Intelligent control method for automatic suction of fork gauge of cold storage shuttle vehicle

Abstract

The application discloses an intelligent control method for automatic adsorption of a fork ruler of a cold storage shuttle vehicle, comprising the following steps: collecting interface state information through a multi-modal sensing data acquisition interface, constructing a contact mechanics digital twin model based on the interface state information; performing anchor point dynamic programming based on the digital twin model, and generating an optimal anchor point coordinate set resistant to dynamic interference; performing a phase change anchoring operation according to the anchor point coordinate set, and monitoring the anchoring process state in real time; collecting multi-dimensional force vector data and performing dynamic force field reconstruction, and performing adaptive and predictive adjustment of the adsorption force field based on the reconstruction result; performing controllable release operation after the task is completed, and performing system performance evaluation and parameter self-evolution. Through interface state perception, digital twin pre-rehearsal and dynamic force field control, the application realizes the transformation from passive adsorption to active intelligent anchoring, solves the problems of insufficient adhesion force of the low-temperature frosting interface and dynamic load adaptability, and improves the operation safety and reliability of the shuttle vehicle in the cold storage environment.

Description

Technical Field

[0001] This invention belongs to the field of automated warehousing and logistics technology, and in particular, it is an intelligent control method for automatically adsorbing forklift rulers on a cold storage shuttle. Background Technology

[0002] In modern cold chain logistics systems, automated storage and retrieval systems (AS / RS) have become core infrastructure. As the primary handling equipment in AS / RS areas, the reliability and efficiency of cold storage shuttles directly impact the throughput capacity of the entire cold chain system. In actual operation, shuttles need to be frequently picked up by forklifts and moved between different tracks or maintenance areas. However, the consistently low temperatures (typically -25°C to -18°C) and high humidity inside cold storage facilities pose a significant challenge to the reliable connection between the shuttles and the forklifts.

[0003] Currently, the industry primarily relies on simple mechanical limiting structures and friction-based anti-slip designs to secure shuttles and forklifts. In extreme cases, permanent magnets are used for auxiliary fixation. However, these traditional solutions have inherent drawbacks under the special conditions of cold storage: First, mechanical limiting has a clearance, which can cause the shuttle to slide relative to the forklift when starting, braking, or turning, posing a risk of falling. Second, the magnetic properties of permanent magnets weaken at low temperatures, and their attraction cannot penetrate the frost layer formed between the shuttle chassis and the forklift contact surface, significantly reducing the adsorption effect. Third, vacuum adsorption solutions fail due to the hardening and freezing of the sealing rings at low temperatures, resulting in low reliability.

[0004] A deeper problem is that existing technologies are all passive fixing solutions, unable to sense and adapt to dynamically changing loads. The acceleration, deceleration, and turning of forklifts on the smooth surfaces of cold storage generate complex inertial forces and torques on the shuttle. Traditional magnetic or friction structures cannot adaptively adjust to these dynamic forces, making them highly susceptible to fixing failure due to sudden changes in force vectors during movement, potentially leading to safety accidents. Furthermore, existing solutions lack the ability to sense interface conditions, failing to determine whether the contact surface is dry metal, frosted, or icy, thus hindering targeted optimization.

[0005] Therefore, there is an urgent need in this field for a new adsorption and fixation technology that can actively adapt to the extreme environment of cold storage, intelligently sense the interface state, and dynamically respond to load changes, in order to overcome the shortcomings of existing passive solutions in terms of reliability, safety, and adaptability. Summary of the Invention

[0006] Purpose of the invention: In view of the above-mentioned problems in the prior art, this application provides an intelligent control method for automatically adsorbing forklifts in a cold storage shuttle.

[0007] Technical solution: An intelligent control method for automatically adsorbing forklift rulers in a cold storage shuttle, comprising:

[0008] By collecting interface state information through multimodal sensing data, a digital twin model of contact mechanics is constructed based on the interface state information;

[0009] Anchor point dynamic planning is performed based on digital twin model to generate optimal anchor point coordinate set resistant to dynamic interference;

[0010] The phase change anchoring operation is performed based on the anchoring point coordinate set, and the anchoring process status is monitored in real time.

[0011] Multidimensional force vector data is collected and dynamic force field is reconstructed. Based on the reconstruction results, adaptive and predictive adjustment of the adsorption force field is performed.

[0012] After the task is completed, a controlled release operation is performed, and system performance is evaluated and parameters are self-evolved.

[0013] Beneficial effects: This invention solves the technical problems of insufficient adhesion and poor dynamic load adaptability of traditional mechanical fixing and passive adsorption solutions in the extreme environment of cold storage, and realizes intelligent, reliable and adaptive connection and fixing between the shuttle and the fork. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the overall technical solution of the present invention.

[0015] Figure 2 This is a flowchart illustrating how the present invention collects interface state information through multimodal sensing data and constructs a digital twin model of contact mechanics based on the interface state information.

[0016] Figure 3 This is a flowchart illustrating the process of generating an optimal set of anchor point coordinates that resists dynamic interference based on a digital twin model, according to the present invention.

[0017] Figure 4 This is a flowchart illustrating the process of collecting multidimensional force vector data and reconstructing a dynamic force field, and then adaptively and predictively adjusting the adsorption force field based on the reconstruction results.

[0018] Figure 5 This is a flowchart illustrating how the present invention performs a controlled release operation after the task is completed, and conducts system performance evaluation and parameter self-evolution. Detailed Implementation

[0019] 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.

[0020] Example 1

[0021] like Figure 1 As shown, this embodiment provides an overall framework for an intelligent control method of automatically adsorbing forklifts in a cold storage shuttle. Specifically, the method of this embodiment includes the following steps:

[0022] S1: Collect interface state information through multimodal sensing data, and construct a digital twin model of contact mechanics based on the interface state information.

[0023] According to one aspect of this application, this embodiment constructs a high-fidelity, computable virtual image of the surface of the goods that the fork ruler will contact and the frost medium attached to it. Its core lies in the deep fusion of multi-source heterogeneous data and the online identification of the physical model. Specifically, the system first simultaneously activates four parallel sensing channels. The first channel is a visual and geometric perception channel: using an active short-focal-length polarized light microscope camera deployed on the side edge of the fork ruler's adsorption surface, it acquires a sequence of original microscopic morphology images with a resolution of up to 1024x768 pixels at a rate of 30 frames per second. The introduction of polarized light effectively suppresses specular reflection, highlighting the ice crystal boundaries and frost texture. The second channel is a thermal perception channel: using a thin-film platinum resistance temperature sensor array embedded in a 5x5 grid under the composite material of the adsorption surface, it simultaneously acquires real-time temperature readings at a sampling frequency of 100Hz for a total of 25 measuring points in the contact area, thereby forming a dense two-dimensional temperature field raw data matrix. The third channel is the mechanical sensing channel: a miniature MEMS six-dimensional force / torque sensor is installed at each of the four load-bearing points connecting the fork to the main structure. At the moment of initial contact (pre-compression) between the fork and the cargo surface, the original triaxial force and triaxial torque voltage signals at each support point are acquired at a frequency of 1kHz. The fourth channel is the dielectric sensing channel: the adsorption panel itself is used as the electrode of a capacitor. A high-frequency capacitance measurement circuit acquires a dielectric constant response signal that alternates at a frequency of 10kHz. This signal is extremely sensitive to the dielectric properties of the aqueous phase at the interface.

[0024] Furthermore, real-time processing and feature extraction are performed on the multimodal raw data. First, spatiotemporal synchronization alignment based on high-precision hardware timestamps is performed on all data streams to ensure spatiotemporal consistency in subsequent fusion analysis. For the original micro-morphology image sequence, key geometric features are extracted using real-time digital image correlation algorithms and edge detection operators, including but not limited to the average ice crystal size distribution vector, frost layer thickness estimation map, surface macro-roughness Ra value, and the orientation and length data of major cracks. For the two-dimensional temperature field raw data matrix, a spatial interpolation algorithm is used to generate continuous temperature contour maps and calculate its temperature gradient vector field, thereby retrieving the transient heat flux density distribution cloud map of the contact interface. For the original voltage signal of the mechanical channel, the force and torque vectors of each support point are obtained after calibration matrix transformation. Through static model fusion, the three-dimensional net external force and net torque acting on the center of the adsorption surface are calculated, and a fast Fourier transform is performed to extract the micro-vibration energy spectrum in the 0-500Hz frequency band, which is used to preliminarily estimate the equivalent stiffness coefficient and damping coefficient of the interface contact. For the dielectric constant response signal, by solving the offset of its amplitude and phase relative to the known dry state, and inputting a pre-calibrated mixed dielectric model, the volume percentage of liquid water in the interface layer is estimated.

