An autonomous analog control and anti-disturbance method for intelligent shuttle vehicles in cold storage

Abstract

The application discloses a kind of self-determination quasi-state control and anti-disturbance method of intelligent shuttle vehicle of cold storage.The method constructs environment space gradient field by distributed optical fiber sensing network, and inverts heat flow disturbance to update dynamic environment model;Synchronous analysis driving current harmonic to identify ground phase change state, and utilize digital twin model to compensate mechanism hysteresis effect;Adopt reflection-adaptation-learning three-layer decision architecture: reflection layer millisecond level responds to burst risk;Adaptation layer adjusts motion parameters for seconds;Learning layer minutes level planning quasi-state navigation path;Finally, through model predictive control, the instructions of each layer are fused, and the output of collaborative control signal is obtained.The application enables the shuttle vehicle to actively perceive, predict and adapt to the non-uniform dynamic environment of the cold storage, improving the operation safety, positioning accuracy and work efficiency in low temperature, slippery and multi-disturbance working conditions.

Description

Technical Field

[0001] This invention belongs to the field of automated warehousing and logistics technology, and in particular, it is a method for autonomous mimicry control and anti-disturbance of intelligent shuttle vehicles in cold storage. Background Technology

[0002] As a core node in modern cold chain logistics, the level of automation within cold storage facilities directly impacts warehousing efficiency and operating costs. Shuttle systems, due to their high-density storage and flexible scheduling capabilities, have become the mainstream solution for automated cold storage. However, the extreme and dynamically changing environment of cold storage poses a significant challenge to the stable operation of shuttles, and existing technologies still exhibit significant bottlenecks in environmental adaptability, operational safety, and operational accuracy.

[0003] First, the cold storage environment is not uniform and steady-state. The uneven airflow from the refrigeration equipment, the heat exchange caused by the opening of the storage doors, and the local heat sources generated by the storage of goods together form a complex three-dimensional temperature and airflow gradient field. Existing shuttle vehicles typically rely on a limited number of point-based temperature and humidity sensors for environmental perception. This discrete measurement cannot capture continuous gradient distributions and microscopic disturbances, causing the control system to treat the environment as a "static background" or only respond to large-scale, low-frequency macroscopic changes. It lacks the ability to perceive and respond to gradual effects such as nonlinear material deformation and sensor drift caused by gradients.

[0004] Secondly, the low temperature and high humidity environment triggers a series of unique physical interaction problems. A thin layer of ice on the ground may undergo a solid-to-quasi-liquid phase transition under the coupling of wheel pressure and temperature, leading to a sudden dynamic change in the adhesion coefficient. Traditional anti-skid control based on wheel speed difference or vehicle posture feedback has a lag and is difficult to prevent slippage. Simultaneously, low temperatures cause the lubricant in mechanical components to become viscous and the elastic modulus of polymer materials to change, resulting in hysteretic nonlinearities in motion mechanisms (such as fork lifting and steering) that are related to temperature and motion history. This severely affects repeatability accuracy, and existing calibration methods are mostly offline static compensation, which cannot adapt to the dynamic characteristic drift during operation.

[0005] Furthermore, the existing control architecture of shuttle vehicles mostly follows a linear paradigm of "perception-planning-execution," with each module handling environmental disturbances in a relatively isolated manner. For example, the navigation module plans the geometric shortest path but may traverse areas of strong airflow disturbance; the stability control module passively intervenes after detecting slippage but cannot proactively adjust torque distribution. This responsive control strategy lacks a holistic understanding and collaborative optimization capability of the "environment-vehicle" coupled system, making it rigid and inefficient in dealing with sudden and variable disturbances in cold storage.

[0006] In summary, current cold storage shuttle technology fails to fully address the fundamental impact of the spatiotemporal heterogeneity of the environment. At the perception level, it lacks refined measurements of microscopic dynamics such as gradient fields and phase transition interfaces; at the control level, it lacks the ability to autonomously adapt and make collaborative decisions by integrating multi-source forward-looking information. Therefore, there is an urgent need to develop a new intelligent control system capable of deeply understanding and proactively adapting to the complex dynamic environment of cold storage, in order to achieve highly reliable and high-precision autonomous operation of the shuttle under extreme conditions. Summary of the Invention

[0007] Purpose of the invention: In view of the above-mentioned problems in the prior art, this application provides an autonomous mimicry control and anti-disturbance method for a cold storage intelligent shuttle vehicle.

[0008] Technical solution: An autonomous mimicry control and disturbance rejection method for an intelligent shuttle vehicle in cold storage, comprising:

[0009] Collect microscale physical quantity data of the vehicle-environment interface, and construct the environmental spatiotemporal gradient field through unstructured grid gradient calculation and spatiotemporal interpolation;

[0010] Based on the environmental spatiotemporal gradient field, the environmental dynamic embedded model, which includes risk feature prediction, is dynamically updated by solving the inverse problem of heat conduction and fusing multi-source data.

[0011] Collect native state data of the drive system, and identify the phase change state of the wheel-ground contact interface in real time by analyzing motor current harmonics and fusing multi-modal features.

[0012] The hysteresis nonlinear digital twin model is updated online based on the historical trajectory deviation data of the motion mechanism, and the model is used to generate feedforward compensation instructions.

[0013] By integrating the dynamic embedded model of the environment, phase transition state, and feedforward compensation instructions, a collaborative control command is generated through a three-layer decision-making architecture of reflection-adaptation-learning, which drives the shuttle to perform mimicry navigation and anti-disturbance operation.

[0014] According to a further improvement of the present invention, constructing an environmental spatiotemporal gradient field includes:

[0015] The raw wavelength data of each node in the distributed fiber Bragg grating sensor network is demodulated and physical quantity is calculated to generate a multi-node instantaneous snapshot data table of physical quantities.

[0016] Spatiotemporal alignment and outlier removal are performed on the data in the instantaneous snapshot table of physical quantities to generate clean physical quantity field mesh data;

[0017] Based on the theory of unstructured triangulated mesh and linear shape function, the temperature gradient vector and strain gradient tensor at the center of each mesh element are calculated to generate gradient vector data of unstructured mesh elements.

[0018] Anisotropic diffusion filtering and node tensor reconstruction are performed on the unit gradient vector data, and high-resolution spatiotemporal gradient field map data covering the vehicle body and near-field space are generated by Kriging space interpolation.

[0019] According to a further improvement of the present invention, dynamically updating the environmental dynamic embedded model that includes risk feature prediction includes:

[0020] Using the temperature field in the high-resolution spatiotemporal gradient field map as the boundary condition, combined with the parameterized vehicle body thermal model, a heat flux inversion optimization problem is constructed and solved using the adjoint method to invert the heat flux density vector distribution data on the vehicle body surface.

[0021] By integrating heat flux density vector distribution data and strain gradient data, a pre-defined rule engine is used to identify and label high heat flux abrupt change zones, thermo-mechanical coupling risk zones, and potential phase change interfaces, generating a list of high-risk environmental features.

[0022] Based on the current high-risk feature list data and historical gradient field sequences, the environmental dynamic embedded model is updated using a data assimilation algorithm.

[0023] According to a further improvement of the present invention, real-time identification of the phase transition state of the wheel-ground contact interface includes:

[0024] The three-phase high-frequency current waveforms, rotor positions, and vehicle vibration data of each drive motor are collected simultaneously to form the original state data stream of the drive system.

[0025] Short-time Fourier transform is performed on the current waveform data to extract the amplitude and phase characteristics of fractional harmonics and sideband frequency components that are not integer multiples of the fundamental frequency, thus forming current nonlinear harmonic characteristic spectrum data.

[0026] Wavelet packet decomposition was performed on the vehicle vibration data to obtain vibration characteristic data;

[0027] Harmonic characteristic spectrum data, vibration characteristic data, local ground temperature data, and wheel dynamic load data are fused into a feature vector, which is then input into an online ensemble learning classifier for real-time inference. After probability calibration, the output is an interface phase probability distribution vector.

[0028] According to a further improvement of the present invention, a feedforward compensation instruction is generated, comprising:

[0029] Record the actual trajectory feedback after the motion mechanism executes the theoretical command trajectory, and calculate a high-fidelity trajectory tracking error profile dataset.

[0030] The error profile dataset, along with the corresponding theoretical command sequence and ambient temperature data, is input into the hysteresis nonlinear digital twin model. The online parameter identification algorithm is used to update the model parameters, resulting in the updated hysteresis model parameter vector.

[0031] The new motion command data to be executed is input into the updated digital twin model for forward simulation to predict its execution deviation, and a feedforward compensation torque / current command sequence is generated by calculating through the mechanism inverse dynamics model.

[0032] The feedforward compensation torque / current command sequence is low-pass filtered and safety-limited, and then superimposed with the new motion command data in the command domain to generate the pre-distortion optimized final execution command data.

[0033] According to a further improvement of the present invention, generating cooperative control instructions includes:

[0034] The system continuously monitors the highest-level alarms in the list of high-risk environmental features and the instability transition signals in the interface phase probability distribution vector. It then directly matches and outputs the basic-level emergency correction command data through the hardware logic rule base.

[0035] Based on the updated dynamic environment model data, short-term predictions, and continuous interface phase probability distribution vectors, dynamic chassis parameter tuning data for adjusting torque distribution, steering sensitivity, and suspension parameters are dynamically calculated and output.

[0036] A multi-level environmental cost map is constructed based on the updated dynamic environment model data. The path with the lowest comprehensive cost is searched through a spatiotemporal three-dimensional algorithm, and the speed curve is optimized to generate global mimicry navigation flow data.

[0037] Based on emergency correction command data, dynamic chassis parameter adjustment data, final execution command data, and global mimicry navigation flow data, the model predictive controller performs conflict arbitration and multi-objective optimization to solve and output the final cooperative control command data package.

[0038] According to a further improvement of the present invention, global mimicry navigation flow data is generated, including:

[0039] The gradient intensity, risk area distribution, static obstacles, and task semantic information in the updated dynamic environment model data are mapped into multi-level cost maps such as thermal disturbance cost layer and basic access layer, and multi-level fused environment cost map is generated through weighted fusion.

[0040] On the integrated environment cost map, a runtime three-dimensional search algorithm is used to find the initial spatiotemporal path sequence with the optimal cumulative cost in both spatial and temporal dimensions from the starting point to the end point.

[0041] The initial spatiotemporal path sequence is smoothed using spline curves, and the smoothed velocity-time curves based on vehicle dynamics constraints and path curvature planning are encapsulated to form the final global mimicry navigation flow data.

[0042] According to a further improvement of the present invention, the adjoint method is used to solve the heat flow inversion optimization problem, including:

[0043] A framework for the inversion problem is constructed, which takes the heat flux density distribution on the vehicle surface as the optimization variable, the minimum difference between the temperature predicted by the forward thermal model and the observed temperature as the objective function, and includes a Tikhonov regularization term.

[0044] The process begins with an initial heat flux hypothesis and iterates. In each iteration, a forward thermal calculation is performed first to obtain the predicted temperature field, followed by an adjoint calculation to efficiently obtain the gradient of the objective function with respect to all heat flux parameters.

[0045] The obtained gradient is used to update the heat flux distribution using the L-BFGS optimization algorithm;

[0046] The optimal regularization parameter is determined based on the L-curve method, and the solution under this parameter is used as the final heat flux density distribution data. At the same time, the variance of the main components of the solution is calculated to quantify the uncertainty.

[0047] According to a further improvement of the present invention, an online parameter identification algorithm is used to update the model parameters, including:

[0048] The parameter update of the hysteresis nonlinear digital twin model is constructed as a recursive least squares optimization problem;

[0049] The objective is to minimize the difference between the model-predicted trajectory and the actual trajectory inferred from the superposition of error contour data with theoretical command sequences;

[0050] After each motion action is completed, the model parameter vector and its covariance matrix are recursively updated using the newly acquired error data to make the model dynamically track the changes in hysteresis characteristics caused by low temperature and wear.

