is less than a preset threshold), they are judged to be "similar." The extracted similar fault strategy skeleton is a "similar fault strategy skeleton" that combines historical fault data with the corresponding scenario feature distribution. It does not simply extract historical fault data or isolated scenario feature distributions, but rather extracts the "strategy skeleton" formed by the relationship between the two—that is, the core logical framework for the system to handle faults in historically similar fault scenarios (e.g., the strategy logic of "isolation of the failed motor + torque compensation of the healthy motor" for Level 1 faults, and the control framework of "torque increase on the same side + torque decrease on the opposite side" for Level 2 faults). These "strategy skeletons" are the crystallization of processing experience with historical fault data under specific scenario features. After the transfer learning module confirms the similarity between the current scenario and historical scenarios through the domain adversarial network, it extracts such skeletons to fine-tune the network parameters of the reinforcement learning module, enabling the current control strategy to quickly adapt to similar fault scenarios.

[0168] The "transfer learning module" of the feedback module is responsible for transferring historical fault experience to the current scenario, while the domain adversarial network is the key technical means of the transfer learning module, used to solve the problem of "inconsistency between historical fault data and current scenario feature distribution" (such as the difference in fault features under different slopes, loads, and vehicle speeds).

[0169] This embodiment continuously optimizes the control strategy of the reinforcement learning module using historical data, improving the system's adaptability to complex faults. The role of the domain adversarial network is:

[0170] Align the feature distributions (such as fault level, slope coefficient, motor temperature, yaw rate, etc.) of historical faults (source domain) with the current scenario (target domain) to eliminate the problem of "historical strategies not being suitable for the current scenario" caused by distribution differences;

[0171] This provides a technical foundation for "extracting similar fault strategy skeletons from the historical database," ensuring that the extracted strategy skeletons can effectively adapt to the current scenario, thereby enabling fine-tuning of the reinforcement learning module network (such as adjusting torque distribution coefficients, path curvature parameters, etc.).

[0172] This embodiment constructs a closed-loop fault tolerance system of "perception-decision-execution-optimization": sensor perceives the state → fault classification → control strategy execution → feedback module optimizes the strategy. Domain adversarial networks, as a key technology in the feedback optimization stage, ensure the effective transfer of historical data, enabling the reinforcement learning module to quickly adapt to new fault scenarios and avoid repeated "exploration from scratch." This improves the system's response speed and control accuracy under diverse faults (such as failures of varying degrees from level 1 to 3), ultimately supporting the entire system to achieve "self-evolution" fault tolerance.

[0173] This is achieved by aligning the feature distribution through Domain Adversarial Networks (DANNs).

[0174] The "features" that need to be aligned include key parameters of historical faults and current scenarios, such as fault level (levels 1-5), operating parameters (gradient coefficient S, vehicle speed coefficient V, load coefficient β), motor status (current I, speed v, temperature T, vibration A), and motion status (yaw rate γ, centroid sideslip angle β, path curvature, etc.).

[0175] What is extracted is a "skeleton of similar failure strategies" that combines historical failure data with the feature distribution of corresponding scenarios:

[0176] It does not simply extract historical fault data (such as temperature and current records of a certain motor) or isolated scene feature distributions (such as vehicle speed distribution under a certain slope), but extracts the "strategy skeleton" formed by the relationship between the two - that is, the core logic framework for the system to handle faults in similar fault scenarios in history (for example: the strategy logic of "isolation of failed motor + torque compensation of healthy motor" in the case of Level 1 fault, the control framework of "torque increase on the same side + torque reduction on the opposite side" in the case of Level 2 fault, etc.).

[0177] These "strategy skeletons" are the culmination of experience in processing historical fault data under specific scenario characteristics. After the transfer learning module confirms the similarity between the current scenario and the historical scenario through the domain adversarial network, it extracts such skeletons to fine-tune the network parameters of the reinforcement learning module, so that the current control strategy can quickly adapt to similar fault scenarios.

[0178] In other words, this embodiment judges similarity by aligning feature distribution through domain adversarial networks, and ultimately extracts a "reusable strategy framework formed by combining historical fault data and scene features", rather than a single data or distribution.