[0025] A further improvement to this embodiment involves mapping the extracted static and dynamic features into a parameterized multiphysics simulation model. This model is a finite element model that couples heat conduction, elasticity, and interfacial fracture mechanics. The system uses feature vectors such as ice crystal size distribution, frost thickness, temperature gradient, equivalent stiffness, and liquid water content as input to a lightweight convolutional neural network. This network, trained on massive amounts of offline simulation data, can output in real-time the most critical and time-varying material property parameters in the digital twin model under the current interface conditions, such as the equivalent elastic modulus of the ice-frost mixture, the calibration coefficient of the interfacial adhesion energy, and the friction coefficient of the ice-metal interface. Finally, the system uses these dynamically identified parameters to instantiate and update the basic finite element model online, resulting in a real-time contact mechanics digital twin model that is highly consistent with the current state of the physical world and can predict mechanical responses under different loading conditions. This model is the virtual heart of the entire intelligent control system, providing predictive capabilities for subsequent planning.

[0026] S2: Based on the digital twin model, perform dynamic planning of anchor points to generate the optimal anchor point coordinate set that resists dynamic interference.

[0027] In this embodiment, the real-time contact mechanics digital twin model constructed in S1 serves as an efficient virtual testing platform to solve for the globally optimal or suboptimal mechanical anchoring strategy, taking into account future dynamic disturbances. When the planning process begins, the system first needs to collect multi-source prediction and real-time data related to the future dynamic environment. This data mainly includes two categories: the first is forward-looking data from the vehicle motion control system, such as obtaining the predetermined acceleration curve of the shuttle and the curvature radius sequence of the planned path within the next 3 seconds via the CAN bus; the second is real-time disturbance data from the environment and the shelving, such as collecting the current high-frequency micro-vibration spectrum through the IMU sensor on the forklift and retrieving the inherent vibration mode parameters of the current shelving segment from the warehouse management system.

[0028] Specifically, the planning engine transforms this dynamic data into boundary conditions that can be loaded onto the digital twin model. For example, the predetermined acceleration curve is decomposed into inertial force loads acting on the centroid of the fork-cargo system; the path curvature is transformed into centrifugal force loads; and the vibration spectrum is applied as the basic excitation load to the model boundary. Furthermore, the planning problem is formalized as a constrained multi-objective optimization problem. In a preferred embodiment, the optimization objective is typically set as follows: under the action of a predetermined dynamic load spectrum, maximizing the minimum safety factor of the entire adsorption interface, while simultaneously homogenizing the force on each anchor point as much as possible, and minimizing total energy consumption. Constraints include: anchor points must avoid packaging weak areas (coordinate restricted areas) identified by S1 image analysis; the distance between any two anchor points must be greater than a minimum threshold (e.g., 15 mm) to prevent excessive overlap of stress fields; and the total number of activated phase change anchoring units is limited by the instantaneous power supply.

[0029] In a preferred embodiment, the system employs a hybrid heuristic optimization algorithm to solve this complex problem, treating the digital twin model as a black-box function. First, within the feasible region of the adsorption surface, hundreds of potential anchor point coordinate combinations (i.e., a population) are randomly generated. For each scheme in the population, the planning engine drives the digital twin model to perform a rapid transient dynamic simulation, calculating the location of the interface most prone to failure and its safety factor under the aforementioned dynamic load, and statistically analyzing the coefficient of variation of the force distribution. Based on these simulation results, the fitness of each scheme is evaluated. Next, by simulating selection, crossover, and mutation operations in biological evolution, a new generation of schemes is generated, and simulation evaluation is performed iteratively. After dozens of iterations, the algorithm converges to one or more Pareto optimal solutions. The central controller ultimately selects a scheme from these optimal solutions that achieves the best balance between safety and energy consumption, outputting a set of optimal anchor point coordinates tailored to the task to resist dynamic interference. This coordinate set is a dynamic, non-uniform spatial lattice, and its distribution strategy directly reflects the pre-countermeasure design against the expected interference mode.

[0030] S3: Perform phase change anchoring operation based on the anchoring point coordinate set, and monitor the anchoring process status in real time.

[0031] This embodiment transforms the virtual plan calculated in S2 into a substantial mechanical connection in the physical world. The key lies in the precise, closed-loop control of the phase change process. After receiving the optimal anchoring point coordinate set from S2, the control system first compares it with the physical position mapping table of the phase change anchoring actuator array on the adsorption surface. Each actuator unit consists of a linear motor with micron-level precision, a miniature thermoelectric cooler, and a conical metal anchor pin embedded in the cold end of the cooler. The system drives the linear motor to ensure that the tip of the anchor pin corresponding to the target coordinate precisely protrudes from the adsorption surface, pre-contacts the cargo surface, and applies a constant, small pre-pressure, for example, 5 Newtons.

[0032] In one alternative implementation, the core of the anchoring operation is the immediate initiation of phase change locking. A central controller sends a precisely designed multi-stage current pulse sequence to a miniature thermoelectric cooler corresponding to the target anchoring point. This sequence typically includes a high-intensity rapid freezing pulse (e.g., 10 amps, lasting 100 milliseconds) designed to rapidly drop the temperature of the anchor tip and its surrounding micro-area (approximately 2 mm in diameter) from ambient temperature (e.g., -25°C) to -50°C or even lower within a very short time (<200 milliseconds). This drastic supercooling causes instantaneous and deep freezing between the anchor and the frost layer on the cargo surface, as well as within the frost layer itself, forming an ice bridge with extremely high mechanical strength. Subsequently, the current transitions to a lower-intensity sustaining pulse, stabilizing the temperature of the micro-area at approximately -35°C to maintain the stability of the ice bridge.

[0033] Real-time monitoring is crucial in this process. Each anchoring unit integrates a miniature thermocouple and circuitry for measuring the insulation resistance between the anchor pin and the substrate, acquiring real-time temperature transient curve T(t) and contact resistance change curve R(t) at a frequency of 1 kHz. These two curves are transmitted to a state analyzer in real time. The state analyzer compares T(t) with a preset ideal cooling curve and evaluates the integrity and speed of the phase transition process by calculating characteristic parameters such as the area under the curve and the time to reach the target temperature. Simultaneously, the steep drop and stable value of R(t) directly reflect the formation quality and conductivity of the ice bridge. Through a lightweight decision logic (such as a rule-based classifier or a miniature neural network), the system can diagnose the state of the anchoring point in real time, for example: successful anchoring - high strength, successful anchoring - medium strength, delayed anchoring, or anchoring failure (such as poor contact). The process state diagnostic labels and process curve characteristic values ​​of all anchoring points are summarized to form a comprehensive anchoring process status report. This report is not only used to confirm the immediate success or failure of the current operation, but more importantly, it serves as a priori knowledge input for the credibility and initial strength of each anchor point when performing force field adaptive adjustment in the next stage, thereby achieving information closure between the preceding and following stages.

[0034] S4: Collect multidimensional force vector data and perform dynamic force field reconstruction, and perform adaptive and predictive adjustment of the adsorption force field based on the reconstruction results.

[0035] Once all planned anchor points have completed initial phase transition locking, the system enters a high-frequency operation phase characterized by fine-tuning and active anti-interference based on force feedback. The core of this embodiment is to transform passive absorption into actively shaping the ideal adsorption force field. At this point, the four six-dimensional force sensors, acting as the system's tactile nerves, begin synchronously acquiring raw, unprocessed multi-channel voltage signals at a frequency of at least 500Hz. These signals, after passing through their respective 24-bit analog-to-digital converters, form six sets of high-frequency digital signal streams (Fx, Fy, Fz, Mx, My, Mz).

[0036] Specifically, the central controller first performs a series of preprocessing steps on these raw signal streams: including unit conversion based on calibration certificates, converting voltage values ​​into force and torque values; performing a coordinate system transformation, converting the measurements of all fulcrums to a global coordinate system with the center of the adsorption surface as the origin; and applying a low-pass digital filter to filter out high-frequency noise, retaining low-frequency force signals related to vehicle motion and load, as well as mid-frequency signals that may be caused by vibration. The preprocessed, spatiotemporally aligned multidimensional force / torque time series data forms the basis for force field reconstruction.

[0037] According to a further improvement in this embodiment, dynamic force field reconstruction is a key inverse problem-solving process. The controller has the following known information: 1) the global net external force and net torque acting on the system (calculated by fusing measurements from four sensors); 2) the spatial coordinates of all effective anchor points (from the confirmation lists of S2 and S3). What needs to be solved is the unknown, dynamic three-dimensional contact force vector at each anchor point. The system employs an iterative algorithm based on weighted least squares estimation and Tikhonov regularization to solve this underdetermined problem. This algorithm uses the global force balance equation and torque balance equation as basic constraints, minimizes the weighted sum of squares of forces at all anchor points (preferring a more uniform and smoother force distribution solution), and utilizes the anchor point process state diagnostic labels obtained in S3 as prior weights (e.g., higher strength anchor points can bear greater forces), to invert a discrete three-dimensional force vector distribution map covering all anchor points in real time, i.e., the reconstructed dynamic adsorption force field F(x, y, t).