[0051] Beneficial effects: This invention transforms environmental disturbances from passive resistance into understandable and predictable cooperative variables for shuttle vehicles, achieving autonomous adaptation based on deep environmental understanding. This technology simultaneously enhances the operational safety, stability, and operational precision of vehicles in complex cold storage environments. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the overall technical solution of the present invention.

[0053] Figure 2 A flowchart for constructing the environmental spatiotemporal gradient field for this invention.

[0054] Figure 3 This is a flowchart illustrating the dynamic updating of the environmental dynamic embedded model that includes risk feature prediction, as described in this invention.

[0055] Figure 4 This is a flowchart illustrating the real-time identification of the phase transition state of the wheel-ground contact interface according to the present invention.

[0056] Figure 5 This is a flowchart illustrating the generation of feedforward compensation instructions for this invention.

[0057] Figure 6 This is a flowchart of the process for generating collaborative control instructions according to the present invention. Detailed Implementation

[0058] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0059] Example 1: This example provides an overall framework for an autonomous mimicry control and disturbance rejection method for an intelligent shuttle vehicle in cold storage. Specifically, the method in this example includes the following steps:

[0060] S1: Acquisition and Preliminary Construction of Environmental Spatiotemporal Gradient Field Data

[0061] According to one aspect of this application, this embodiment abandons the traditional approach of relying on a small number of discrete point sensors to obtain general environmental information. Instead, it performs high-resolution, full-field gradient measurements of the microenvironment surrounding the shuttle vehicle, thereby providing a detailed environmental topographic map for subsequent intelligent decision-making. Specifically, this embodiment does not simply obtain the absolute values ​​of temperature or humidity, but focuses on capturing the rate of change of these physical quantities in space (i.e., gradient) and their dynamic characteristics of evolution over time. This is because gradient is the fundamental physical driving force that drives heat flow, causes non-uniform material deformation, and triggers local anomalies (such as condensation and strong convection). In this embodiment, a highly integrated distributed fiber Bragg grating sensor network is deployed. This network consists of hundreds of micron-sized fiber Bragg grating sensor nodes. These nodes are not uniformly or randomly arranged, but are preferentially deployed at geometric abrupt changes in the vehicle body, critical sections of the airflow path, and stress concentration areas of the mechanical structure, such as the fork root, the inner side of the wheel guards, the surface of the battery compartment, and the windward and leeward sides of the vehicle top, based on computational fluid dynamics simulation results. In a preferred embodiment, the sensors are combined in the form of arrays and clusters, which can sense both large-scale trends and focus on local hotspots.

[0062] Furthermore, the sensor network operates continuously in the low-temperature environment of the cold storage. Each sensor node collects raw optical wavelength shift data streams caused by temperature or strain changes in the grating reflection center wavelength, at a sampling frequency of no less than 100Hz. This raw data is transmitted to the onboard signal processing unit via low-temperature resistant optical fiber. Subsequently, the system demodulates and processes the raw data. Specifically, based on the pre-calibrated "wavelength-temperature" and "wavelength-strain" relationship curves of each sensor in a high-low temperature test chamber before leaving the factory, the processing unit converts the wavelength shift into corresponding instantaneous physical quantity data, namely, the specific coordinates of each sensor node in three-dimensional space, the precise temperature value (unit: °C) and micro-strain value (unit: με) sensed at the current millisecond. Thus, an instantaneous physical state snapshot of hundreds or thousands of "tactile points" on the vehicle surface is obtained.

[0063] However, snapshots of discrete points are insufficient to describe the environmental field. Therefore, according to a further improvement in this embodiment, these discrete instantaneous physical quantity values ​​are injected into a pre-defined spatial grid model that precisely matches the three-dimensional digital model of the vehicle body. Based on this grid model, the algorithm uses spatial interpolation techniques and finite difference methods to calculate the changes in physical quantities between adjacent sensor nodes. For example, for the temperature field, the algorithm calculates the temperature change per unit distance (e.g., millimeters) in the X, Y, and Z directions, thereby generating vector-based temperature gradient data covering the vehicle surface and the space near the vehicle body; for the strain field, it generates more complex tensor data. Finally, these gradient data are integrated and visualized to output a high-resolution environmental gradient field map. This map not only includes the magnitude (intensity) of the gradient, but more importantly, it includes the direction of the gradient. In terms of form, it is a multi-dimensional data matrix or point cloud that updates over time, with a spatial resolution reaching the centimeter level, thus achieving a precise digital depiction of the invisible microscopic heterogeneous structures such as "airflow gullies" and "temperature cliffs" within the cold storage.

[0064] S2: Dynamic Environment Model Update Based on Physical Inversion and Data Fusion

[0065] After successfully acquiring detailed environmental gradient field data, the dynamic environmental model update step begins. This step goes beyond simple data presentation, shifting towards a mechanistic understanding of the physical processes. The superficial gradient data obtained in the previous step is transformed into a dynamic cognitive model that reveals what kind of disturbances are actively applied to the environment and possesses a certain predictive capability through reverse deduction of physical laws. For example, the temperature gradient itself is a result, but this step seeks to understand the cause of this gradient distribution—that is, how heat flow is distributed on the vehicle surface. To this end, the output high-resolution environmental gradient field data is used as known boundary conditions, and an inverse problem solver based on unsteady-state heat conduction partial differential equations is introduced.

[0066] Specifically, the solver, based on the vehicle's surface geometry and material thermal properties, employs optimization algorithms such as the conjugate gradient method and Tikhonov regularization to iteratively calculate and find the heat flux input most likely to generate the currently observed temperature gradient distribution at the boundary. Through this complex calculation process, the system successfully traces the cause (heat flux) from the effect (gradient) back to the cause, retrieving the vector distribution of heat flux density acting on each surface element of the vehicle at the current moment. This data is crucial because it directly quantifies the active heating or cooling power of the environment on the vehicle and is a direct indicator for assessing the intensity of environmental thermal disturbances. For example, at the instant the cold storage door is opened, the system can immediately retrieve the high heat flux impact zone formed on the right side of the vehicle by the incoming hot air, whose intensity far exceeds the background cooling load.

[0067] Furthermore, to form a more comprehensive environmental threat assessment, this embodiment performs multi-source data fusion. The system performs spatiotemporal alignment and correlation analysis on the inverted heat flux density vector distribution data and strain gradient field data. In an optional implementation, the algorithm sets a series of adaptive thresholds: for example, when the heat flux density amplitude of a certain region exceeds three standard deviations of the average background value, and the tensile strain gradient of that region also shows an abnormal peak, the algorithm marks that region as a "high heat-high stress coupling risk zone". Similarly, the system can also identify "strong convective shear zones" (based on abrupt changes in heat flux direction in a specific pattern) and "potential condensation zones" (based on the localized appearance of small positive heat flux under low temperature background). All these identified regions, along with their type labels, spatial boundaries, risk levels, and key parameters, are integrated into a structured environmental risk characteristic list.

[0068] Finally, the system performs model assimilation and updating, inputting the current gradient field map, inverted heat flow data, and risk feature list, along with historical sequence data from the past tens or even minutes, into the embedded environmental dynamic model. This model is a simplified digital twin environment. Through data assimilation techniques such as Kalman filtering or particle filtering, it continuously corrects its internal state variables, enabling the model not only to remember past environmental disturbance patterns but also to extrapolate the evolution trend of the environmental field in the next few seconds based on physical laws. The updated dynamic environmental model data it outputs is a comprehensive data package containing current state estimates, short-term future predictions, and uncertainty assessments, providing a forward-looking environmental cognitive foundation for subsequent real-time decision-making.

[0069] S3: Real-time online identification of phase transition state at the ground interface

[0070] According to another aspect of this application, a real-time online identification method for the phase transition state of the ground interface is proposed to solve the highly nonlinear stick-slip dynamics problem caused by the thin ice or water film between the wheel and the cold storage floor, which is imperceptible to the naked eye. Traditional anti-skid strategies rely on macroscopic hysteresis information such as wheel speed difference or vehicle posture, which is a post-event remedy. However, this embodiment realizes pre-event warning and in-event diagnosis, that is, without adding additional contact sensors, it can understand the physical phase change of the microscopic contact interface between the tire and the ground by interpreting the motor current and vibration signals of the drive system itself.

[0071] In a preferred embodiment, the system synchronously acquires raw data of the three-phase high-frequency current waveforms of all drive wheels' motors, high-precision motor rotor position / speed feedback data, and triaxial high-frequency vibration acceleration data mounted on the axle or vehicle frame, at a sampling rate of no less than 2kHz. These data collectively constitute the raw sensory signal stream for perceiving the state of the contact interface. Among these, the motor current signal is considered the core information source because any minute slippage, shearing, or adhesive force change occurring on the tire-ground contact patch will immediately be transmitted back to the motor through the transmission system, causing subtle fluctuations in the motor's output torque, which will ultimately be reflected in changes in the harmonic components of the current waveform.

[0072] Specifically, online spectrum analysis is performed on the raw data of the collected three-phase high-frequency current waveforms of the motor. Traditional fault diagnosis only focuses on harmonics that are integer multiples of the fundamental frequency, while this embodiment focuses on those non-integer multiples, fractional harmonic components, such as the 0.5, 1.5, and 2.5 harmonics. The appearance and enhancement of these fractional harmonics are typical frequency domain characteristics of phenomena such as nonlinear dry friction and stick-slip oscillations in the system. The algorithm extracts the amplitude, phase, and modes of these specific fractional harmonics as a function of motor speed in real time, forming a current nonlinear harmonic characteristic spectrum data. At the same time, the high-frequency vibration acceleration data, after bandpass filtering and envelope demodulation, is used to extract the characteristic frequency energy related to wheel tangential slip and vertical runout, generating vibration characteristic energy data.

[0073] Furthermore, to achieve accurate state classification, a lightweight deep learning model (such as a one-dimensional convolutional neural network) is introduced. The model's input is a feature vector incorporating multi-dimensional information, including but not limited to: the current nonlinear harmonic characteristic spectrum data, vibration characteristic energy data, local ground temperature gradient data for the area directly beneath the wheel, and the wheel's current dynamic load estimation data. In the offline phase, the model is trained using massive amounts of data collected in the laboratory simulating different ice surface conditions (dry ice, wet ice, thin water film, mixed state), learning to identify the physical phase state of the microscopic contact interface from these complex signals. During online operation, the model independently performs millisecond-level inference for each drive wheel, outputting an interface phase probability vector. This vector probabilistically quantifies the likelihood of the current contact interface being in different states such as "stable solid friction," "quasi-liquid lubrication film formation," "stick-slip oscillation initiation," or "macroscopic slippage occurrence." This data provides the control system with a real-time, fine-grained estimate of the ground adhesion coefficient, serving as a sensory input for achieving precise torque control and preventing instability.

[0074] S4: Digital twin compensation for hysteresis in motor mechanisms

[0075] To address the deep-seated problem of hysteresis and reduced repeatability accuracy in motion mechanisms caused by viscoelastic changes and lubricant solidification in materials such as metals and polymers at low temperatures, this embodiment designs a digital twin compensation method for the hysteresis effect of motion mechanisms. Instead of creating a separate static compensation table for each mechanical error source, this embodiment constructs a dynamic digital twin model for the entire motion execution chain (such as the fork lifting mechanism and the steering servo mechanism), enabling online learning and adaptive evolution. This uses the predictability of software to overcome the non-repeatability of physical processes.