[0179] Domain adversarial networks learn the feature distribution patterns of "historical fault data (source domain)" and "current scenario data (target domain)". When the feature distributions of the two types of data tend to be consistent (the distribution difference is less than a preset threshold), they are judged as "similar".

[0180] Physical system: Transfer robot entity (labeled hub motor, sensor, controller).

[0181] Virtual model: High-precision dynamic model (based on a three-degree-of-freedom vehicle model), output comparison module.

[0182] Transfer learning module: Historical fault database (source domain). Real-time fault scenario domain adversarial network (DANN) (target domain).

[0183] The digital twin platform dynamically optimizes the MPC weight matrix and fault recovery threshold by comparing the outputs of the physical system and the virtual model. For example, if the virtual model predicts that a certain motor has a high probability of failure at a specific slope, RL will increase the health monitoring weight of the motor in advance and adjust the fault classification threshold to trigger recovery measures more quickly.

[0184] Transfer learning enhances cross-scenario adaptability: Historical fault data is aligned to new scenarios (target domains) through Domain Adversarial Networks (DANNs), enabling MPC strategies to generalize. For example, if a certain type of combined fault (such as Level 2 + communication delay) occurs frequently in historical data, transfer learning will optimize RL strategies and communication switching logic, shortening response time.

[0185] The feedback module compares the virtual model with the physical model and feeds back the output error information of both to the control command generation module (Model Predictive Control Module MPC + Reinforcement Learning Module RL). Specifically, this includes:

[0186] Motion state error: The deviation between the ideal yaw rate, center of mass sideslip angle, path tracking accuracy, etc., predicted by the virtual model and the corresponding parameters in the actual operation of the physical model (transfer robot entity) (such as whether the yaw rate deviation exceeds the 5% threshold).

[0187] Motor condition error: The difference between the motor torque distribution, temperature change, vibration amplitude, etc. simulated by the virtual model and the actual sensor data measured in the physical model (such as the deviation between the actual temperature rise of a healthy motor and the predicted value).

[0188] Fault response error: The difference between the fault recovery effect predicted by the virtual model (such as the stability after torque redistribution) and the actual fault handling result of the physical model (such as the degree of path deviation).

[0189] The feedback error information is used to update the core parameters of the control command generation module, specifically including:

[0190] Model Predictive Control (MPC): Updates the weight matrices (Q, R) in its optimization objective function and adjusts the priority of constraints such as yaw stability, path tracking accuracy, and motor torque limit (for example, if the yaw rate deviation is too large, increase the weight of the yaw stability term in the Q matrix).

[0191] The reinforcement learning module (RL) updates the action output parameters of the policy network (such as torque distribution coefficient α and path curvature adjustment amount κ) and reward function weights (for example, if the motor temperature rises too quickly, increase the weight of the "motor temperature rise" penalty term), making RL more adaptable to the fault handling needs of actual working conditions.

[0192] The transfer learning module in the feedback module is responsible for extracting historical data, specifically in the following way:

[0193] Based on Domain Adversarial Network (DANN), historical fault data is aligned with the feature distribution of the current scenario (features include fault level, slope, vehicle speed, load, motor status, etc.) to eliminate the data distribution differences under different scenarios;

[0194] Extract "strategy skeletons" (such as torque distribution logic under similar level 1 faults and path degradation schemes under similar level 3 faults) from the historical fault database as reference templates for the current control strategy.

[0195] System performance optimization is achieved through a dual approach of "real-time calibration + historical migration":

[0196] Digital twin calibration (real-time optimization): The digital twin calibration module uses the error between the virtual model and the physical model to dynamically adjust the weights and constraints of MPC, making the control strategy more in line with the actual dynamic characteristics of the physical system, and improving yaw stability, path tracking accuracy and torque distribution efficiency under fault conditions.

[0197] Transfer learning (cross-scenario optimization): The transfer learning module fine-tunes the RL policy network based on extracted historical similar policies, reducing the exploration cost in new fault scenarios and accelerating fault response speed (such as quickly matching torque allocation schemes for similar faults). At the same time, it iteratively optimizes the reward function through historical data to balance the relationship between "efficiency-safety-energy consumption" (such as reducing unnecessary speed reductions and improving transfer efficiency while ensuring stability).