[0038] Based on this reconstructed force field, the system initiates adaptive and predictive adjustments. For adaptive adjustment: the algorithm analyzes F(x, y, t) in real time, identifying stress hotspots (anchor points experiencing shear or peeling forces close to preset thresholds) and stress voids (anchor points experiencing forces far below the average). For stress hotspots, the controller can send a fine-tuned enhancement pulse to the corresponding thermoelectric cooler, further reducing its local temperature by 2-3°C to instantaneously increase the shear strength of the ice bridge at that point; or slightly adjust the force direction of the surrounding anchor points. For stress voids, their maintaining power can be appropriately reduced, allowing for a slight increase in the force they bear, thus making the force field distribution more uniform. For predictive adjustment: the system continuously receives motion commands from the vehicle controller for the next few hundred milliseconds. Using a simplified vehicle-load dynamics model, it predicts the distribution pattern of the additional force field generated by the upcoming inertial load (such as centrifugal force at the start of a turn) on the adsorption force field. The controller proactively provides a small-scale strength enhancement to the anchor point group on the side that will bear tension in this additional force field. This combined feedforward and feedback control strategy makes the entire adsorption system act like a resilient smart biofilm, capable of actively adapting to and counteracting external dynamic disturbances, thus firmly locking the cargo in place.

[0039] S5: After the task is completed, perform a controlled release operation and conduct system performance evaluation and parameter self-evolution.

[0040] This embodiment is used to continuously improve system performance through experience learning. Upon receiving a release command from the upper-level system, the controller initiates a controlled release operation. This operation is not a simple power cut-off, but a reverse, precise phase change control process. Specifically, the controller sends a reverse heating release pulse sequence sequentially or in groups to all active anchor points. This sequence begins with a slowly rising ramp current, causing the anchor tip temperature to rise at a controllable rate (e.g., 10°C / second) from the operating temperature (-35°C) to slightly below the freezing point (e.g., -2°C). During this stage, the six-dimensional force sensor continuously monitors the attenuation of the constraint forces in each direction. When the total adsorption force is detected to drop below 20% of the initial value, the system determines that the ice bridge has softened sufficiently but has not completely melted. A brief, strong final release pulse is then applied, causing the temperature to instantly jump to above 0°C (e.g., +5°C), completely breaking the remaining phase change connection and achieving complete and smooth separation of the cargo from the forklift, while avoiding excessive condensation due to rapid melting.

[0041] In a preferred embodiment, after release, the system immediately enters the offline analysis and learning phase, namely system performance evaluation and parameter self-evolution. The system automatically packages the complete data chain of this task from S1 to S5, including the initial interface feature vector, planning scheme, anchoring process curve and diagnostic results, reconstructed force field time series data, adjustment command records, and the final task success indicators (such as whether slippage occurred, release time, etc.), into a structured task instance data package and stores it in the historical database.

[0042] An evolutionary learning engine running in the background periodically analyzes these historical data packets. At the core of this engine is a meta-optimizer, which, for example, might employ a Bayesian optimization framework. It defines a set of hyperparameters to be optimized, including weights of the feature extraction network in S1, cost function weights in the planning algorithm in S2, the threshold for successful anchoring in S3, and regularization parameters for force field reconstruction in S4. Each successful task instance is considered an experiment, with the input being the current operating conditions (summarized from initial feature vectors and dynamic disturbance data), and the output being a task performance score (calculated comprehensively from safety margin, energy consumption, and time). The goal of the evolutionary learning engine is to find a set of hyperparameters that maximizes the system's expected performance score under various operating conditions. It achieves this by establishing a probabilistic surrogate model between these hyperparameters and the performance score, and intelligently proposing new parameter combinations for virtual testing (based on historical data simulation or triggering new real-world tasks), continuously iterating and optimizing. Finally, when a new configuration significantly superior to the current parameters is discovered, the learning engine generates a parameter update proposal, which, after security review, is updated to the corresponding module in the main control system. Through this continuous, data-driven self-evolution mechanism, the entire intelligent adsorption system can gradually adapt to more complex cargo types, more severe environmental conditions, and unforeseen dynamic scenarios, becoming smarter with use. Its control strategy is continuously refined over time, which is the core guarantee for the long-term robustness and superiority of this solution.

[0043] Example 2

[0044] like Figures 1 to 5 As shown, this embodiment, based on Embodiment 1, describes in detail the data processing flow of an intelligent control method for automatically adsorbing forklifts on a cold storage shuttle, specifically including:

[0045] S1: Collect interface state information through multimodal sensing data, and construct a digital twin model of contact mechanics based on the interface state information, specifically including:

[0046] S1.1: Synchronous acquisition of multimodal sensor data and preprocessing of raw data.

[0047] In this embodiment, to achieve comprehensive perception of the interface state, the system first initiates a strictly synchronized multi-channel data acquisition process. Specifically, the central controller sends a unified synchronous trigger pulse to all sensing units to ensure data alignment in time base. In the visual channel, the active polarized light microscope camera acquires one or more consecutive frames of raw polarized images upon triggering, and transmits the raw RAW format data to the image processing unit via a high-speed interface. In the thermal channel, the analog signals from the temperature sensor array are synchronously sampled by a multiplexed ADC to obtain a set of raw temperature and voltage value arrays containing 25 channels. In the mechanical channel, the analog outputs of four six-dimensional force sensors are acquired in parallel by a synchronously sampled ADC to obtain four sets of six-channel raw force / torque voltage signal arrays. In the dielectric channel, the AC response signal output from the capacitance measurement circuit is demodulated by a lock-in amplifier to obtain raw data of in-phase (I) and quadrature (Q) components characterizing dielectric properties. After acquisition, all raw data are stamped with a unified high-precision microsecond-level timestamp and encapsulated into a timestamped multimodal raw data packet, stored in a circular buffer for subsequent processing.

[0048] S1.2: Parallel extraction and fusion of multi-source heterogeneous features.

[0049] After obtaining the time-aligned raw data packets, the system initiates a parallel feature extraction pipeline. Specifically, for visual data, the original polarized image is first de-mosaiced, white-balanced, and denoised to obtain a clear RGB image. Then, two images with different polarization angles are differentially analyzed to extract an enhanced texture image that suppresses specular highlights. This enhanced image is then processed using local binarization and contour detection algorithms to calculate the percentage of ice crystal coverage area and the average equivalent diameter of ice crystals. The frost thickness distribution is estimated using stereo vision principles (if using dual cameras) or photometric stereo methods. For thermal data, the original temperature and voltage value array is converted to Celsius temperature values ​​using a calibration formula, and a temperature field distribution image is generated using bilinear interpolation. The gradients of this temperature field in the X and Y directions are calculated to obtain a two-dimensional temperature gradient vector field matrix representing the direction of heat flow. For the mechanical data, the four sets of six-channel raw voltage signals are transformed using their respective calibration matrices to obtain the force vectors (Fx_i, Fy_i, Fz_i) and torque vectors (Mx_i, My_i, Mz_i) of the four pivot points in the global coordinate system. Through static resultant force calculation, the total force vector (F_total) and total torque vector (M_total) acting on the center of the adsorption surface are fused together. Simultaneously, a short-time Fourier transform is performed on the force signals to extract the dominant vibrational frequency components and their amplitudes with prominent energy in the 0-200Hz frequency band. For the dielectric data, the changes in the real and imaginary parts of the dielectric constant are calculated based on the in-phase (I) and quadrature (Q) components. These changes are input into a pre-trained model based on the Maxwell-Garnett mixed medium theory, outputting the estimated liquid water volume fraction (Φ_w) of the current interface layer. Finally, all extracted features are normalized and combined into a structured multi-dimensional feature vector, which comprehensively describes the geometric, thermal, mechanical, and dielectric states of the interface.

[0050] S1.3: Dynamic parameter identification and digital twin model instantiation based on neural networks.

[0051] In this embodiment, the system maintains a parameterized basic finite element model template, which includes the basic geometry and material definitions of the fork, cargo packaging, and frost layer. However, its key material properties (such as the elastic modulus E, Poisson's ratio ν, and interfacial adhesion strength σ_bond of the frost layer) are set as variables to be identified. According to a further improvement of this embodiment, a lightweight deep neural network is used to complete this identification task. This network takes multi-dimensional feature vectors as input. Specifically, the network's input layer receives dozens of feature scalars, including ice crystal size, frost layer thickness, mean temperature gradient, dominant vibration frequency, and liquid water content. After a nonlinear transformation through 3-5 fully connected hidden layers, the network output layer directly predicts the optimal estimates of those time-varying parameters in the basic finite element model template, such as the equivalent elastic modulus E_eff and the interfacial strength coefficient k_bond. These predicted parameters are injected into the basic finite element model template in real time, thereby generating a personalized, instantiated digital twin model of contact mechanics that perfectly matches the current physical interface state. This model is ready to be used for simulating loading and predicting failure.