[0076] In one alternative implementation, the system initiates the learning loop for this step whenever the shuttle completes a specific motion command, such as "lifting the forks from height A to height B". First, it reads the timing data of the theoretical command trajectory issued for that motion from the motion controller. This data fully records the position, speed, and current value that the controller expects the motor to reach along the entire motion timeline. Simultaneously, the system collects actual execution trajectory feedback data from a high-precision absolute encoder or miniature laser rangefinder installed at the end of the actuator. Through precise timestamp alignment and coordinate transformation, the algorithm calculates the deviation between the actual position and the theoretical position at each sampling moment throughout the entire motion, thereby generating a detailed trajectory following error profile. This error data not only includes the final steady-state error but, more importantly, includes dynamic error components related to speed and acceleration during the ascent, descent, and hold phases.

[0077] Subsequently, this trajectory-following error profile data, along with the theoretical command trajectory timing data that triggered the action, and the local average temperature data of the mechanism during the action execution, are fed into a parameterized hysteresis nonlinear mathematical model. In a preferred embodiment, this model employs a physical foundation model, such as the Prandtl-Ishlinskii model or the Bouc-Wen model, which can effectively describe rate-dependent and temperature-dependent hysteresis phenomena. The system invokes an online parameter identification algorithm (such as recursive least squares) to dynamically adjust key parameters in the model, aiming to minimize the difference between the model prediction error and the actual trajectory-following error profile data. These parameters include, for example, the Coulomb damping coefficient representing friction, the weighting function representing elastic aftereffects, and the relaxation time constant, which is strongly correlated with ambient temperature. This process outputs updated hysteresis model parameter set data, enabling the digital twin model to more accurately mimic the state of the real physical mechanism under the current temperature and mechanical wear conditions.

[0078] According to a further improvement of this embodiment, the digital twin model is not only used for historical analysis, but its main value lies in feedforward prediction. When the control system needs to issue a new motion command, before sending the command to the physical motor, the pre-release data of the new command and the current local average temperature data of the mechanism are input into the hysteresis digital twin model that has completed parameter updates. Based on the physical laws it has learned, the model quickly simulates the execution of the command and predicts the expected trajectory deviation that the physical mechanism will produce. Then, based on this predicted deviation, the system generates a feedforward compensation correction command data of equal magnitude and opposite direction. Finally, the original pre-release data of the new command and the feedforward compensation correction command data are superimposed to form a "pre-distorted" optimized final execution command data, which is then sent to the physical actuator. In this way, the hysteresis and deformation of the physical mechanism are pre-canceled at the command level before they are actually generated, thereby achieving high repeatability positioning accuracy in variable temperature environments.

[0079] S5: Reflective Autonomous Decision-Making and Navigation Flow Generation Integrating Multi-Source Information

[0080] Having achieved comprehensive and in-depth perception and modeling of the environment, itself, and the interactive interface, this embodiment ultimately converges and transforms all information into safe, smooth, and efficient action commands through a reflective autonomous decision-making and navigation flow generation process that integrates multi-source information. This embodiment abandons the traditional, linear "perception-planning-control" pipeline, replacing it with a hierarchical, parallel "reflection-adaptation-learning" decision-making architecture inspired by biological nervous systems. This ensures that the system can optimally respond to various challenges, from sudden dangers to long-term tasks, at different time scales of milliseconds, seconds, and minutes.

[0081] Specifically, the bottom layer of the decision-making architecture is a millisecond-level reflection layer. This layer directly receives alarms about emergency risks (such as sudden high-heat flow impacts) from the structured environmental risk feature list data, as well as high-probability judgments about impending stick-slip from the interface phase probability vector data. This information is fed into a reflection rule library implemented by hardware logic or minimal code. For example, the rule stipulates that when "the macroscopic slip probability of the right front wheel > 90%", a preset reflection of "right front wheel torque to zero, left rear wheel torque increased by 30%" is immediately triggered. This layer does not perform any complex calculations, only pattern matching and command mapping. Its output is instinctive emergency correction command data that ensures survival safety, and the response latency is strictly controlled within milliseconds.

[0082] Above the reflection layer lies the second-level adaptation layer. This layer is the core of real-time control, continuously processing updated dynamic environment model data (especially short-term predictions) and continuous interface phase probability vector data. Its core task is to dynamically optimize the vehicle's driving style parameters. For example, when the model predicts that the path ahead will pass through a strong convective shear zone, the adaptation layer calculates and generates a set of dynamic chassis parameter tuning data, including: appropriately reducing vehicle speed to increase stability, increasing the differential sensitivity of the four-wheel drive system to prepare for uneven ground adhesion, and even fine-tuning the damping coefficient of the steering system to counteract crosswind effects. Simultaneously, it dynamically allocates the torque limit of each motor based on the real-time phase probability of each wheel, implementing preventative torque limiting to actively avoid slippage rather than remedying it after it occurs.

[0083] Furthermore, at the top level, the minute-level learning layer is responsible for the strategic planning of the global task. Based on dynamic environmental model data, the environment is treated as a potential field composed of heat flow, temperature, and risk areas. The path planning algorithm of this layer does not prioritize the geometric shortest path, but rather calculates the path with the lowest overall perturbation potential energy, i.e., prioritizing travel along quiet corridors with uniform temperature, stable airflow, and away from risk areas. Its output is a global mimicry navigation flow data that considers environmental friendliness, a high-level task sequence that includes recommended waypoints, velocity curves, and energy-saving strategies.

[0084] Ultimately, the central arbitration and fusion unit is responsible for coordinating the commands that may be issued from these three layers, sometimes even conflicting ones. It integrates the instinctive emergency correction command data (highest priority), dynamic chassis parameter tuning data, optimized final execution command data, and smoothed global mimicry navigation flow data. The arbitrator dynamically adjusts the weight of each layer's commands based on the real-time context (such as whether it is in an emergency or performing a high-precision docking mission). Through an optimization controller (such as a model prediction controller), it solves for the optimal solution that simultaneously satisfies multiple objectives such as safety, comfort, accuracy, and efficiency at the current moment. This final collaborative control command data package is then distributed to each actuator, driving the shuttle to operate freely and stably in the complex and harsh cold storage environment, much like an intelligent organism with conditioned reflexes, skilled expertise, and long-term planning.

[0085] Example 2: Based on Example 1, this example provides a more detailed description of the data processing flow for an autonomous mimicry control and disturbance rejection method for an intelligent shuttle vehicle in a cold storage facility. Specifically, it includes:

[0086] S1: Acquisition and preliminary construction of environmental spatiotemporal gradient field data, including:

[0087] S1.1: Collect raw sensor data from each node of the distributed sensor network.

[0088] According to one aspect of this application, this embodiment first acquires the most basic raw signals through a carefully designed and deployed distributed sensor network. Specifically, the network consists of more than 128 high-sensitivity fiber Bragg grating sensor nodes with micron-level writing cycles. These nodes are not randomly attached, but strategically integrated into key functional parts of the vehicle body based on prior simulation analysis of the airflow and temperature fields of the vehicle body in a cold storage environment. For example, a longitudinal array is deployed at a density of one node per centimeter on the upper and lower surfaces and tooth tips of the forks to monitor microscale temperature fluctuations caused by contact conduction when the forks pick up goods; sensors are distributed in a mesh pattern at the four corner pillars and top of the vehicle body to capture spatial three-dimensional temperature gradients; and dense ring monitoring clusters are deployed on the surfaces of heat source components such as the battery compartment and controller housing. All sensors are connected in series through a specially made armored optical fiber that is resistant to -60°C and bending, forming a complete network. During the operation of the shuttle, each sensor node continuously captures the center wavelength value of its grating reflection at a constant sampling frequency of 100Hz. These wavelength values ​​are transmitted in real time to the vehicle-mounted edge computing unit via optical fiber in the form of digital signals, forming a continuous stream of raw optical wavelength offset data. This data stream is the cornerstone of all subsequent environmental sensing; its high density and high frequency characteristics ensure that transient environmental disturbances can be captured.

[0089] S1.2: Solve for the physical quantity data of each node.

[0090] Upon receiving the raw data stream, the system immediately performs precise physical quantity conversion. Specifically, the onboard edge computing unit stores the unique identifier of each fiber Bragg grating sensor and its corresponding rigorously calibrated three-dimensional response surface data of "wavelength-temperature-strain". In a preferred embodiment, this calibration is performed in a professional high and low temperature humidity test chamber, covering a temperature range of -50°C to 25°C and a strain range of 0 to 1000 με, thereby ensuring the accuracy of the conversion relationship throughout the entire cold storage working area. The data processing algorithm matches the sensor ID of each incoming raw optical wavelength offset data packet and queries the corresponding calibration surface. Through a three-dimensional interpolation algorithm, the current wavelength offset is calculated into two core physical quantities: the absolute temperature value (unit: °C) at the node and the micro-strain value (unit: με) at the node. This process is performed in parallel and in real time. For example, for a sensor located on the right front wheel arch guard, the algorithm might output "-23.7℃, 12.5με", indicating that the temperature at that point is -23.7 degrees Celsius and it is subjected to a tensile strain of 12.5 με. Finally, the system aggregates the solution results from all nodes at the same timestamp, generating a structured multi-node instantaneous snapshot data table of physical quantities. This table contains the spatial coordinates (X, Y, Z), temperature value, and strain value for each node, providing a discrete-point dataset for spatial field analysis.

[0091] S1.3: Calculate the local gradient tensor data

[0092] After obtaining snapshots of physical quantities at discrete points, this embodiment further calculates gradient information that more profoundly reflects environmental inhomogeneity. Specifically, the system incorporates a digital surface mesh model that precisely matches the three-dimensional shell of the vehicle body, with the center point of each mesh cell theoretically corresponding to a sensor node. The algorithm first maps the "multi-node physical quantity instantaneous snapshot data table" onto this mesh model. For the temperature field, within each mesh cell, the system selects sensor node data (usually 3 or 4) at its vertices and performs calculations using a first- or second-order spatial finite difference method. For example, for a quadrilateral mesh, the algorithm calculates the rate of temperature change along two orthogonal tangent directions: and These two partial derivatives form a two-dimensional temperature gradient vector. This vector not only contains the magnitude of the gradient ||▽T|| (characterizing the drasticness of temperature change), but more importantly, it contains its direction (pointing to the direction of the fastest temperature increase). For the strain field, the calculation is more complex, requiring the handling of the strain tensor. The system calculates the strain gradient tensor at a point from strain data in multiple directions. This is a higher-order mathematical object that describes the spatial inhomogeneity of material deformation. The output of this step is a dataset of local gradient tensors calculated for the center point of each effective grid cell. This dataset serves as a crucial bridge connecting discrete point measurements and continuous field analysis.

[0093] S1.4: Generate the environmental gradient field map data at the current moment.

[0094] Finally, the system synthesizes the discrete gradient tensor data into an intuitive and continuous environmental map. Further, the system employs a Kriging space interpolation algorithm, using the calculated local gradient tensor dataset as known sample points to fill the entire vehicle surface and its surrounding space (a hexahedral area centered on the vehicle with sides of 2 meters). The algorithm interpolates not only the magnitude of the gradient but, more importantly, its direction, ensuring the generated vector field is physically continuous. In an optional implementation, for visualization and subsequent processing, the system normalizes the interpolation results into a 5cm x 5cm x 5cm three-dimensional voxel grid. Each voxel stores two core attributes: 1) the temperature gradient vector (▽T); 2) the direction and magnitude of the principal strain gradient. Ultimately, the system generates and outputs a high-resolution spatiotemporal gradient field map. This data can be in the form of a multidimensional matrix or a point cloud vector field. It not only clearly displays macroscopic features such as "a strong cold airflow descends from the left side of the third shelf (forming a high gradient zone)" or "a localized temperature rise zone forms at the rear of the shuttle due to motor heat dissipation (gradient direction outward)," but also reveals minute hot spots or stress concentrations caused by microstructures such as weld seams and screw connection points. This map is the most direct data input for the Environmental Dynamic Embedded Model (EDEM), realizing a leap from point perception to field cognition of the environment.