[0198] Ultimately, through a continuous feedback-update-learning cycle, the system's fault tolerance, adaptability, and operational efficiency are gradually optimized.

[0199] Example 2

[0200] This embodiment is described in conjunction with Embodiment 1. The fault containment system for implementing the fault containment method of a five-axis heavy-duty AGV described in this embodiment includes a sensor group, a fault level judgment module, a control command generation module, and a feedback module.

[0201] The sensor group is used to acquire the status data of the five-axis heavy-duty AGV in real time.

[0202] The fault level judgment module is used to calculate dynamic fault factors based on status data, and to determine the fault level of the motor based on the dynamic fault factors and the operating condition weight factor.

[0203] The control command generation module is used to generate dynamic control commands according to a preset control strategy.

[0204] The feedback module is used to compare the simulation results with the actual situation, and then feed the comparison results back to the control command generation module, thereby optimizing the dynamic control commands.

[0205] like Figure 1 As shown, the five-axis heavy-duty AGV includes a front section, a rear section, and a power unit. The front section includes a first axle 1 and a second axle 2, the rear section includes a third axle 3 and a fourth axle 4, and the power unit includes a fifth axle 5. This constitutes a five-axis, ten-wheel structure, and each wheel uses a hub motor with independently controlled torque. The first four axles are fixed-direction wheels, and each wheel is driven independently by a hub motor.

[0206] The fifth axle, 5, is the rear axle, equipped with a steering motor 7, which drives the rear wheels 8 to achieve steering angle. adjust;

[0207] The clamping mechanism 9 is symmetrically arranged on both sides of the AGV centerline. When clamping the load, the center of mass is located on the centerline, and the longitudinal position changes dynamically.

[0208] The front and rear panels have 8 wheels corresponding to 16 hub motors. Two hub motors are arranged side by side at the position of each wheel on the front and rear panels.

[0209] I. Addressing the power gap in "high load + complex terrain"

[0210] Roll-on / roll-off vehicles need to meet stringent requirements for load capacity, climbing ability, and driving speed; the power / torque of a single hub motor is often insufficient to cover extreme operating conditions.

[0211] Load stacking: The total weight of the vehicle and cargo exceeds 3.5 tons. The dual motors can output power in parallel (equivalent to "doubling the power") to make up for the "insufficient power" under heavy load and ensure continuous torque when climbing a 10° slope (avoiding "getting stuck on the slope" due to insufficient power).

[0212] Dynamic adaptation: The ramp of the roll-on / roll-off ship has a slope angle of 4°-8°, and the parking lot has uneven road surface (vertical overload of 1g). The dual motors can independently adjust the torque (similar to "intelligent differential"), so that the wheels still have power output when "slipping / weightless" (for example, if one wheel is suspended in the air, the other motor can still drive it, avoiding the whole vehicle from stopping).

[0213] II. Constructing "Redundant and Reliable" Operational Support

[0214] Ro-Ro terminals require 24-hour uninterrupted operation (loading and unloading over 1280 vehicles per day), and the failure of any single component could cause the entire process to stop. Dual motors are a typical example of "redundant design".

[0215] Fault tolerance: If a single motor fails due to "prolonged high load and harsh environment (intense heat inside the ship / exhaust gas)," another motor can temporarily take over (maintain basic power output), avoiding "the entire vehicle being grounded and blocking the passage due to the failure of one motor";

[0216] Extended lifespan: The dual motors can work alternately or in coordination (single motor operates under light load, and dual motors work together under heavy load), reducing the wear rate of a single motor and indirectly improving the stability and economy of the equipment throughout its entire life cycle (reducing maintenance frequency and extending replacement cycle).

[0217] III. Adapting to the intelligent requirements of "precise control + kinetic energy recovery"

[0218] Intelligent transfer systems prioritize automation and low energy consumption, with dual motors offering unique value in terms of precise control and energy efficiency optimization.