[0052] According to another aspect of this application, the parallel extraction and fusion of multi-source heterogeneous features includes:

[0053] S1.2.1: Visual feature extraction and enhancement processing.

[0054] In this embodiment, the acquired original polarization image sequence is first preprocessed. Specifically, the system performs demosaicing, adaptive white balance, and nonlocal mean denoising on two images with different polarization angles (e.g., 0° and 90°) to obtain two clear RGB images. Further, pixel-level difference operations are performed on these two images to obtain an enhanced texture image with significantly suppressed specular reflection components. Subsequently, an adaptive threshold segmentation algorithm is applied to this enhanced image to separate the ice crystal region from the background. Then, morphological opening operations are used to remove noise points, and connected component analysis is used to calculate the area, perimeter, and principal axis direction of each ice crystal region. The system statistically analyzes the area distribution of all ice crystal regions to obtain the percentage of ice crystal coverage, the average equivalent diameter of ice crystals, and a histogram of ice crystal orientation distribution. Simultaneously, using a binocular stereo vision matching algorithm (if using dual cameras) or photometric stereo (if using multi-angle illumination), the depth map of the frost surface is recovered based on the parallax or brightness variations of pixels. Then, through plane fitting and difference calculation, a frost thickness distribution map in micrometers is obtained. All of the above visual feature data are organized into structured visual feature vectors.

[0055] S1.2.2: Thermal feature extraction and gradient field calculation.

[0056] For thermal data, the system first converts the raw temperature and voltage value arrays of 25 measuring points into Celsius temperature values ​​using calibration coefficients (each sensor calibrated independently) stored in non-volatile memory, resulting in a temperature value array T_array. Then, using a bilinear interpolation algorithm, the discrete 25-point temperature data is expanded into a pseudo-color temperature field distribution image with a resolution of 100x100 pixels, covering the entire adsorption region. Based on this temperature field image, the system uses the Sobel operator to calculate the temperature gradients in the X and Y directions, i.e., G_x = dT / dx, G_y = dT / dy, thus obtaining a two-dimensional temperature gradient vector field matrix. Further, the system calculates the divergence of this gradient field to assess the convergence or divergence of local heat flow, and calculates the average and variance of the gradient amplitude as indicators of heat exchange uniformity. These calculated gradient field data, divergence maps, and statistical indicators together constitute the thermal feature vector.

[0057] S1.2.3: Mechanical feature extraction and spectrum analysis.

[0058] For the raw voltage signal arrays acquired by four six-dimensional force sensors, the system first decouples the voltage values ​​using the 6×6 calibration matrix of each sensor (obtained through laboratory calibration), converting the voltage values ​​into corresponding force and torque values. Specifically, for the i-th sensor, its output six-dimensional vector V_i is multiplied by matrix multiplication F_i = C_i × V_i (where C_i is the calibration matrix) to obtain the force and torque in the sensor's coordinate system. Then, using a pre-calibrated installation attitude matrix, the force / torque values ​​of all sensors are transformed to a unified global coordinate system. Next, the system adds the force vectors of the four sensors to obtain the total force vector F_total, and adds the torque vectors plus the additional torque generated by the lever arm to obtain the total torque vector M_total. Simultaneously, to extract dynamic characteristics, the system performs a windowed short-time Fourier transform on the force signal of each sensor (especially in the Z-axis direction), with a window length of 256 sampling points and 50% overlap. The system extracts the three highest-energy frequency components and their amplitudes from the spectrum and calculates the energy proportion of the 0-50Hz low-frequency band. These frequency characteristics, together with the total force / torque, constitute the mechanical characteristic vector.

[0059] S1.2.4: Dielectric feature extraction and liquid water content inversion.

[0060] For the dielectric channel, the system first performs a moving average filter on the acquired raw data of the in-phase (I) and quadrature (Q) components to suppress random noise. Then, its amplitude A = sqrt(I) is calculated. 2 +Q 2The amplitude and phase are φ = atan2(Q, I). The current amplitude and phase are compared with reference values ​​measured on a dry, clean surface to obtain the relative changes ΔA and Δφ. According to a further improvement of this embodiment, the system maintains a lookup table based on mixed-medium theory, which establishes the mapping relationship between ΔA and Δφ for different ice-water mixing ratios and at different temperatures. By inputting the currently measured ΔA and Δφ and the current average temperature into this lookup table, and using bilinear interpolation, the estimated liquid water volume fraction Φ_w of the current interface layer is retrieved. Furthermore, the system calculates the standard deviations of the I and Q components during the measurement period as an indicator of the dielectric signal stability. These data collectively constitute the dielectric eigenvector.

[0061] S1.2.5: Multi-feature vector fusion and normalization.

[0062] Finally, the system concatenates the visual, thermal, mechanical, and dielectric feature vectors extracted from the four channels to form a high-dimensional original multi-dimensional feature vector. To eliminate the influence of different feature dimensions and numerical ranges, the system uses offline calculated mean and standard deviation to perform Z-score standardization on the original vector, i.e., subtracting the mean from each feature value and dividing by the standard deviation. The standardized feature vector is organized into a fixed-format array, labeled with a timestamp and task ID, and stored in a shared memory area for use by the neural network in step S1.3.

[0063] S2: Based on the digital twin model, dynamic programming of anchor points is performed to generate an optimal set of anchor point coordinates resistant to dynamic disturbances, specifically including:

[0064] S2.1: Quantitative modeling and loading spectrum generation of dynamic disturbance loads.

[0065] Before planning begins, the potential mechanical disturbances must be quantitatively described. Specifically, the system acquires the planned velocity curve v(t) and the planned path curvature κ(s) within a future time window (e.g., 3 seconds) from the vehicle control system in real time. Based on Newton's second law and the formula for circular motion, the resulting inertial force load spectrum F_inertial(t) and centrifugal force load spectrum F_centrifugal(t) acting on the cargo's center of mass are calculated. Simultaneously, the main environmental vibration excitation frequencies f_env and their estimated amplitudes A_env are extracted from the shelf database and the current IMU spectrum. All this load information is integrated and arranged in a time series into a multi-channel dynamic load input spectrum, which defines the time-varying boundary conditions that the digital twin model must withstand within the planning time window.

[0066] S2.2: Optimization solution for anchor point layout based on digital twin simulation.

[0067] In a preferred embodiment, the system employs a simulation-based optimization process. The workflow of the optimization algorithm (such as a genetic algorithm) is as follows: First, the algorithm generates a random initial population, where each individual represents a possible combination of anchor point coordinates. For each individual in the population, the planning engine performs the following operations: 1) In the digital twin model, based on the individual's coordinates, activate the corresponding node constraints to simulate anchoring; 2) Apply the dynamic load input spectrum as an external load to the model; 3) Run transient dynamic simulations to calculate the stress-strain response of the entire adsorption interface under the entire load spectrum; 4) Extract key indicators from the simulation results, such as the maximum principal stress peak, the maximum interface peeling energy, and the stress distribution uniformity index. These indicators are combined into an objective function value to evaluate the merits of the layout scheme. Based on the evaluation results of all individuals, the optimization algorithm performs selection, crossover, and mutation to generate a new generation of population and iterates the above simulation evaluation process. After multiple generations of evolution, the algorithm converges to one or more schemes that perform well on the objective function.

[0068] S2.3: Post-processing of the optimal solution for anti-interference and output of coordinate set.

[0069] After the optimization algorithm converges, the optimal or Pareto front solution set needs post-processing to ensure its engineering feasibility. First, the system performs collision detection on the anchor point coordinates in the solution set based on the weak zone coordinate exclusion map, removing or fine-tuning points that fall into the exclusion zone. Second, it checks the Euclidean distance between any two points to ensure it is greater than a preset minimum spacing threshold, avoiding excessive stress concentration. Finally, it evaluates whether the number of simultaneously activated anchor points required by the scheme exceeds the system's instantaneous power limit; if so, it removes them according to their importance. Through this post-processing, the system selects a comprehensively optimal scheme from the solution set and formats the coordinates of all its anchor points into a clearly structured final anchor point coordinate instruction set containing coordinate numbers, X, Y, and Z values, and expected priorities. This instruction set can directly drive subsequent actuators.

[0070] According to another aspect of this application, the anchor point layout optimization solution based on digital twin simulation includes:

[0071] S2.2.1: Optimization problem initialization and population generation.

[0072] In this embodiment, the optimization algorithm first defines the search space based on the effective area of ​​the adsorption surface and the minimum anchor point spacing constraint. Specifically, the adsorption surface is modeled as a rectangular region, the coordinate range of which is determined by the physical dimensions. The algorithm randomly generates an initial population containing N individuals (e.g., N=100). Each individual is represented by a binary matrix or real number vector encoding. For example, an individual can be a binary string of length M (M is the total number of selectable anchor point positions), where 1 indicates that an anchor point is placed at that position, and 0 indicates that no anchor point is placed. In an optional implementation, to obtain finer position control, the (x, y) coordinates of each anchor point can also be directly represented by real number encoding. When generating the population, it is necessary to ensure that each individual meets basic constraints, such as the number of anchor points being within a preset range and the distance between any two points being greater than a minimum threshold.