[0095] S2: Dynamic environment model update based on physical inversion and data fusion, including:

[0096] S2.1: Retrieval of environmental heat flow disturbance data

[0097] This embodiment reveals the active environmental driving forces hidden behind the gradient field. Specifically, the system uses the temperature gradient field from the generated high-resolution spatiotemporal gradient field map data as a boundary condition, inputting it into a physical inversion model constructed based on an unsteady-state heat conduction equation. This model treats the vehicle surface as a solid boundary with specific material properties (such as thermal conductivity, specific heat capacity, and density). According to a further improvement of this embodiment, the inversion problem is constructed as an optimization problem: finding a heat flux density function q(x,y,z,t) distributed on the vehicle surface such that the surface temperature distribution calculated by the positive heat conduction model based on this heat flux input minimizes the error between the measured temperature gradient field (which can be converted into a temperature field after one integration). In a preferred embodiment, the Tikhonov regularization method is used to solve this ill-posed problem to suppress the instability of the solution caused by measurement noise. The solution process is completed using an iterative algorithm (such as the conjugate gradient method). Finally, the algorithm outputs a vector distribution of heat flux density on the vehicle surface. This data has a clear physical meaning: for example, it may be displayed as a negative value in a certain region (e.g., -200 W / m). 2 A positive value indicates that the environment is absorbing heat from the area (cooling effect); a positive value indicates that the environment is heating the vehicle body (possibly from heat dissipation from adjacent equipment or radiation from personnel). This directly transforms passive temperature results into active thermal disturbance input, which is a key basis for assessing the level of environmental threat.

[0098] S2.2: Identify and label high-risk characteristic areas data

[0099] After obtaining heat flux data, the system initiates a multi-source information fusion and pattern recognition process to automatically mark potential threats in the environment. Specifically, the algorithm simultaneously retrieves three sets of data: 1) the inverted heat flux density vector distribution data of the vehicle body surface; 2) the corresponding strain gradient field data; and 3) the statistical characteristics of the same location area in historical data. The system has a series of pre-set intelligent judgment rules. For example, rule one: "Judgment of high heat flux abrupt change zones." The algorithm calculates the spatial divergence of heat flux density. When the absolute value of the divergence in a certain area exceeds a threshold (e.g., |▽·q| > 150 W / m²), the system identifies the threat. 3If the heat flux vector direction is disordered, the area is marked as a "strong convection vortex zone," which may cause drastic fluctuations in sensor readings. Rule 2: "Thermo-mechanical coupling risk zone determination." The algorithm performs spatial superposition analysis on high-density heat flux zones (|q| > preset value) and high-value strain gradient zones. If the two highly overlap spatially, the area is determined to have a risk of high internal stress due to uneven thermal expansion and contraction, and is marked as a "thermal stress concentration zone," which should be avoided for extended periods. Rule 3: "Potential phase change interface determination." Combining areas where the local temperature is close to the freezing point and the heat flux is a small positive value, the algorithm will mark the area as a "condensation risk zone." All these determination results, including the spatial polygonal boundary of the area, risk type (such as vortex, stress, condensation), risk level (high, medium, low), and core quantitative indicators, are integrated into a structured list of high-risk environmental features. This list is like adding clear warning labels to an environmental map.

[0100] S2.3: Assimilate historical data and update the dynamic environment model

[0101] To endow the system with predictive capabilities, this embodiment introduces data assimilation technology to maintain a dynamically evolving environmental model. In a preferred implementation, the core model is a simplified, reduced-order coupled computational fluid dynamics and heat conduction model. The model's state variables include temperature and airflow velocity in key areas. At each step, the model predicts the current state from the previous state based on its internal physical laws. Then, the system performs an assimilation operation: comparing the model's prediction for the current state with the observed and calculated heat flux density distribution data and high-risk feature list data (converted into observations that can be assimilated by the model). Using Kalman filtering or ensemble Kalman filtering algorithms, the difference between the prediction and observation (i.e., the innovation vector) is calculated, and this is used to adjust the model's state variables and some uncertain parameters to make the model's prediction optimally approximate the actual observation. This closed-loop process of "prediction-observation-correction" is data assimilation. Its final output is updated dynamic environmental model data. This data not only contains the optimal estimate of the current environmental state, but more importantly, it contains the model's short-term forecast of environmental evolution over the next few seconds to tens of seconds based on the assimilated state. For example, the model might predict: "In 10 seconds, the cold air front from channel 3 will arrive at the current position, causing the temperature on the left to drop by 5°C and potentially forming a momentary strong gradient region." This forward-looking information forms the basis for decision-making in achieving mimicry navigation and active disturbance rejection.

[0102] S3: Real-time online identification of phase transition state at the ground interface

[0103] S3.1: Acquire native state data of the drive system

[0104] Specifically, the system synchronously acquires the native physical signals of the drive motor and vehicle body with high fidelity. For each wheel-side drive motor, instantaneous waveform data of the motor's three-phase AC current is acquired via a high-precision current sensor at a sampling frequency of no less than 5kHz. Simultaneously, high-resolution rotor position feedback data (typically provided by an encoder) is directly read from the motor controller; this data, after differentiation, yields accurate instantaneous speed. Furthermore, an industrial-grade IMU (Inertial Measurement Unit) installed on the drive axle or frame near the wheels acquires triaxial high-frequency vibration acceleration data at a frequency of 2kHz. In an optional implementation, DC bus voltage data of the motor driver is also acquired to eliminate the influence of power supply fluctuations on current analysis. All these signals are strictly synchronized via timestamps, forming a multi-dimensional native state data stream of the drive system. This data stream fully records all echoes generated by the interaction force between the wheel and the ground, transmitted back through the transmission system to the motor and vehicle body, providing the raw acoustic signature for diagnosing the condition of the ground interface.

[0105] S3.2: Extracting nonlinear harmonic characteristic data of current signals

[0106] The system performs deep spectrum mining on the acquired current signals to find the characteristic fingerprints of stick-slip oscillations. Further, the algorithm performs a windowed short-time Fourier transform on the instantaneous waveform data of each phase current to obtain a dynamic time-frequency spectrum. Traditional motor fault diagnosis focuses on integer multiples of harmonics (e.g., the 5th, 7th), while this embodiment focuses on fractional harmonics and sideband frequencies. Specifically, the algorithm first calculates the current fundamental frequency *fe* based on real-time speed data. Then, it searches the spectrum for frequency components centered on fractional multiples of *fe* (e.g., 0.5fe, 1.5fe, 2.5fe) and sidebands centered on *fe* and modulated by mechanical characteristic frequencies (e.g., *fe ± fr*, where *fr* is the estimated tire rolling frequency or its harmonics). The presence of these components is direct evidence of nonlinear friction, periodic slip, or vibration coupling in the system. The algorithm accurately extracts the amplitude, phase, and their trends with speed of these specific frequency components, forming a current nonlinear harmonic characteristic spectrum data. For example, when a tire begins to microslip on thin ice, the amplitude of the 0.5fe harmonic may increase significantly; and when periodic stick-slip oscillations occur, a distinct symmetrical sideband appears near fe. This characteristic spectrum is one of the core inputs for subsequent phase classification.

[0107] S3.3: Calculate the comprehensive state index data of the current contact interface.

[0108] Finally, the system integrates multimodal features for intelligent classification, providing a health check report of the contact interface. In a preferred embodiment, a lightweight one-dimensional convolutional neural network is used as the classifier. The input of this network is a carefully constructed feature vector that integrates: 1) current nonlinear harmonic characteristic spectrum data (as frequency domain features); 2) specific frequency band energy related to tangential and vertical vibrations obtained after wavelet packet decomposition of triaxial high-frequency vibration acceleration data (as time-frequency domain features); 3) local temperature estimation data of the ground directly beneath the wheel; and 4) real-time dynamic vertical load estimation of the wheel (estimated based on vehicle attitude and acceleration). In the offline phase, this CNN model is trained using a large amount of data collected on simulated ice surfaces, wet and slippery surfaces, and other road surfaces with different adhesion coefficients, learning to map complex input features to several typical contact phases. During online operation, the model performs independent and rapid real-time inference for each wheel, outputting an interface phase probability distribution vector. The vector is of the form [P_solid, P_quasi_liquid, P_stick_slip, P_macro_slip], where each element represents the probability of the corresponding state, and the sum is 1. For example, an output of [0.85, 0.10, 0.05, 0.00] indicates that the system is 85% confident that the current state is a stable solid-state contact, but there is a 10% probability that a lubricating water film is forming. This quantified, probabilistic state data provides extremely fine and forward-looking input for subsequent torque distribution and stability control, enabling a leap from sensing slippage to predicting adhesion.

[0109] S4: Digital twin compensation for hysteresis in motor mechanisms

[0110] S4.1: Record the deviation data between the instruction and the execution trajectory.

[0111] This embodiment aims to establish a precise record of the gap between the actual behavior of the physical mechanism and the ideal model. Specifically, whenever the motion control system completes a closed-loop action, such as "performing a linear interpolation motion from coordinate point A to coordinate point B" or "raising the forks to the target height H," a high-precision data recording and comparison process is initiated. The system first extracts the complete theoretical control command timing sequence data issued for executing the action from the motion controller's command buffer. This data typically includes the target position, target speed, and even target current value issued to the servo motor in each control cycle (e.g., every 1 millisecond), forming the gold standard for describing the desired motion trajectory. Simultaneously, the system collects the actual motion trajectory feedback timing sequence data within the same time period from a high-precision feedback device (such as a linear encoder, absolute multi-turn encoder, or laser tracker) installed at the execution end. In a preferred embodiment, to ensure the accuracy of time alignment, the system employs a hardware interrupt-based synchronous acquisition mechanism, stamping both sets of data with microsecond-level timestamps generated by the same clock source.

[0112] Subsequently, the comparison and analysis algorithm is initiated. The algorithm compares the target position curve in the theoretical control command time-series data with the actual position curve in the actual motion trajectory feedback time-series data, point-by-point along a strictly aligned time axis. The calculation not only focuses on the steady-state position error after the motion ends but also delves into the dynamic error profile throughout the entire motion process. For example, the algorithm calculates whether the actual position lags behind the target position during the acceleration phase; whether there are high-frequency micro-oscillations during the constant speed phase; and whether there is overshoot or undershoot during the deceleration and stopping phase. These errors are quantified into a high-fidelity trajectory tracking error profile dataset containing time, error magnitude, and error derivative (velocity error). Specifically, the system correlates the environmental context at the time of the action, such as the average mechanism temperature at that time, and appends this context as metadata to the error dataset. This detailed medical record provides irreplaceable primary observational evidence for subsequent diagnosis of the mechanism's health status (i.e., hysteresis characteristics).

[0113] S4.2: Update parameters of the hysteresis nonlinear digital twin model

[0114] After obtaining the error history, this data is used to perform a health check and parameter tuning on the digital twin model characterizing the mechanism's properties. The high-fidelity trajectory tracking error profile dataset and the timing sequence data of the theoretical control commands that triggered the motion are input into the corresponding hysteresis nonlinear digital twin model of the motion mechanism. In a preferred embodiment, this model employs a differential equation model with a well-defined physical background, such as the Bouc-Wen or Prandtl-Ishlinskii model, which contains a series of parameters, such as stiffness coefficients characterizing elasticity, damping coefficients characterizing friction, and parameters describing the shape and width of the hysteresis loop. These parameters are not fixed but are functions of ambient temperature (especially low temperatures) and the long-term operating state of the mechanism.