[0219] Torque Refinement: The dual motors can precisely distribute the torque of each wheel through independent control algorithms (for example, when turning, the inner motor reduces torque and the outer motor increases torque, assisting the power unit to achieve "flexible steering coordination"), allowing "non-steering wheels" to cooperate with the power unit and more flexibly adapt to "turning on the spot and narrow road maneuvering";

[0220] Enhanced kinetic energy recovery: During braking / downhill driving, both motors can simultaneously switch to power generation mode (equivalent to "dual generators"), recovering more energy than a single motor (improving energy recovery efficiency by 20%-30%). Combined with automatic battery swapping, this extends the vehicle's range (meeting the requirement of "3 hours+ continuous operation").

[0221] Faced with "heavy loads and complex terrain", dual motors fill the power gap;

[0222] Facing the challenge of "24 / 7 operation and zero fault tolerance," dual motors form a reliable barrier.

[0223] To meet the demands for "intelligent and low-energy consumption", dual motors enable precise control and energy efficiency optimization.

[0224] This is precisely the core logic of "automated transfer" at roll-on / roll-off terminals—using the "redundancy" of hardware design to ensure the "smoothness" of the operational process.

[0225] like Figure 2As shown, sensor modules 1-5 are current, encoder, wheel speed, vibration, and temperature sensors, respectively, used to monitor the motor's current, speed, torque, linear speed, vibration, and temperature rise. Controller 6 receives the sensor inputs and determines the fault level. Based on the fault level, the MPC and RL redistribute the motor torque or execute a three-level safety chain.

[0226] The front and rear wafers exchange data wirelessly. The exchanged data includes the position and motion status data of the front and rear wafers, control commands and action synchronization data, power and energy status data, sensor and environmental perception data, and fault diagnosis and status monitoring data. The purpose of the data exchange is to ensure that the two wafers remain coordinated in the separated or combined state, integrate the environmental perception information of the front and rear wafers, share fault information in real time, improve the robot's navigation accuracy and obstacle avoidance ability, and facilitate the rapid location of problems by the mobile control console or scheduling system.

Claims

1. A fault containment method for a five-axis heavy-duty AGV, characterized in that, Includes the following steps: Step 1: Acquire the status data of the five-axis heavy-duty AGV in real time through the sensor group, obtain the working status data of the hub motor of each wheel, and calculate the dynamic fault factor h based on the acquired working status data; The dynamic fault factor is calculated based on the acquired working status data. The specific process is as follows: ； in For current, Let A be the rotational speed and A be the vibration. For temperature; h is the dynamic fault factor, i is the speed weighting factor, j is the vibration weighting factor, and k is the temperature weighting factor. Step 2: Calculate the working condition weighting factor w based on the current working conditions, and determine the fault level of each wheel hub motor based on the dynamic fault factor h obtained in Step 1. In step 2, the specific steps for determining the fault level of the motor are as follows: Step 21: Calculate the operating condition weighting factor : ； Where a is the gradient weight factor, b is the vehicle speed weight factor, c is the load weight factor, S is the gradient coefficient, V is the vehicle speed coefficient, and β is the load coefficient. Step 22: Based on dynamic fault factors With operating condition weighting factor Obtain the fault level of the motor; like This means that one motor of a single wheel in the front or rear section fails, which is considered a Level 1 fault. like This means that either a single-wheel dual-motor failure occurs completely in a single area or a single-wheel dual-motor failure occurs in each of two areas, which is considered a Level 2 fault. like This means that two wheels on opposite sides of the same area fail, or two wheels on opposite sides of each of two areas fail, which is considered a level 3 fault. like This means that there are two motors on the same side of a certain area that have failed, or at least three motors in a certain area that have failed, or a power unit motor that has failed. This is considered a level 4 fault. If the packet loss rate L between the front and rear chips is ≥30%, indicating sensor failure, it is considered a level 5 fault. , , , These are the judgment thresholds corresponding to fault levels 1 to 4, respectively. Step 3: Generate dynamic control commands through the control command generation module to provide input commands for fault response and control strategies for dynamic fault tolerance; Step 4: Fault recovery and safety chain execution. Based on the fault level and control strategy, the torque of the motor in the non-faulty wheel hub is redistributed, and the system status is fed back to the control command generation module in real time to adjust the control strategy. In step 4, the specific steps of the control strategy for level 4 and level 5 faults are as follows: A three-tiered safety chain of "communication emergency response - combined braking - mechanical locking" is adopted to ensure equipment safety under extreme failure conditions; Emergency Communication: Switch to the backup channel, which is a LoRa channel, used to broadcast fault codes to the central control system; Combined braking: Redundancy activation of electric motor braking combined with hydraulic braking; Mechanical locking: Dual electromagnetic locks lock synchronously, and the optimal locking point is predicted by the feedback module to prevent slippage; Step 5: The feedback module compares the virtual model with the physical model and then feeds back to the control command generation module, while extracting historical data to continuously optimize system performance. In step 5, the feedback module compares the virtual model with the physical model using the following specific steps: Step 51: The feedback module includes a digital twin calibration module and a transfer learning module; the digital twin calibration module optimizes and enhances the model prediction control module based on the virtual model, compares the output of the virtual model with that of the physical system, and adjusts the weight of the model prediction control module in real time based on the error value obtained from the comparison. Step 52: The transfer learning module aligns the distribution of historical faults with the current scene features based on the domain adversarial network; it extracts similar fault policy skeletons from the historical database, thereby enabling fine-tuning of the reinforcement learning module network.