[0073] S2.2.2: Individual Decoding and Digital Twin Model Instantiation.

[0074] For each individual in the population, the optimization engine needs to decode it into a specific set of anchor point coordinates. Taking binary encoding as an example, the decoder converts the bits that are 1 in the binary string into corresponding (x, y) coordinates according to a predefined position mapping table. Then, the system calls the constructed instantiated contact mechanics digital twin model and updates the model with these coordinates as additional boundary conditions (e.g., applying fixed constraints or couplings at these coordinate points), thereby obtaining a personalized digital twin model instance tailored to that individual for simulation testing.

[0075] S2.2.3: Dynamic load spectrum application and transient dynamic simulation.

[0076] For each personalized digital twin model instance, the optimization engine applies a dynamic load input spectrum as an external load condition to the model. Specifically, the inertial and centrifugal forces in the load spectrum are converted into time-varying force loads acting on the cargo's center of mass, while vibration excitation is applied as a basic acceleration excitation to the model boundaries. Subsequently, the transient dynamics solver is activated to simulate the model's stress, strain, and displacement responses over the entire load spectrum duration (e.g., 3 seconds). The simulation employs an explicit integration algorithm to ensure computational speed, with the time step automatically determined based on the model's minimum feature size. During the simulation, the system monitors the stress distribution at the interfaces in real time, particularly the normal and shear stresses.

[0077] S2.2.4: Post-processing of simulation results and fitness calculation.

[0078] After the simulation, the system performs post-processing on the results to extract key performance indicators. First, it extracts the maximum principal stress σ_max and maximum shear stress τ_max for all element nodes at the interface throughout the entire simulation time from the results file. The peak values ​​of these maximum values ​​are calculated, and their locations and times are recorded. Second, it calculates the stress distribution uniformity of all nodes at the interface at the end of the simulation, for example, by calculating the coefficient of variation (standard deviation / mean) of the normal stress. Further, based on material failure criteria (e.g., the maximum tensile stress criterion and the Mohr-Coulomb criterion), it calculates the safety factor for each node at each time step and identifies the minimum safety factor for the entire interface and the entire time window. Finally, it weights the minimum safety factor (the larger the better), stress distribution uniformity (the smaller the coefficient of variation the better), and the number of anchor points (the fewer the better) to calculate the fitness value for that individual. The weighting coefficients can be dynamically adjusted according to task priority; for example, under extreme conditions, the weight of the safety factor can be set to the highest.

[0079] S2.2.5: Evolutionary Operations and Iterative Optimization.

[0080] Based on the fitness values ​​of all individuals, the optimization algorithm performs evolutionary operations to generate a new generation of population. First, a roulette wheel selection method is used, where individuals with higher fitness have a higher probability of being selected for the mating pool. Then, single-point crossover is performed on the selected individuals; that is, a crossover point is randomly chosen, and the partial codes after the crossover point of two individuals are swapped, generating two new individuals. Next, individuals are mutated with a small probability, for example, by randomly flipping a bit in a binary string (binary encoding) or making a small perturbation to the real number coordinates (real number encoding). The newly generated population will repeat processes S2.2.2 to S2.2.4 for simulation and evaluation. This iterative process continues until a preset maximum number of generations (e.g., 50 generations) is reached or the fitness value no longer significantly increases over several consecutive generations. Finally, the algorithm selects the individual with the highest fitness from the last generation of the population as the optimal solution.

[0081] S3: Perform phase change anchoring operation based on the anchoring point coordinate set, and monitor the anchoring process status in real time, specifically including:

[0082] S3.1: Actuator precise positioning and phase change anchoring energy pulse application.

[0083] After parsing the anchoring point coordinate instruction set, the control system first converts it into motion commands for a two-dimensional high-precision linear motor platform. This platform moves the entire adsorption surface (or a local array), precisely aligning each target coordinate point specified in the command with the physical center of a phase change anchoring actuator. Once aligned, the control system drives a miniature linear actuator within the unit, causing the conical metal anchor pin to extend at a controllable speed until it contacts the cargo surface and reaches a preset micro-contact pre-pressure value (e.g., 5N). This pressure is confirmed by feedback from a miniature force sensor integrated within the actuator. Subsequently, the core phase change anchoring stage begins: the temperature control unit applies a pre-optimized multi-stage current pulse waveform to the miniature thermoelectric cooler at that point. This waveform typically includes a high-amplitude pulse for rapid thermal barrier breaking, a platform pulse for deep freezing to form an ice bridge, and a low-power steady-state current for maintenance. This process is executed sequentially or in parallel at each anchoring point.

[0084] S3.2: Multi-parameter synchronous monitoring and state feature extraction during the anchoring process.

[0085] Simultaneously with the application of energy pulses, high-frequency process monitoring is initiated. Embedded sensors in each anchoring unit acquire changes in two key physical quantities at a frequency of at least 1 kHz: 1) the anchor tip temperature T(t), measured by a miniature thermocouple; and 2) the insulation resistance R(t) between the anchor and the substrate. These two raw signals are transmitted to a process analyzer in real time. The process analyzer first filters the signals to remove noise. For the temperature curve T(t), the analyzer calculates its descent slope dT / dt_max, the time t_reach to reach the target temperature (e.g., -35°C), and dynamic characteristics such as overshoot amplitude. For the resistance curve R(t), the analyzer identifies the starting point of its steep drop from a high-resistance (insulation) state and its final stable value R_final. These extracted dynamic characteristic values ​​collectively constitute the process feature vector describing the anchoring process.

[0086] S3.3: Real-time anchored quality diagnosis and classification based on process characteristics.

[0087] The extracted process feature vectors are immediately fed into a real-time diagnostic classifier. This classifier can be a rule-based expert system or a lightweight machine learning model (such as a support vector machine or decision tree). The classifier compares and matches the current process features with a pre-stored feature library of the gold anchoring process, built up through extensive experiments. Based on the comparison results, the diagnosticator outputs a diagnostic label for the current anchoring point state, such as: "Class A: Optimal Anchoring" (features highly match the gold process), "Class B: Acceptable Anchoring" (some features are slightly off, but the strength is expected to be sufficient), "Class C: Weak Anchoring" (insufficient cooling or insignificant resistance decrease), or "Class D: Anchoring Failure" (no temperature response or no change in resistance). Simultaneously, the diagnosticator also outputs a confidence score predicting the relative anchoring strength of the anchoring point based on the process features. All the diagnostic labels and confidence scores for the anchoring points are aggregated to form a comprehensive anchoring process status report, serving as crucial prior knowledge for the next stage of force field adjustment.

[0088] According to another aspect of this application, the anchoring process includes multi-parameter synchronous monitoring and state feature extraction, comprising:

[0089] S3.2.1: Synchronous acquisition and buffering of multi-channel raw signals.

[0090] The monitoring system starts simultaneously with the application of the phase-change anchoring energy pulse. The temperature sensor and resistance measurement circuitry of each anchoring unit operate synchronously at a sampling frequency of 1 kHz. The central data acquisition card uses a multiplexer to cyclically acquire the raw temperature voltage signal V_T(t) and raw resistance voltage signal V_R(t) of all active anchoring points. These signals are buffered in real-time into a first-in-first-out buffer of length L (e.g., L = 5000, corresponding to 5 seconds of data) and aligned with the global system clock to ensure accurate timestamps for each sampling point.

[0091] S3.2.2: Signal preprocessing and feature time point detection.

[0092] For the temperature signal at each anchor point, a moving average filter is first applied to remove high-frequency noise. Then, by searching for the zero-crossing point of the first derivative of the filtered signal, the starting time point t_start of the temperature drop is detected. Similarly, for the resistance signal, its logarithm is first calculated to expand the dynamic range, and then a threshold method is applied: when the resistance value drops from an initial high resistance (e.g., >10MΩ) by more than a preset proportion (e.g., 50%), that moment is marked as the starting time point t_drop of the resistance drop. These key time points will be used for subsequent feature calculations.

[0093] S3.2.3: Dynamic feature parameter calculation and vector construction.

[0094] Based on the preprocessed signal and the detected time points, a set of dynamic characteristic parameters are calculated. For the temperature signal: calculate the maximum slope S_max = max(dT / dt) for the fastest temperature drop starting from t_start; calculate the time span Δt_cool required for the temperature to drop from the initial value to the target temperature (e.g., -35°C); calculate the possible overshoot amplitude after the temperature reaches its lowest point. For the resistance signal: calculate the time span Δt_settle for the resistance to drop to a stable value (e.g., <10kΩ) starting from t_drop; calculate the final stable value R_final of the resistance; calculate the average slope K_R of the logarithmic resistance versus time curve during the resistance drop. All calculated characteristic parameters (S_max, Δt_cool, Overshoot, Δt_settle, R_final, K_R), together with the key time points (t_start, t_drop), constitute the process characteristic vector of this anchor point.