[0115] Specifically, the parameter update process is constructed as an optimization problem. The algorithm uses the current model parameters as initial values, inputs the theoretical control command time-series data into the model for simulation, and obtains the trajectory predicted by the model. Then, it calculates the difference between the model-predicted trajectory and the actual trajectory inferred from the observed high-fidelity trajectory tracking error profile dataset (i.e., the theoretical command superposition error). The goal is to minimize this difference by iteratively adjusting the model's internal parameters. In an optional implementation, this process is achieved using online parameter identification algorithms such as recursive least squares or extended Kalman filtering. These algorithms can efficiently process time-series data and provide optimal estimates of the model parameters and their uncertainties. For example, after training with forklift action data from low to high, the algorithm might find that the static friction parameter value in the model needs to be increased to explain the observed start-up delay. The output of this learning process is a set of updated hysteresis model parameter vectors, enabling the digital twin model to achieve a new and higher level of fidelity in its imitation of the physical mechanism at the current temperature and operating point.

[0116] S4.3: Generate feedforward compensation command data

[0117] After the digital twin model is precisely calibrated, its predictive capabilities are used to eliminate problems before they occur. When the motion planner generates a new motion requirement and produces raw motion command data to be executed, the command is not directly sent to the physical servo driver. According to a further improvement in this embodiment, the command data is first fed into a hysteretic nonlinear digital twin model that has undergone parameter updates for a preliminary simulation. Based on its internal mathematical description of the physical mechanism's characteristics such as "inertia," "friction," and "elastic deformation," the digital twin model quickly simulates and calculates the actual motion trajectory that the physical end effector will produce if the raw command is executed directly. Then, the model compares this predicted actual trajectory with the expected ideal trajectory (i.e., the raw command itself) and calculates the predicted trajectory deviation.

[0118] Next, the compensation algorithm generates a mirrored correction command based on the predicted deviation. Specifically, if the model predicts a 5-pulse lag in the acceleration phase, the compensation algorithm injects an additional 5 pulses in advance during the acceleration phase of the original command; if it predicts a 3-pulse overshoot at the stopping point, the algorithm reduces the corresponding command amount in advance during the deceleration phase. This process generates a feedforward compensation correction amount. Finally, the system superimposes the original motion command data to be executed with the feedforward compensation correction amount in the command domain in real time, thereby synthesizing a pre-distortion optimized final execution command. This final command already contains the correction amount needed to offset all predicted hysteresis and nonlinear effects. When it is sent to the physical servo system, the inherent defects of the physical mechanism are precisely neutralized by the pre-distortion part in the command, allowing the final execution result to reproduce the initially expected ideal trajectory. This achieves an intelligent compensation effect with ultra-high mechanical precision even in harsh low-temperature environments.

[0119] S5: Reflective Autonomous Decision-Making and Navigation Flow Generation Integrating Multi-Source Information

[0120] S5.1 (Millisecond-level reflective layer): Generates emergency evacuation command data

[0121] This layer establishes a direct, dedicated high-speed data channel to continuously monitor the highest-level alarms in the environmental high-risk characteristic list data (such as "sudden strong convective vortex intrusion"), as well as state transitions in the interface phase probability distribution vector that indicate imminent instability (such as "macroscopic slip probability" surging from 5% to 95% within 2 milliseconds). This data is fed into a reflection rule lookup table composed of hardware logic (such as FPGA) or a minimal state apparatus.

[0122] Specifically, each reflection rule is a "condition-action" pair. The condition is a Boolean judgment based on the input data, such as: "IF (right rear wheel macroscopic slip probability > 90%) AND (vehicle yaw rate > threshold) THEN…". The action is a direct, quantitative hard control command to the actuator. When the input data meets the condition of a rule, this layer bypasses all upper-level decision loops and directly outputs the corresponding instinctive emergency correction command data within a millisecond delay. For example, for the slip condition mentioned above, the action might be: "Instantly reduce the torque of the right rear wheel to zero, increase the torque of the left front wheel by 50%, and apply differential braking for 0.1 seconds." These commands are pre-programmed and have been validated through extensive simulations and experiments as optimal emergency responses. The output of this layer is discrete, pulse-like control commands, the purpose of which is to quickly suppress the spread of dangerous conditions, buy valuable reaction time for the upper-level system, and serve as the last and fastest line of defense for safety.

[0123] S5.2 (Second-level Adaptation Layer): Generates local motion parameter adjustment data.

[0124] This layer is responsible for smooth and continuous motion optimization and parameter tuning in non-emergency situations, continuously updating the dynamic environment model data (especially its short-term prediction function) and the continuously evolving interface phase probability distribution vector. Unlike the reflection layer, the adaptation layer does not generate direct motion trajectory commands, but dynamically adjusts the behavior parameters of the underlying controllers (such as wheel torque controllers, steering controllers, and suspension controllers) to proactively adapt to the environment.

[0125] Furthermore, an adaptive parameter mapper algorithm operates within this layer. This algorithm calculates an optimal set of vehicle dynamic control parameters in real time based on environmental predictions and ground conditions. For example, when it predicts entering a cold, windy area 10 meters ahead, the algorithm might generate dynamic chassis parameter tuning data based on table lookups or simple calculations. This includes adjusting the front-to-rear torque distribution ratio of the four-wheel drive system from 50:50 to 40:60 to increase rear-wheel drive stability; slightly sluggishly adjusting the steering ratio of the steering system to counteract crosswind interference; and lowering the vehicle height by 5 millimeters to lower the center of gravity. Simultaneously, based on the real-time adhesion probability of each wheel, the upper limit of its drive / braking torque is dynamically adjusted to implement preventative torque management, ensuring that power is always distributed to the wheel with the greatest adhesion potential. These adjustments are gradual and smooth, optimizing the vehicle's driving feel and stability under known or predicted environmental disturbances, demonstrating the system's proactive adaptive capabilities based on environmental understanding.

[0126] S5.3 (Minute-level Learning Layer): Generates global mimicry navigation flow data

[0127] This layer is responsible for planning the optimal action route from a macro-level task perspective. Its main input is dynamic environmental model data covering the entire operational area. This model data is treated here as an environmental disturbance potential field composed of heat flow disturbances, temperature gradients, and risk area density.

[0128] Specifically, the path planning algorithm in this layer employs an innovative minimum perturbation potential energy path search strategy. Unlike the traditional Dijkstra's algorithm, which searches for the geometric shortest path, this algorithm assigns a travel cost to each location on the map, which is positively correlated with the intensity of environmental perturbation at that location (e.g., high heat flux gradient areas and known risk areas have high costs; corridors with uniform temperature and stable airflow have low costs). The algorithm then searches for the path from the starting point to the target point with the lowest cumulative travel cost. This derives global mimicry navigation flow data. This path may not be the shortest, but it guides the shuttle to automatically avoid air conditioner vents, stay away from corners where condensation frequently occurs, and choose the shelf aisles with the most stable temperature. The output navigation flow not only contains a series of waypoint coordinates but also includes recommended cruising speeds, acceleration curves, and energy-saving strategy suggestions for each path segment. For example, when traversing an unavoidable micro-gradient area, the navigation flow suggests reducing speed for a smoother passage. This navigation flow reflects the high-level mission intent, guiding vehicles on a macro level on how to coexist gracefully with the environment and achieve the overall optimization of efficiency, safety, and energy consumption.

[0129] S5.4: Fusion to generate final control command data

[0130] This embodiment is responsible for integrating instructions from three different levels, which may have different objectives and time scales, into a coordinated, consistent, and conflict-free final control command. The central arbiter receives instinctive emergency correction command data (highest priority) from the reflection layer, dynamic chassis parameter tuning data from the adaptation layer, pre-distortion optimized final execution command data from S4 (for high-precision point positioning), and smoothed global mimicry navigation flow data from the learning layer.

[0131] Specifically, the arbitrator internally runs a context-based mixed-integer optimizer. This optimizer first performs conflict detection and resolution. For example, when the learning layer plans a straight path, but the reflection layer issues a slight right-turn correction instruction due to a sudden strong wind on the right, the optimizer prioritizes the reflection layer's instruction and dynamically adjusts the learning layer's short-term path points. Under normal, conflict-free conditions, the optimizer (in a preferred embodiment, employing a model predictive control framework) uses the instructions from each layer as weights for the optimization objective. It uses the vehicle dynamics model as constraints to solve for a future, finite-time control sequence that: 1) maximizes the tracking of the learning layer's global path; 2) satisfies the vehicle dynamic characteristics adjusted by the adaptation layer; 3) accurately executes the compensated point instructions from S4; and 4) ensures absolute safety (respecting all red lines defined by the reflection layer). The resulting optimal control sequence is encapsulated into a final cooperative control instruction data package and distributed to various actuators such as drive, steering, and lift. This data package enables the shuttle to operate like a well-trained expert in complex cold storage environments, demonstrating instinctive avoidance of sudden dangers, skillful adaptation to normal disturbances, meticulous precision in operation, and optimal planning for long-term tasks, thus achieving truly intelligent and autonomous operation.

[0132] According to another aspect of this application, calculating the local gradient tensor data further includes:

[0133] S1.3.1: Perform spatiotemporal alignment and outlier removal on physical quantity snapshot data.

[0134] Before proceeding to the core gradient calculation, the multi-node physical quantity instantaneous snapshot data table is first cleaned and preprocessed. Specifically, the algorithm checks the timestamp synchronization of the data from each sensor node. For minor deviations caused by transmission delays, an interpolation algorithm based on network time protocol and sensor ID is used for alignment, generating time-synchronized physical quantity field data. Further, a statistical outlier detection algorithm is run. For example, for temperature data, the algorithm calculates the mean and standard deviation of the temperature for all nodes, marking nodes whose data deviates from the mean by more than three times the standard deviation as suspected outliers. These outliers may originate from transient sensor malfunctions or localized extreme interference (such as brief contact with cargo). In an alternative implementation, the system does not simply discard outliers but initiates a spatial correlation-based repair procedure: using the normal sensor data adjacent to the node, a reasonable alternative value is estimated using inverse distance weighted interpolation. Finally, a cleaned and repaired, highly reliable clean physical quantity field mesh data is output, laying the foundation for subsequent precise differential calculations.

[0135] S1.3.2: Gradient Operator Construction and Computation Based on Unstructured Mesh

[0136] Given the non-uniform distribution of sensor nodes on the vehicle surface, a gradient calculation method suitable for unstructured meshes is adopted. First, each sensor node in the clean physical quantity field mesh data is connected to its nearest neighbor node, constructing a triangulated mesh (Delaunay triangulation) covering the vehicle surface. For each triangular element in the mesh, the algorithm performs the following calculations: obtaining the physical quantity values ​​(e.g., temperatures Ti, Tj, Tk) and their three-dimensional coordinates of the three vertices i, j, and k of the triangle. Based on the linear shape function theory in the finite element method, it can be assumed that the physical quantities change linearly within the triangular element. By solving a system of linear equations, the gradient vector at the center of the triangular element can be directly calculated. Specifically, for the temperature field, the element gradient ▽T can be obtained by the following formula (taking a two-dimensional projection as an example, but the algorithm is performed in three-dimensional space): ▽T=(1 / 2A)[(Ti(yj-yk)+Tj(yk-yi)+Tk(yi-yj))i+(Ti(xk-xj)+Tj(xi-xk)+Tk(xj-xi))j]; where A is the directed area of ​​the triangle, (xi,yi), (xj,yj), and (xk,yk) are the coordinates of the three vertices, and Ti,Tj,Tk are the physical quantities (such as temperature) at the corresponding vertices. This calculation process is performed in parallel on all triangular elements, ultimately generating the gradient vector data of the unstructured mesh elements corresponding to each element. This method can better adapt to complex vehicle geometry and ensure the accuracy of gradient calculation even at boundaries and in areas with high curvature.