2. The fault containment method for a five-axis heavy-duty AGV according to claim 1, characterized in that, In step 1, the sensor group includes a current sensor, an encoder, a wheel speed sensor, a vibration sensor, and / or a temperature sensor; The status data includes motor operating current, motor speed, wheel linear speed, mechanical vibration, and motor temperature.

3. The fault containment method for a five-axis heavy-duty AGV according to claim 1, characterized in that, In step 4, the specific steps of the control strategy for level 1 faults are as follows: First, the faulty motor is isolated using hardware redundancy within the wheel to prevent the failure from propagating; the control command generation module includes a model prediction control module and a reinforcement learning module. The dual in-wheel motors are switched to the healthy motor. The healthy motor is monitored by a temperature sensor. The model predictive control module adjusts the additional yaw torque to compensate for the torque shortfall of the failed motor. The reinforcement learning module balances efficiency and temperature rise, thereby limiting the temperature of the healthy motor. The faulty motor enters "hot standby mode," maintaining only a portion of the current for status monitoring.

4. The fault containment method for a five-axis heavy-duty AGV according to claim 1, characterized in that, In step 4, the specific steps of the control strategy for level 2 faults are as follows: The model prediction control module tightens the yaw rate error constraint to control vehicle stability; The reinforcement learning module uses a "torque increase on the same side + torque decrease on the opposite side" strategy to increase the weight of "yaw stability" in the reward function and output actions to make the healthy motor on the same side fully loaded, thereby increasing the path curvature.

5. The fault containment method for a five-axis heavy-duty AGV according to claim 1, characterized in that, In step 4, the specific steps of the control strategy for level 3 faults are as follows: The torque is evenly distributed to the other working wheels where the dual motors have not failed. The model prediction control module keeps the torque of all working wheels equal and minimizes it to avoid unwanted yaw torque. At the same time, the reinforcement learning module forces the electronic differential lock to lock the normal working wheels, replans the path downgrade, and selects low curvature road sections with curvature less than the preset curvature and road sections without slopes.

6. A fault-containment system for implementing a fault-containment method for a five-axis heavy-duty AGV as described in any one of claims 1-5, characterized in that, It includes a sensor group, a fault level judgment module, a control command generation module, and a feedback module; The sensor group is used to acquire the status data of the five-axis heavy-duty AGV in real time. The fault level judgment module is used to calculate dynamic fault factors based on status data, and to determine the fault level of the motor based on the dynamic fault factors and the operating condition weight factor. The control command generation module is used to generate dynamic control commands according to a preset control strategy. The feedback module is used to compare the simulation results with the actual situation, and then feed the comparison results back to the control command generation module, thereby optimizing the dynamic control commands.

7. The fault containment system of the fault containment method for a five-axis heavy-duty AGV according to claim 6, characterized in that, The five-axis heavy-duty AGV includes a front section, a rear section, and a power unit. The front section includes a first axis and a second axis, the rear section includes a third axis and a fourth axis, and the power unit includes a fifth axis, thus forming a five-axis, ten-wheel structure. Each wheel uses a hub motor with independently controlled torque. The front section and the rear section exchange data via wireless communication.

Images