[0095] S3.2.4: Eigenvector normalization and outlier handling.

[0096] Since different anchor points may have different feature value ranges due to slight differences in contact, the feature vectors need to be normalized to facilitate subsequent classification. In this embodiment, minimum-maximum value normalization is used to scale each feature parameter to the [0,1] interval. The minimum and maximum values ​​used for normalization are obtained based on a large amount of historical data. At the same time, the system checks whether there are obvious outliers in the feature vector (e.g., Δt_cool is negative or extremely large). If an outlier is found, the feature vector is marked as suspicious and given special treatment in subsequent classification, such as being directly classified as anchoring failure.

[0097] S4: Collect multidimensional force vector data and perform dynamic force field reconstruction. Based on the reconstruction results, perform adaptive and predictive adjustment of the adsorption force field, specifically including:

[0098] S4.1: Global multidimensional force data fusion and dynamic load decoupling.

[0099] During the closed-loop regulation phase, four six-dimensional force sensors continuously provide data streams. The central controller first fuses this data in real time: the forces / torques measured by the four sensors are converted and synthesized into a total resultant force F_global(t) and a total resultant torque M_global(t) acting on the center of the adsorption surface, based on the coordinates of their installation positions relative to the center of the adsorption surface. This is a six-dimensional vector that varies with time in a global coordinate system. Next, the system needs to decouple the quasi-static loads generated by the vehicle's basic motion (uniform linear motion, uniform acceleration) and the dynamic disturbance loads generated by vibration and impact from these total loads. This can be achieved through adaptive filtering techniques, for example, using the desired acceleration from the vehicle controller as a reference signal to filter out the low-frequency components associated with it in the load, thereby separating the high-frequency disturbance load components. The decoupled load information provides a clearer analytical basis for force field reconstruction.

[0100] S4.2: Real-time reconstruction of microscopic force fields based on inverse problem solving.

[0101] This is a crucial step in transforming global measurements into local understanding. The system has the following known information: the global load (F_global, M_global) and the spatial locations of all valid anchor points. The local force f_i at each anchor point needs to be solved. This is a classic inverse problem in mechanics. The system constructs a system of linear equations: global load = vector sum of all local forces. Since the number of anchor points is usually greater than the number of force equilibrium equations, the system of equations is underdetermined. This embodiment uses the Tikhonov regularization method to solve it. Specifically, the objective function is to minimize the global load residual while also minimizing the weighted sum of squares of all local force magnitudes (i.e., seeking the smoothest force distribution solution). The diagonal elements of the weighting matrix W are set according to the confidence score of each anchor point: anchor points with high confidence are allowed to bear greater forces (smaller weights); anchor points with low confidence are limited in the forces they bear (larger weights). By solving this optimization problem, the system outputs in real-time the estimated three-dimensional force components of each anchor point at the current time, i.e., the reconstructed microscopic force field distribution map F_micro(t).

[0102] S4.3: Rule-based and model-based adaptive and feedforward composite regulation.

[0103] Based on the reconstructed microscopic force field F_micro(t), the system performs dual-mode regulation. Adaptive (feedback) regulation: The system continuously monitors the force state of each anchor point. If the shear force or normal tensile force at a point exceeds its dynamic safety threshold (which can be adaptively adjusted according to its diagnostic label), that point is marked as an overload risk point. The controller then sends a small compensating current pulse to the corresponding TEC at that point, causing a slight drop in its local temperature to instantaneously enhance the ice bridge strength. Conversely, for points with forces far below average, their sustaining power can be appropriately reduced. Predictive (feedforward) regulation: The system continuously receives future motion commands from the vehicle. Using a simplified mass-spring-damper model, it predicts the spatial distribution pattern of the inertial load at the adsorption interface during an upcoming action (such as the initiation of braking) – the "additional force field increment ΔF_predict". The controller pre-enhances the anchor point group predicted to bear additional tensile force in ΔF_predict, for example, by applying a small reinforcing pulse hundreds of milliseconds in advance. This strategy, which combines feedback to eliminate current errors with feedforward to cancel expected disturbances, constitutes a high-performance composite controller, greatly improving the system's robustness against dynamic disturbances.

[0104] According to another aspect of this application, real-time reconstruction of the microscopic force field based on inverse problem solving includes:

[0105] S4.2.1: Constructing the force balance equations and regularization model.

[0106] The system first constructs a geometric matrix G based on the spatial positions (coordinates (x_i, y_i, z_i)) of all valid anchor points. This matrix relates the local force vector f (with three force components f_xi, f_yi, f_zi for each anchor point) to the global resultant force and resultant moment, satisfying the equation: G × f = F_global, where F_global is a 6 × 1 global load vector (three force components and three moment components). Since the number of anchor points n is usually greater than 6, this equation is underdetermined and has infinitely many solutions. According to this embodiment, Tikhonov regularization is used to obtain a stable solution, i.e., solving a minimization problem: min ||G × f - F_global|| 2 + λ× ||W×f|| 2 , where λ is the regularization parameter, W is the diagonal weight matrix, and its diagonal elements w_i are set according to the confidence score of each anchor point. The lower the confidence score, the larger w_i is, so as to limit the force estimate of that point.

[0107] S4.2.2: Real-time solution of regularized least squares problems.

[0108] The above minimization problem has an analytical solution: f = (G T× G + λ× W^T × W) -1 × G T × F_global. To improve real-time performance, the system pre-computes the matrix H = (G) whenever the set of effective anchor points changes (i.e., at the start of each task). T ×G + λ×W T ×W) -1 × G T During operation, the local force estimation vector f can be obtained by multiplying the real-time measured and fused F_global with the pre-calculated H. This matrix multiplication has a very small computational cost, easily meeting real-time requirements (such as a 100Hz update rate). When an anchor point fails (e.g., the diagnostic label is marked as failed), the system removes it from the valid list and recalculates matrix H.

[0109] S4.2.3: Post-processing and visualization of force field solutions.

[0110] The obtained local force vector f contains the three-dimensional force components of each anchor point. The system first calculates the resultant force magnitude |f_i| and force direction angle at each anchor point. Then, based on the positions and force vectors of all anchor points, it generates a two-dimensional force field cloud map covering the entire adsorption region using an interpolation algorithm (such as natural neighborhood interpolation), including the distribution of normal force and shear force. Simultaneously, the system calculates statistical indices for the entire force field: average normal force, normal force non-uniformity coefficient, maximum shear force, etc. These cloud maps and statistical indices are displayed in real-time on the monitoring interface and serve as input for subsequent adjustments.

[0111] According to another aspect of this application, rule-based and model-based adaptive and feedforward composite regulation includes:

[0112] S4.3.1: Real-time identification and priority ranking of risk anchor points.

[0113] The system analyzes the reconstructed microscopic force field data in real time. For each anchor point, a real-time risk score is calculated based on its stress conditions (normal tensile force, shear force) and its strength confidence level. For example, the risk score = (current shear force / predicted maximum shear strength) × weight 1 + (current normal tensile force / predicted maximum normal strength) × weight 2. Then, all anchor points are sorted in descending order of risk score to obtain a risk priority list. Simultaneously, the system also identifies points in the force field where the stress is significantly below the average level (idle points), which may not be fully utilized.

[0114] S4.3.2: Execution of adaptive feedback adjustment strategy.

[0115] For points in the risk priority list, the system employs a tiered adjustment strategy. Specifically, if the risk score of an anchor point exceeds threshold 1 (high risk), a strong reinforcement pulse (e.g., a 20% increase in current, lasting 100ms) is immediately sent to its TEC (Dynamic Energy Center), and the force direction of its surrounding anchor points may be finely adjusted (by controlling minor temperature changes in the surrounding anchor points, altering local stiffness, and thus influencing the force transmission path). If the risk score exceeds threshold 2 (medium risk) but is below threshold 1, a medium-intensity reinforcement pulse (a 10% increase in current, lasting 50ms) is sent. For idle points, the system can appropriately reduce their sustaining power, allowing a slight temperature increase, thus slightly increasing the force they bear, thereby optimizing the overall force field distribution. All adjustment commands are implemented through a closed-loop PID controller to ensure precise adjustment.

[0116] S4.3.3: Active feedforward adjustment based on prediction model.

[0117] Meanwhile, the feedforward adjustment channel is also operational. The vehicle control system provides the expected acceleration changes (e.g., a_x, a_y) for the next few hundred milliseconds. The system uses a simplified rigid body dynamics model to predict how these accelerations will generate additional force distributions at the adsorption interface. Specifically, treating the cargo as a rigid body, the system calculates the distribution of inertial forces generated by accelerations a_x and a_y at each anchorage point (based on the point's position relative to the center of mass). The system then preemptively reinforces those anchorage points predicted to experience additional tension, for example, by sending a slight pre-reinforcement pulse (a 5% increase in current, lasting until the expected acceleration event ends) to these points 50 ms in advance. In this way, the system pre-compensates before the disturbance actually occurs, significantly reducing the impact of dynamic disturbances.