[0137] S1.3.3: Smoothing of Gradient Fields and Reconstruction of Tensor Fields

[0138] Due to noise from the original sensors and high-frequency fluctuations that may be introduced by discrete computation, directly calculated cell gradient data may not be smooth enough. Therefore, an edge-preserving smoothing filter is performed on the gradient vector data of unstructured mesh cells. In a preferred embodiment, an anisotropic diffusion filter algorithm is used. This algorithm performs strong smoothing along the isotherm (or isostrain line) direction to suppress noise, while performing weaker smoothing in the direction perpendicular to the isotherm (i.e., the gradient direction) to preserve the true boundary and abrupt change characteristics. After smoothing, smoothed cell gradient field data is obtained. For strain gradients, the system needs to reconstruct the complete strain gradient tensor. The algorithm gathers multiple gradient estimates calculated from different neighboring cells for each node and fits the most consistent first-order strain tensor (a 3x3 matrix) at that node using the least squares method. Finally, the system outputs a structured local gradient tensor field dataset, which is organized in both node and cell forms, containing both vector-form temperature gradients and tensor-form strain gradients, fully describing the first-order variation characteristics of the physical field in space.

[0139] According to another aspect of this application, the inverted environmental heat flow disturbance data also include:

[0140] S2.1.1: Construct a parameterized forward model of vehicle body heat transfer and set up an inversion problem.

[0141] The inversion process begins with an accurate forward model. First, a parametric digital twin model of the vehicle body is loaded, containing detailed geometry, material layers (metal shell, insulation layer, internal components), and thermal properties (thermal conductivity λ, density ρ, specific heat capacity c). This model is discretized into a finite element mesh. Based on this, a finite element solver for the unsteady heat conduction equations is constructed as the forward model F. The inversion problem is formalized as follows: given the temperature field T_obs observed on the vehicle body surface Γ, find a heat flux density distribution q defined on Γ such that the difference between the output F(q) of the forward model and T_obs is minimized. This is expressed as an optimization problem with an objective function J(q): min J(q) = ||F(q) - T_obs|| 2 + αR(q). Here, the first term is the data fitting term, the second term is the regularization term R(q), used to introduce prior knowledge about the solution q (such as smoothness), and α is the regularization parameter used to balance the degree of fit and the reasonableness of the solution. The system initializes this heat flow inversion optimization problem framework.

[0142] S2.1.2: Employing the adjoint method to efficiently calculate the gradient and solve iteratively.

[0143] Since the forward calculation of the heat conduction equation is already very time-consuming, directly using optimization methods that require repeated forward calculations of a large number of parameters q (such as genetic algorithms) is not feasible. Based on a further improvement in this embodiment, an efficient gradient optimization algorithm based on the adjoint method is adopted. The specific process is as follows: 1) Start from the initially hypothesized heat flow distribution q0; 2) Perform one round of forward calculation to obtain the predicted temperature field T_pred = F(q0), and calculate the residual T_pred - T_obs; 3) Run an adjoint calculation. The adjoint equation is mathematically the "inverse" version of the forward equation, allowing the system to obtain the gradient ▽J(q0) of the objective function J relative to thousands of heat flow parameters q on the entire surface in just one calculation; 4) Use the calculated gradient ▽J(q0) to update the heat flow distribution using the conjugate gradient method or a quasi-Newton method (such as L-BFGS) to obtain q1; 5) Repeat steps 2-4 until the residual satisfies the convergence condition or reaches the maximum number of iterations. This iterative process outputs the heat flow sequence and residual data during the iteration process.

[0144] S2.1.3: Determining the optimal regularization parameter and quantifying uncertainty based on the L-curve method

[0145] The choice of regularization parameter α is crucial to the rationality and stability of the inversion results. The system employs the L-curve method for automated selection. Specifically, in each optimization iteration or for a series of different α values, the algorithm performs a complete inversion process, recording the data fitting residual norm ||F(q) - T_obs|| and the regularization term norm ||R(q)|| corresponding to each set of α. On a graph with these two norms as logarithmic coordinates, these points typically form an "L"-shaped curve. The α value at the inflection point of the curve is considered the optimal trade-off between fitting the observed data and maintaining the stability of the solution. The system automatically identifies this inflection point and uses the heat flux solution under the corresponding α value as the final result, generating heat flux density distribution data under optimal regularization. Furthermore, to evaluate the reliability of the inversion results, the system performs simple uncertainty quantification. For example, by calculating the approximate inverse of the Hessian matrix (the second derivative of the objective function), variance estimates of the major components in the solution q (such as the average heat flux in different regions) can be obtained. Finally, the system outputs environmental heat flux disturbance data and a confidence report that includes the optimal heat flux distribution and its uncertainty range.

[0146] According to another aspect of this application, the calculation of the comprehensive state index data of the current contact interface also includes:

[0147] S3.3.1: Synchronization, Alignment, and Fusion Vector Construction of Multi-Source Heterogeneous Features

[0148] First, the system performs strict spatiotemporal alignment: nonlinear harmonic characteristic spectrum data from the current sensor, vibration characteristic energy data from the IMU, and local ground temperature gradient data are uniformly interpolated onto the same phase axis based on the wheel rotation period, according to the timestamps of their acquisition hardware and the known fixed transmission delay, generating a time-phase synchronized multimodal feature stream. Next, for each wheel, the algorithm extracts the synchronized feature stream within a fixed time window (e.g., 100 milliseconds). For each data mode within the window, a set of statistical features is extracted, such as: the mean, variance, and peak value of each key component in the harmonic characteristic spectrum; the root mean square and kurtosis of the vibration energy; and the mean and trend of the temperature gradient. All these statistical features are concatenated into a high-dimensional original fused feature vector. In a preferred embodiment, contextual features, such as the current wheel slip ratio estimate and the dynamic vertical load estimate, are also added.

[0149] S3.3.2: Dynamic Feature Space Optimization Based on Online Feature Selection and Dimensionality Reduction

[0150] Directly inputting the original fused feature vector into the classifier may lead to the curse of dimensionality and overfitting, and the importance of features may differ under different operating conditions. Therefore, the system introduces an online feature selection and dimensionality reduction module. This module maintains a feature importance scoring table, with scores based on the mutual information between the feature and the historically labeled phase labels or on the feature importance based on a tree model. In each time window, the algorithm dynamically selects the Top-K most important features from the feature pool based on the current vehicle state (e.g., speed range, steering state), forming a concise currently effective feature vector. Furthermore, online principal component analysis can be used to reduce the dimensionality of the selected features, compressing the feature vector to a lower dimension while retaining most of the information, forming the dimensionality-reduced core feature vector. This not only improves the computational efficiency and generalization ability of the subsequent classifier but also enables the system to adaptively focus on the most relevant signal features at present.

[0151] S3.3.3: State reasoning and confidence calibration using an integrated incremental learning classifier

[0152] The system employs a lightweight online ensemble learning model as its core classifier, such as an online random forest based on Hoeffding trees. The advantage of this model lies in its ability to perform incremental learning. During state inference, the dimensionality-reduced core feature vector is input into this ensemble model. Each tree provides an independent class probability prediction, and the outputs of all trees are ultimately integrated through a soft voting mechanism to generate a preliminary raw phase probability distribution vector. To calibrate the confidence of the probability output and make it more realistically reflect uncertainty, the system is followed by an online Platt Scaling or Isotonic Regression calibrator. This calibrator uses the correspondence between the classifier's predicted probabilities and actual results in recent historical data to learn a mapping function and recalibrate the raw probabilities. Finally, it outputs a calibrated interface phase probability distribution vector with reliable confidence information. Furthermore, if the system receives operator feedback or obtains an accurate "ground truth" label through other high-confidence methods (such as significant wheel slippage), this label, along with the corresponding feature vector, is used for online incremental updates of the model, enabling the classifier to continuously adapt to slow changes in characteristics caused by vehicle wear, tire aging, etc.

[0153] According to another aspect of this application, generating feedforward compensation instructions further includes:

[0154] S4.3.1: Multi-step forward trajectory prediction and bias mapping based on digital twin

[0155] Upon receiving the raw motion command data to be executed (e.g., a position-velocity-time curve), the compensation algorithm first performs multi-step forward prediction. The system inputs the raw command sequence and the current mechanism temperature into an updated hysteresis nonlinear digital twin model. This model executes the entire process of the command in the digital domain with a simulation step size faster than the actual control cycle (e.g., 0.1 ms simulation step size for an actual control cycle). The model outputs not only the predicted end trajectory but also internal state variables, such as simulated friction states and elastic deformation. By comparing the predicted trajectory with the original ideal command at each simulation step, the algorithm generates a high-time-resolution time series of predicted trajectory deviation data. This data details how the deviation evolves over time: where the hysteresis begins, how the hysteresis increases, and how it exhibits backlash during velocity reversal.

[0156] S4.3.2: Compensation Quantity Synthesis Based on Inverse Model and Feedforward Control

[0157] After obtaining the predicted trajectory deviation time series data, it needs to be converted into compensation control quantities. In a preferred embodiment, the system employs a feedforward synthesis method based on an inverse model. The algorithm treats the predicted deviation data as a disturbance trajectory that needs to be canceled. This disturbance trajectory is input into a simplified, linearized mechanism inverse dynamics model. This inverse model calculates the additional motor torque or current command that needs to be injected based on the displacement deviation to be canceled. For example, to cancel a predicted position deviation Δx(t) at time t, the inverse model calculates the required compensation force F_comp(t) = K_p * Δx(t) + K_d * Δx(t) based on the equivalent stiffness and damping of the mechanism. Here, K_p and K_d are coefficients tuned according to the current state of the mechanism. By calculating the entire time series, an original feedforward compensation torque / current command sequence is generated.

[0158] S4.3.3: Smoothing, Limiting, and Final Overlay of Compensation Commands

[0159] Directly synthesized compensation commands may contain high-frequency components, and direct superposition may excite mechanical resonance. Therefore, the system performs low-pass filtering on the feedforward compensation torque / current command sequence to remove components higher than the first-order natural frequency of the mechanism, resulting in a smoothed compensation command. Simultaneously, the algorithm implements strict amplitude limiting protection: the amplitude of the compensation command must not exceed the maximum overload capacity of the motor at the current speed, and the total command after superposition with the original command must also fall within a safe range. Finally, in the time domain, the smoothed compensation command is superimposed point-by-point with the original motion command data to be executed (in the current or torque command domain). Before superposition, it must be ensured that the time bases of both are completely synchronized. After superposition, the pre-distortion optimized final execution command data is generated. The system can also set a compensation activation threshold; that is, when the prediction deviation is less than a certain value (e.g., one encoder pulse), no compensation is applied to avoid introducing unnecessary jitter at standstill or extremely low speeds. At this point, an intelligent control command that has anticipated and offset its own hysteresis defect has been generated and is ready for execution.

[0160] According to another aspect of this application, generating global mimicry navigation flow data also includes:

[0161] S5.3.1: Construction and Integration of Multi-Level Environmental Cost Maps

[0162] First, the various information contained in the dynamic environment model data is transformed into cost maps of different levels. Specifically, the following levels are created:

[0163] Basic access layer: Based on a static map, it identifies fixed obstacles, shelf locations (access cost is infinite), and walkable passages (basic cost is low).

[0164] Thermal perturbation cost layer: The magnitude of the temperature gradient ||▽T|| is mapped to cost, with high gradient regions incurring high costs. Simultaneously, regions in the environmental high-risk feature list data, such as "strong convective vortex regions," are assigned additional punitive high costs.

[0165] Energy efficiency cost layer: This considers factors such as travel distance, frequent starts and stops (at intersections), and gradient. For example, constant speed straight-line travel has low cost, while frequent turns or acceleration / deceleration zones increase cost.

[0166] Task semantic layer: Based on the current task (such as picking up goods, delivering goods, charging), assign different cost adjustment factors to the area near the target point, the charging station route, etc.

[0167] A weighted fusion algorithm is used to overlay the cost maps of each layer to generate a comprehensive, multi-layered fused environmental cost map. The weights of each layer can be dynamically adjusted according to the global strategy. For example, when emphasizing safety, the weight of the thermal disturbance layer is increased; when emphasizing efficiency, the weight of the energy efficiency layer is increased.

[0168] S5.3.2: Mimicry Path Search and Initial Trajectory Generation Based on Spatiotemporal 3D Algorithm

[0169] On a multi-layered fused environmental cost map, the system runs an improved spatiotemporal 3D search algorithm. Unlike traditional algorithms that search only in two-dimensional space, the spatiotemporal 3D search algorithm searches in a three-dimensional space (x, y, time), allowing the algorithm to "wait" or "detour" in the time dimension to avoid instantaneous environmental disturbances. For example, the algorithm might discover that delaying departure by 2 seconds can avoid a periodic strong cold air current sweeping across the predetermined path. The heuristic function considers not only spatial distance but also the cumulative environmental cost. The search algorithm outputs a preliminary spatiotemporal path sequence with the lowest overall cost in the spatiotemporal dimension from the starting point to the destination. This sequence contains a series of (x, y, t, v) points, where v is the suggested speed.

[0170] S5.3.3: Path smoothing and velocity curve optimization to generate the final navigation flow

[0171] The paths directly searched may be formed by connecting grid center points, resulting in abrupt transitions. The system smooths the initial spatiotemporal path sequence, for example, by fitting Bézier curves or spline curves in the spatial dimension to ensure continuous path curvature and smooth vehicle tracking. Next, based on the smoothed spatial path and environmental cost distribution, speed planning is performed. The speed planner considers vehicle dynamics constraints (maximum acceleration, deceleration), path curvature (curve deceleration), and strategies for traversing high environmental cost areas (such as choosing low-speed, smooth passage to reduce disturbance, or choosing high-speed passage to reduce exposure time). By solving an optimization problem, a smooth and feasible speed-time curve matching the spatial path is generated. Finally, the smoothed spatial path and optimized speed curve are encapsulated to form the final global mimicry navigation flow data. This data flow not only guides the vehicle "where to go," but also suggests "when to go" and "at what speed," achieving global spatiotemporal collaborative optimization in dynamic heterogeneous environments. The system also periodically replans and continuously optimizes unexecuted navigation flow segments based on the latest predictions from the environmental model.

[0172] This invention represents a leap from discrete-point perception to continuous-field cognition of the environment. Traditional shuttle vehicles rely on a limited number of sensors to acquire scalar data such as local temperature and humidity. This discrete and static information cannot characterize the real three-dimensional dynamic gradient fields (such as temperature gradients and airflow gradients) and their physical effects within the cold storage facility. This invention, through innovative deployment of a high-density distributed fiber optic sensor network and pioneering a method based on unstructured grid gradient calculation and physical inversion, deconstructs the environment surrounding the vehicle into a dynamic digital field containing thermal flow disturbance vectors, strain gradient tensors, and risk feature maps. This is not merely an increase in data volume, but a significant improvement in cognitive dimension: the system can "see" how the conical, strong cold airflow from the refrigeration unit outlet moves, can "understand" the invisible temperature steps formed by heat accumulation in the gaps between shelves, and can predict the time-varying impact of these gradient fields on vehicle materials, sensors, and stability. This deep environmental understanding capability allows all subsequent decisions and controls to be based on solid physical facts rather than empirical assumptions, solving the control blind spots and response lag problems caused by the superficial environmental perception of existing technologies.

[0173] Furthermore, this invention constructs a highly collaborative, bio-inspired "perception-decision-execution" closed loop. Unlike the traditional hierarchical control architecture where information flows unidirectionally and modules operate independently, this invention achieves organic collaboration across multiple time scales and task objectives through a three-layer parallel processing architecture of reflection, adaptation, and learning. The millisecond-level reflection layer, like a biological spinal reflex, directly handles emergency threats such as phase transition slip, ensuring a survival baseline; the second-level adaptation layer, like the cerebellum, dynamically optimizes chassis parameters based on environmental predictions, achieving smooth and seamless proactive adaptation; and the minute-level learning layer, like the cerebral cortex, plans a mimicry navigation path that balances efficiency and safety based on a global perturbation potential field. More importantly, the decisions of these three layers are integrated and arbitrated by a central model predictive controller, enabling the vehicle to simultaneously and optimally handle multiple coupled tasks such as "avoiding sudden strong winds," "adjusting torque distribution on ice surfaces," and "performing high-precision enhancements." This architecture enables the shuttle to exhibit near-biological intelligent behavior: it has both instinctive reactions to danger and skillful routine operations, as well as strategic thinking for long-term planning, thus demonstrating extraordinary robustness, flexibility and efficiency in extremely dynamic environments.

[0174] Finally, this invention endows the shuttle with the ability of "self-awareness" and "continuous evolution" through digital twin and online learning technologies. Addressing the long-standing challenge of mechanical hysteresis at low temperatures that has hampered precise positioning, this invention does not stop at static compensation but instead establishes parameterized nonlinear digital twin models for each key motion mechanism. This model can use the execution deviation data of each action to perform online parameter identification and track the characteristic drift of the mechanism caused by temperature changes and wear and aging in real time. Based on this, the system can perform high-fidelity simulations before executing new instructions, predict and proactively compensate for hysteresis errors, achieving "instruction pre-distortion," thereby achieving sub-millimeter-level repeatability positioning accuracy at the physical level. This paradigm of endowing fixed hardware with adaptive and self-calibrating capabilities through software algorithms not only significantly improves current performance but also, through continuous data accumulation and model optimization, enables the system's performance to continuously evolve over time, setting a new technological benchmark for the intelligent and long-term reliable operation of cold chain storage equipment.

[0175] Example 3: In this example, an intelligent shuttle equipped with the present invention will be simulated to perform a retrieval task from "starting point A of the channel" to "shelf S03-B2 storage location" in a cold storage at -25°C. The key calculation process of its core data processing unit will be tracked within a complete task cycle.

[0176] Scenario Setting and Initialization: Drive from the starting point (coordinates (0,0)) to the target storage location (coordinates (15, 5)), raise the forks, and pick up the pallet. Environmental Anomalies: 1) Due to the periodic operation of the fan near rack S03, there is a moving "strong cold airflow disturbance zone". 2) Due to previous watering, there is a thin, barely noticeable area of ​​ice on the ground near coordinates (8, 2). Vehicle Status: Four-wheel independent drive, the fork lifting mechanism has been loaded with the initial parameters of the Bouc-Wen hysteresis model, and all sensors have been calibrated.

[0177] 1. Environmental gradient field sensing calculation

[0178] Data Acquisition and Calculation: The vehicle is powered on, and the system starts. The distributed fiber optic sensor network begins operation. Real-time wavelength data streams from 128 FBG sensors are input. The edge computing unit reads data from all channels at a frequency of 100Hz. For sensor ID #47 (located on the windward side of the right front wheel arch guard), the instantaneous wavelength is read as 1540.112nm. Its calibration curve is consulted; at a reference temperature of -25°C, this wavelength corresponds to a temperature shift of +0.5°C and a micro-strain of +15με. The instantaneous calculated data for sensor #47 is output: temperature -24.5°C, micro-strain +15με. All 128 node data points together constitute the physical snapshot table at time t0.

[0179] Gradient calculation: Based on the physical snapshot table at time t0, the algorithm identifies sensors #47, #48, and #53 as forming a triangular mesh unit. Their coordinates and temperature values ​​are: T 47 =-24.5°C, T 48 =-25.1°C, T 53 =-24.8°C. Using the unstructured mesh gradient formula, the temperature gradient vector ▽T = [0.8, -0.3, 0.1] °C / cm at the center point of this element (direction pointing upwards and to the right of the vehicle's front). Parallel computation of all elements generates a local gradient tensor field, showing a specific temperature gradient region with an intensity of approximately 0.86°C / cm in the right front of the vehicle.

[0180] Field map generation: Based on the local gradient tensor field, the Kriging interpolation algorithm is used to fill the discrete unit gradient data into a 5cm resolution voxel grid, generating a 3D vector field. A high-resolution gradient field map is output, which is displayed in subsequent frames. The gradient region located near coordinates (7, 1) is moving towards (8, 2) with a slow increase in intensity, and is initially recorded by the system as a dynamic gradient feature.

[0181] 2. Environmental dynamic model update calculation

[0182] Each time a full-field map is generated, a model assimilation cycle is initiated (approximately 100ms), which includes:

[0183] Heat flux inversion: Based on the generated vehicle surface temperature field (obtained by gradient field integration), the current surface temperature field is used as a boundary condition and substituted into the parameterized vehicle thermal model (finite element mesh with approximately 10,000 nodes). Using the L-BFGS optimizer combined with the adjoint method, the difference between the predicted and observed temperatures is minimized after 5 iterations. The calculation shows that a heat flux of -120 W / m² exists on the right front side surface of the vehicle. 2 A localized area with strong heat absorption, approximately 0.1 m². 2 By combining data from multiple frames, it was determined that the heat flow zone was moving.

[0184] Risk Labeling: Based on the data of the moving thermal disturbance zone and strain gradient, the following rule engine is executed to determine: 1) Thermal flux intensity |-120| > threshold 100 W / m 2 2) The strain gradient in this area increases synchronously; 3) The moving speed matches the known wind turbine cycle. Trigger rule: "Identified as a moving strong convective vortex risk area". Judgment: Mark a dynamic risk object in the environmental model with the attributes {Type: strong convective vortex, Current position: (7.5, 1.8), Predicted trajectory: moving towards (8.5, 2.5), Expected to affect (8, 2) in 10 seconds, Risk level: high}.

[0185] Model Assimilation and Prediction: Based on all currently observed risk features and the historical gradient field sequence over the past 20 seconds, the state vector of the environmental dynamic embedded model is updated using an ensemble Kalman filter algorithm. The model predicts based on simplified hydrodynamic equations that the "dynamic risk object" will cover the region near coordinates (8,2) between t0+9s and t0+15s. The updated environmental model, incorporating this prediction information, is then communicated to the decision-making layer.

[0186] 3. Ground phase transition identification calculation

[0187] Feature extraction: Based on the three-phase current (instantaneous value), speed, and vibration data of the right front wheel drive motor, the current signal was analyzed using a 1024-point FFT when the vehicle reached coordinates (8, 1.5). The algorithm found a sideband of fe±6Hz near the fundamental frequency fe=35Hz, and the amplitude of the 0.5fe harmonic increased by 40% compared to the previous second. A current harmonic feature vector was generated, containing 12 features including sideband energy and fractional harmonic amplitude.

[0188] State reasoning: Based on harmonic feature vectors and local ground temperature (-25.5°C) + wheel load (350kg), the feature vectors are input into an online random forest model (containing 50 Hoeffding trees). 30 trees output "quasi-liquid phase transition", 15 output "stable solid state", and 5 output "stick-slip initiation". After soft voting and probability calibration, the phase probabilities generated for the right front wheel interface are: P(quasi-liquid state) = 0.65, P(stick-slip initiation) = 0.25, P(stable solid state) = 0.10. The system determines that the adhesion coefficient of this wheel is decreasing rapidly, indicating a high risk of slippage.