[0118] S5: After the task is completed, perform a controlled release operation and conduct system performance evaluation and parameter self-evolution.

[0119] S5.1: Verification of the execution and release process of the multi-stage controllable release strategy.

[0120] The release operation is designed as a multi-stage controlled process. The first stage is the pre-release stage: the controller applies a slow, linear heating ramp current to all active anchor points, causing the anchor tip temperature to rise uniformly from the operating temperature (e.g., -35°C) to a preset first release temperature point (e.g., -5°C). During this process, force sensors continuously monitor the decay of the adsorption force to ensure a smooth descent. The second stage is the main release stage: when the total adsorption force drops below a safe threshold, the controller applies a rapid heating pulse, quickly pushing the temperature above freezing to a positive temperature release point (e.g., +3°C), ensuring all ice bridges completely melt. The third stage is the separation and reset stage: the controller retracts all anchors and drives the adsorption fork to detach from the cargo surface. Simultaneously, sensors confirm that the forces in all directions have returned to zero, verifying complete release. The entire process's time-temperature curve is fully recorded, forming a release process log.

[0121] S5.2: Full task lifecycle data archiving and multi-dimensional performance evaluation.

[0122] After the task is completed, the system automatically initiates the data archiving and analysis process. The complete data chain from S1 to S5.1, including raw sensor data, feature vectors, model parameters, planning schemes, process curves, control commands, force field data, release logs, etc., is packaged and compressed into a structured task file and stored in the historical database. Subsequently, the performance evaluation module automatically calculates the performance score of this task based on preset indicators. These indicators include: task success rate (whether slippage or damage occurred), total energy consumption, average anchoring strength, force field uniformity index, release time, and the degree of suppression of dynamic disturbances. These indicators are weighted and synthesized to obtain an overall performance score, which is then stored in association with the task file.

[0123] S5.3: Parameter self-evolution and strategy optimization based on group experience.

[0124] The system's self-evolution occurs in a separate, low-priority background learning thread. This thread periodically extracts a batch of recent task files and their performance scores, forming a learning dataset. Further improvements to this embodiment can employ Bayesian optimization or deep reinforcement learning as the evolutionary engine. Taking Bayesian optimization as an example: the engine defines a set of adjustable parameters in the system control (such as the architecture hyperparameters of the feature network in S1, the weights of the optimization algorithm in S2, the threshold of the diagnostic classifier in S3, and the gain of the regulator in S4) as the search space. Each historical task is considered an experiment, with the input being the initial conditions of the task (summarized by the feature vector) and the output being the performance score. The Bayesian optimization engine constructs a probabilistic surrogate model (Gaussian process) on these historical data points regarding "parameter combination → expected performance." Then, based on the acquisition function (such as the desired improvement in EI), it recommends a new parameter combination within the search space that may bring performance improvement. This new combination can be verified in subsequent simulation tests, or, under strict safety monitoring, deployed to a test vehicle for actual verification. The validation results (new task profiles and scores) are fed back into the learning dataset to update the agent model. Through continuous iteration, the system can gradually discover and converge to a better set of global control parameters, thereby achieving continuous autonomous evolution of performance and adapting to new working conditions and challenges without human intervention.

[0125] According to another aspect of this application, parameter self-evolution and policy optimization based on group experience include:

[0126] S5.3.1: Preprocessing and feature engineering of historical task data.

[0127] The self-evolutionary engine periodically (e.g., daily) extracts all structured task files from a recent period (e.g., one month) from a historical database. First, each task file is preprocessed to extract features and objectives for evolutionary learning. Specifically, input features include: a multi-dimensional feature vector at the start of the task and key statistical features of the dynamic disturbance load spectrum during the task (e.g., maximum acceleration, major vibration frequencies). The objective variable is the overall performance score for the task. Additionally, the system records the set of control parameters used for the task (i.e., the values ​​of all currently adjustable parameters). All task data is organized into a table, with each row representing one task and columns including input features, control parameter set, and performance score.

[0128] S5.3.2: Constructing a proxy model and performing parameter sensitivity analysis.

[0129] The engine employs Gaussian process regression as a surrogate model to fit the complex mapping relationship from "input features + control parameter set" to "performance score". Specifically, the input features and control parameters are first standardized. Then, a composite kernel function (such as the product of a linear kernel and a radial basis kernel) is used to capture linear and nonlinear relationships. The Gaussian process model is trained using historical data. After training, the model can predict the expected performance score for any given set of input features and control parameters. Furthermore, the engine performs global sensitivity analysis, such as using the Sobol exponent, to assess the average impact of each control parameter on the performance score, thereby identifying the key parameters with the greatest impact on performance for focused adjustment in subsequent optimizations.

[0130] S5.3.3: Bayesian optimization of search and parameter update.

[0131] Based on a trained Gaussian process surrogate model, the engine uses Bayesian optimization to find a better set of control parameters. The optimization objective is to maximize the expected performance score for various possible future operating conditions (characterized by the distribution of input features). Specifically, Expected Improvement (EI) is used as the acquisition function. The optimization process iterates within the search space of the control parameters: the surrogate model predicts the performance score and its uncertainty based on existing data; the EI function balances exploration (trying regions with high uncertainty) and utilization (trying regions with high predicted scores) to propose the next parameter combination to be evaluated. Since performing each evaluation on the actual system is costly, the engine prioritizes batch evaluation in a simulation environment composed of digital twin models. That is, a batch of representative operating conditions is randomly sampled from historical input features, and the performance of the new parameter combination is tested in simulation to obtain an average simulation performance score. When a new parameter combination performs significantly better than the current parameters in simulation (e.g., the average simulation score is improved by more than 5%), and passes safety checks (e.g., ensuring that all safety factors meet the standards), the parameter combination will be marked as a candidate update parameter.

[0132] S5.3.4: Online verification and deployment of candidate parameters.

[0133] Candidate updated parameters are not immediately deployed to all vehicles. According to a further improvement in this embodiment, the system employs a gradual deployment strategy. First, small-scale online validation is conducted on one or a few test shuttles. During validation, the system meticulously records task data for these vehicles using the new parameters and compares it with similar tasks using the old parameters. If the online validation results show that the new parameters do indeed statistically significantly improve performance scores (e.g., t-test p-value < 0.05) and do not introduce new failure modes, then the new parameters will be approved for use on a larger fleet. Finally, the new parameters are safely and batch-wise deployed to all shuttles in the entire fleet via over-the-air download technology. After each deployment, the system continues to collect data, forming a complete closed loop of "evaluation-optimization-validation-deployment," thereby achieving continuous autonomous evolution of system performance.

[0134] Example 3

[0135] In this embodiment, a typical operation of the automatic adsorption of the forklift ruler by the cold storage shuttle car is used to demonstrate the data processing flow of this solution.

[0136] The environmental conditions are as follows: the cold storage temperature is -25°C, and the relative humidity is 85%. The goods to be transported by the shuttle are food boxes packaged in corrugated cardboard, with external dimensions of 600mm (length) × 400mm (width) × 300mm (height) and a total weight of 20kg. The forklift adsorption panel has dimensions of 620mm × 420mm, with 100 independently controllable phase change anchoring units evenly distributed on the surface, with a unit spacing of 40mm.

[0137] The specific implementation steps and data calculations are as follows:

[0138] S1: Construction of multimodal sensing and digital twin models, including:

[0139] 1.1 Multimodal data acquisition, including:

[0140] 1) Visual data acquisition:

[0141] A Basler acA2440-75um polarization camera was used, with polarization angles set to 0° and 90° respectively; acquisition parameters: resolution 1024×768, exposure time 10ms, gain 1.2; two raw polarization images I1 and I2 were acquired, with pixel values ​​ranging from [0, 255].

[0142] 2) Temperature data acquisition:

[0143] An array of 25 PT1000 thin-film temperature sensors, sampling at a frequency of 100Hz, yields a temperature data matrix T_raw(5×5).

[0144] [-24.8, -25.2, -25.1, -24.9, -25.0]

[0145] [-25.1, -25.3, -25.4, -25.2, -25.1]

[0146] [-24.9, -25.0, -25.5, -25.3, -25.2]

[0147] [-24.7, -24.8, -25.2, -25.0, -24.9]

[0148] [-24.8, -24.9, -25.1, -24.8, -24.7].