[0189] 4. Calculation of motor lag compensation

[0190] The vehicle arrives at the target storage location (15,5) and prepares to perform the precision operation of "lifting the forks to 1.2 meters," including:

[0191] Model Learning (Historical Actions): Based on the command trajectories and actual error data of 10 previous fork lifting actions in environments ranging from -20°C to -28°C, the Bouc-Wen model parameters were continuously optimized using a recursive least squares algorithm. It was found that when the temperature is below -25°C, the parameter "r" (controlling the hysteresis loop width) needs to be increased by approximately 15% to fit the observed start-up delay. The updated fork mechanism hysteresis model is output, which includes a temperature compensation term.

[0192] Feedforward Compensation Generation: Based on the new lifting command (target: 1.200 meters) and the current mechanism temperature (-26°C), the digital twin model simulates the execution of the command in 0.1ms steps. Predictions show that due to the viscosity of the low-temperature lubricant, a positional hysteresis of 3.5mm will occur during startup, and an overshoot of 0.8mm will occur during shutdown. The inverse model calculates a compensation current curve accordingly. A pre-compensation current command is generated, which injects an additional pulse at startup and performs current decay in advance before shutdown. Finally, these are combined into an optimized execution command sent to the lifting motor.

[0193] 5. Autonomous decision-making and navigation flow generation calculation

[0194] Based on all the above information, plan the path from the current location (avoiding risks) to the target point and control it in real time, including:

[0195] Global Path Replanning (Learning Layer): Based on the updated environment model (predicting that area (8,2) will be a high-risk zone in the next 10-15 seconds), the global path planner receives the task start and end points. The original shortest path requires passing through (8,2). Considering the risk prediction, the planner calculates two new paths: 1) Accelerate in advance to pass through the risk zone before it forms within 9 seconds; 2) Detour via the left-hand channel, increasing the distance by 2 meters but completely avoiding the risk zone. Calculated using the cost function (higher safety weight), path 2 has a lower cost. A new global navigation flow is generated: "Abandon the straight path, turn left at (7,1), and detour via channel B."

[0196] Local motion adjustment (adaptation layer): Based on the high slip probability of the right front wheel (0.65) + the new navigation flow (left turn command), the adaptive parameter mapper calculates: 1) Based on the slip probability, the upper limit of the right front wheel torque is dynamically limited to 60% of the normal value. 2) To execute the left turn, the differential torque distribution is calculated, but due to the limitation of the right front wheel, more torque is distributed to the right rear wheel to balance the turning torque. A dynamic chassis parameter adjustment command is generated: "Limit the torque of the right front wheel, increase the torque of the right rear wheel by 25%, and reduce the steering angular velocity by 10% to ensure stability in wet and slippery corners."

[0197] Reflective Hazard Avoidance (Reflective Layer) - Simulated Accident: While the vehicle is traveling along a new path, the left rear wheel accidentally runs over an unmodeled oil slick. An instantaneous increase in the macroscopic slip probability of the left rear wheel to 95% within 3ms is detected. This signal is directly sent to the reflective layer FPGA, executing the rule matching: "IF Single rear wheel macroscopic slip probability > 90% THEN Initiate differential braking and reduce torque on the same-side front wheel." A millisecond-level reflective command is output and sent directly to the driver: a short 5ms braking pulse is applied to the left rear wheel, and the torque on the left front wheel is instantly reduced by 20%. The vehicle's attitude stabilizes within 50ms, and no sideslip occurs.

[0198] Final fusion control (arbitrator): Based on the global navigation flow, chassis parameters limiting torque, pre-compensation commands for fork lifting, and occasional reflective braking pulses (completed). The Model Predictive Controller (MPC) optimizes the solution with a window of 10 control cycles (10ms). Optimization objectives: Track the detour path, meet torque limits, and smoothly execute the lifting action. A series of optimal motor torque and steering angle commands are obtained. The final output is a coordinated control command package sent to each actuator. The vehicle smoothly avoids high-risk areas, passes through slippery curves without slipping, precisely stops at the target storage location, and smoothly and accurately lifts the forks to 1.200 meters in one go, completing the pallet retrieval.

[0199] This embodiment demonstrates a complete closed loop, from microscopic sensor data processing (such as FBG wavelength calculation) to mesoscopic model prediction (such as heat flow inversion and phase identification) and then to macroscopic decision-making and control (such as path replanning and multi-level command fusion). The system demonstrates its ability to safely, accurately, and efficiently complete complex tasks in dynamic and harsh cold storage environments through deep environmental understanding, self-state awareness, and forward-looking decision-making.

[0200] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. An autonomous mimicry control and disturbance rejection method for an intelligent shuttle vehicle in a cold storage facility, characterized in that, include: Collect microscale physical quantity data of the vehicle-environment interface, and construct the environmental spatiotemporal gradient field through unstructured grid gradient calculation and spatiotemporal interpolation; Based on the environmental spatiotemporal gradient field, the environmental dynamic embedded model, which includes risk feature prediction, is dynamically updated by solving the inverse problem of heat conduction and fusing multi-source data. Collect native state data of the drive system, and identify the phase change state of the wheel-ground contact interface in real time by analyzing motor current harmonics and fusing multi-modal features. The hysteresis nonlinear digital twin model is updated online based on historical trajectory deviation data of the motion mechanism, and feedforward compensation commands are generated using this model. The generation of feedforward compensation commands includes: recording the actual trajectory feedback after the motion mechanism executes the theoretical command trajectory, and calculating a high-fidelity trajectory tracking error profile dataset; inputting the error profile dataset, along with the corresponding theoretical command sequence and ambient temperature data, into the hysteresis nonlinear digital twin model, updating the model parameters using an online parameter identification algorithm, and obtaining the updated hysteresis model parameter vector; inputting the new motion command data to be executed into the updated digital twin model for forward simulation, predicting its execution deviation, and generating a feedforward compensation torque / current command sequence through the mechanism inverse dynamics model; performing low-pass filtering and safety limiting on the feedforward compensation torque / current command sequence, and then superimposing it with the new motion command data in the command domain to generate the pre-distortion optimized final execution command data. By integrating an embedded dynamic environmental model, phase transition states, and feedforward compensation commands, a three-layer decision-making architecture of reflection-adaptation-learning is used to generate collaborative control commands, driving the shuttle to perform mimicry navigation and anti-disturbance operations. The generated collaborative control commands include: continuously monitoring the highest-level alarms in the high-risk environmental feature list data and the instability transition signals in the interface phase probability distribution vector, directly matching and outputting instinctive emergency correction command data through a hardware logic rule base; dynamically calculating and outputting dynamic chassis parameter tuning data for adjusting torque distribution, steering sensitivity, and suspension parameters based on short-term predictions of the updated dynamic environmental model data and continuous interface phase probability distribution vectors; constructing a multi-level environmental cost map based on the updated dynamic environmental model data, searching for the path with the lowest overall cost through a spatiotemporal three-dimensional algorithm, optimizing the speed curve, and generating global mimicry navigation flow data; and using the emergency correction command data, dynamic chassis parameter tuning data, final execution command data, and global mimicry navigation flow data, performing conflict arbitration and multi-objective optimization through a model predictive controller to solve and output the final collaborative control command data package.

2. The method according to claim 1, characterized in that, Constructing the environmental spatiotemporal gradient field includes: The raw wavelength data of each node in the distributed fiber Bragg grating sensor network is demodulated and physical quantity is calculated to generate a multi-node instantaneous snapshot data table of physical quantities. Spatiotemporal alignment and outlier removal are performed on the data in the instantaneous snapshot table of physical quantities to generate clean physical quantity field mesh data; Based on the theory of unstructured triangulated mesh and linear shape function, the temperature gradient vector and strain gradient tensor at the center of each mesh element are calculated to generate unstructured mesh element gradient vector data. Anisotropic diffusion filtering and node tensor reconstruction are performed on the unit gradient vector data, and high-resolution spatiotemporal gradient field map data covering the vehicle body and near-field space are generated by Kriging space interpolation.

3. The method according to claim 1, characterized in that, Dynamically updated environmental embedded models that include risk feature predictions include: Using the temperature field in the high-resolution spatiotemporal gradient field map as the boundary condition, combined with the parameterized vehicle body thermal model, a heat flux inversion optimization problem is constructed and solved using the adjoint method to invert the heat flux density vector distribution data on the vehicle body surface. By integrating heat flux density vector distribution data and strain gradient data, a pre-defined rule engine is used to identify and label high heat flux abrupt change zones, thermo-mechanical coupling risk zones, and potential phase change interfaces, generating a list of high-risk environmental features. Based on the current high-risk feature list data and historical gradient field sequences, the environmental dynamic embedded model is updated using a data assimilation algorithm.

4. The method according to claim 1, characterized in that, Real-time identification of the phase transition state at the wheel-ground contact interface, including: Collect the three-phase high-frequency current waveforms, rotor position, and vehicle vibration data of each drive motor to form the original state data stream of the drive system; Short-time Fourier transform is performed on the current waveform data to extract the amplitude and phase characteristics of fractional harmonics and sideband frequency components that are not integer multiples of the fundamental frequency, thus forming current nonlinear harmonic characteristic spectrum data. Wavelet packet decomposition was performed on the vehicle vibration data to obtain vibration characteristic data; Harmonic characteristic spectrum data, vibration characteristic data, local ground temperature data, and wheel dynamic load data are fused into a feature vector, which is then input into an online ensemble learning classifier for real-time inference. After probability calibration, the output is an interface phase probability distribution vector.

5. The method according to claim 1, characterized in that, Generate global mimicry navigation flow data, including: The gradient intensity, risk area distribution, static obstacles, and task semantic information in the updated dynamic environment model data are mapped into multi-level cost maps such as thermal disturbance cost layer and basic access layer, and multi-level fused environment cost map is generated through weighted fusion. On the integrated environment cost map, a runtime three-dimensional search algorithm is used to find the initial spatiotemporal path sequence with the optimal cumulative cost in both spatial and temporal dimensions from the starting point to the end point. The initial spatiotemporal path sequence is smoothed using spline curves, and the smoothed velocity-time curves based on vehicle dynamics constraints and path curvature planning are encapsulated to form the final global mimicry navigation flow data.

6. The method according to claim 3, characterized in that, The adjoint method is used to solve the heat flow inversion optimization problem, including: A framework for the inversion problem is constructed, which takes the heat flux density distribution on the vehicle surface as the optimization variable, the minimum difference between the temperature predicted by the forward thermal model and the observed temperature as the objective function, and includes a Tikhonov regularization term. The process begins with an initial heat flux hypothesis and iterates. In each iteration, a forward thermal calculation is performed first to obtain the predicted temperature field, followed by an adjoint calculation to efficiently obtain the gradient of the objective function with respect to all heat flux parameters. The obtained gradient is used to update the heat flux distribution using the L-BFGS optimization algorithm; The optimal regularization parameter is determined based on the L-curve method, and the solution under this parameter is used as the final heat flux density distribution data. At the same time, the variance of the main components of the solution is calculated to quantify the uncertainty.

7. The method according to claim 1, characterized in that, The model parameters are updated using an online parameter identification algorithm, including: The parameter update of the hysteresis nonlinear digital twin model is constructed as a recursive least squares optimization problem; The objective is to minimize the difference between the model-predicted trajectory and the actual trajectory inferred from the superposition of error contour data with theoretical command sequences; After each motion action is completed, the model parameter vector and its covariance matrix are recursively updated using the newly acquired error data to make the model dynamically track the changes in hysteresis characteristics caused by low temperature and wear.

Images