[0149] 3) Mechanical data acquisition:

[0150] Four ATI Mini40 six-dimensional force sensors, sampling frequency 1kHz; data collected during the pre-compression phase (applying a 5N normal force):

[0151] Sensor 1: Fx=0.12N, Fy=0.08N, Fz=5.05N, Mx=10.5N·mm, My=9.2N·mm, Mz=0.32N·mm; Sensor 2: Fx=0.09N, Fy=-0.04N, Fz=4.98N, Mx=9.8N·mm, My=-8.3N·mm, Mz=0.28N·mm; Sensor 3: Fx=-0.11N, Fy=0.07N, Fz=5.02N, Mx=-10.8N·mm, My=9.0N·mm, Mz=-0.35N·mm; Sensor 4: Fx=0.06N, Fy=-0.05N, Fz=4.96N, Mx=-9.9N·mm, My=-8.5N·mm. Mz = 0.31 N·mm.

[0152] 4) Dielectric data acquisition:

[0153] AD7746 capacitor digital converter, excitation frequency 16kHz; acquired data: in-phase component I=1.28V, quadrature component Q=0.35V; drying reference value: I0=1.00V, Q0=0.20V.

[0154] 1.2 Feature extraction calculation, including:

[0155] 1) Visual feature calculation:

[0156] Polarization difference: I_diff = |I1- I2|, enhances ice crystal texture;

[0157] Adaptive threshold segmentation: Threshold T = 0.7 × mean(I_diff) = 0.7 × 85.6 = 59.92;

[0158] Ice crystal coverage calculation: Binary image B(x,y) = 1 if I_diff(x,y) > T else 0; Coverage C_ice = ΣB(x,y) / (1024×768) = 0.342 (34.2%);

[0159] Frost layer thickness estimation: The depth map D(x,y) is calculated by parallax, and the residual of the reference plane fitting is the thickness; the average thickness t_avg = 0.32mm, and the standard deviation σ_t = 0.12mm.

[0160] 2) Temperature characteristic calculation:

[0161] Bilinear interpolation generates a 100×100 temperature field: using the interpolation formula: T(x,y) = a0 + a1x + a2y +a3xy; resulting in a continuous temperature distribution T_map(100×100);

[0162] Gradient calculation: G_x(i,j) = [T(i+1,j) - T(i-1,j)] / (2Δx); G_y(i,j) = [T(i,j+1)- T(i,j-1)] / (2Δy); Average gradient magnitude |G|_avg = 0.48°C / mm;

[0163] Mechanical characteristic calculation: Global resultant moment: F_total = ΣF_i = [0.16, 0.06, 20.01] N; M_total = Σ(M_i + r_i × F_i) = [1.2, -0.8, 0.56] N·mm;

[0164] Spectrum analysis: Perform a 256-point FFT on the Fz signal with a Hanning window; Dominant frequencies: f1 = 9.8 Hz (amplitude 0.12 N), f2 = 15.2 Hz (amplitude 0.08 N); Low-frequency energy ratio: E_0-50 Hz / E_total = 0.76.

[0165] 3) Dielectric characteristic calculation:

[0166] Amplitude and phase: ;φ=arctan(Q / I) = arctan(0.35 / 1.28) = 0.267rad;

[0167] Amount of change: ΔA = A / A0- 1 = 1.327 / 1.020 - 1 = 0.301; Δφ=φ-φ0 = 0.267 -0.197 = 0.070rad;

[0168] Liquid water content inversion: From the table (temperature -25°C): Φ_w = 0.048±0.005.

[0169] 3) Feature vector construction:

[0170] After standardization, a 100-dimensional feature vector is obtained: X = [0.342, 0.320, 0.120, 0.480, 0.160, 0.060, 20.010, 1.200, -0.800, 0.560, 9.800, 0.120, 15.200, 0.080, 0.760, 0.301, 0.070, 0.048, ...].

[0171] 1.3 Instantiation of the digital twin model, including:

[0172] 1) Neural network parameter identification:

[0173] Network structure: 100 nodes in the input layer, 64-32-16 nodes in the hidden layer, and 5 nodes in the output layer;

[0174] Activation function: ReLU;

[0175] Input feature vector X;

[0176] Output parameters: E_eff = 2.47 GPa (equivalent elastic modulus); k_bond = 0.79 (interfacial strength coefficient); μ = 0.16 (friction coefficient); λ = 0.52 W / (m·K) (thermal conductivity); c = 1820 J / (kg·K) (specific heat capacity).

[0177] 2) Finite element model update:

[0178] Number of mesh elements: 60 × 40 × 3 = 7200 hexahedral elements;

[0179] The material properties have been updated to the identification values ​​mentioned above;

[0180] Boundary conditions: bottom fixed, top free;

[0181] A personalized digital twin model is obtained.

[0182] S2: Anchor point dynamic programming, including:

[0183] 2.1 Dynamic load spectrum generation, including:

[0184] 1) Vehicle motion planning:

[0185] Time series: t = [0, 1, 2, 3] s;

[0186] Acceleration: a = [0.3, 0.3, 0.0, -0.2] m / s 2 ;

[0187] Velocity: v = [0.0, 0.3, 0.6, 0.4] m / s.

[0188] 2) Load calculation:

[0189] Inertial force: F_inertial(t) = m × a(t) = 20 × a(t) N; F_inertial = [6.0, 6.0, 0.0, -4.0] N;

[0190] Vibration load: Dominant frequency: f_vib = 15Hz, Amplitude A_vib = 0.1g = 0.98m / s 2 ;F_vib(t) = m× A_vib × sin(2πf_vib t) = 19.6 × sin(94.2t) N;

[0191] Load spectrum matrix L(4×3):

[0192] Time (s) Fx(N) Fy(N) Fz(N) 0.0 0.0 0.0 196.0 1.0 6.0 0.0 196.0 2.0 0.0 0.0 196.0

[0196] 3.0 -4.0 0.0 196.0

[0197] (Note: Fz includes gravity 196N).

[0198] 2.2 Optimization solution, including:

[0199] 1) Genetic Algorithm Parameters: Population size: 100; Encoding method: binary, 100 bits (corresponding to 100 candidate points); Crossover probability: 0.8; Mutation probability: 0.05; Maximum number of generations: 50;

[0200] 2) Example of fitness calculation (an individual is coded with 12 anchor points):

[0201] Digital twin simulation: Load spectrum L; Calculation of stress σ_ij(t) for each element;

[0202] Extracted indicators: Minimum safety factor: S_min = min(σ_yield / σ_max) = 2.48; Number of anchor points: N = 12; Stress variation coefficient: CV = std(σ_avg) / mean(σ_avg) = 0.18;

[0203] Fitness value: Fitness = 0.6×S_min / 3 + 0.2×(1-N / 50) + 0.2×(1-CV)= 0.6×2.48 / 3 + 0.2×(1-12 / 50) + 0.2×(1-0.18)= 0.496 + 0.152 + 0.164 = 0.812;

[0204] After 50 generations of evolution, the optimal solution is:

[0205] Number of anchor points: 12; Coordinates (unit: mm):

[0206] (50,50), (50,150), (50,250), (50,350)

[0207] (250,100), (250,200), (250,300)

[0208] (450,50), (450,150), (450,250), (450,350)

[0209] (550,200)

[0210] Minimum safety factor: 2.65;

[0211] Fitness: 0.843.

[0212] 2.3 Post-processing verification, including:

[0213] Spacing check: Minimum spacing 41.2mm > 40mm (pass);

[0214] Avoid seams: all points are more than 30mm from the edge (pass);

[0215] Power check: Simultaneous power consumption at 12 points <200W (pass).

[0216] S3: Phase change anchoring and process monitoring, including:

[0217] 3.1 Anchored execution, including:

[0218] Taking the anchor point (50,50) as an example:

[0219] 1) Positioning: Linear motor movement, error <0.1mm;

[0220] 2) Preload: Apply a normal force of 5.02N (sensor feedback);

[0221] 3) Current pulse sequence:

[0222] Phase 1 (0-100ms): 10A, T drops from -25°C to -51°C;

[0223] Phase 2 (100-600ms): 5A, T maintained at -36°C;

[0224] Phase 3 (after 600ms): 1.2A, T maintained at -35°C.

[0225] 3.2 Process monitoring data, including:

[0226] 1) T(t): The temperature drops from -25°C to -50°C within 100ms, and then stabilizes at -35°C;

[0227] 2) R(t): The R value drops from 10 MΩ to 5 kΩ within 200 ms and then stabilizes.

[0228] 3.3 Feature extraction, including:

[0229] Calculate dynamic characteristics:

[0230] Cooling slope: S_max = (-51.2 - (-25.0)) / 0.1 = -262 °C / s;

[0231] Cooling time: Δt_cool = 100ms;

[0232] Overshoot amplitude: Overshoot = 0°C;

[0233] Resistance fall time: Δt_settle = 80ms;

[0234] Final resistance: R_final = 5.1kΩ;

[0235] Resistance decrease rate: K_R = (log10(5.1)-log10(10000)) / 0.08 = -26.2 / s;

[0236] Eigenvector: [-262, 0.100, 0, 0.080, 5.1, -26.2].

[0237] 3.4 Diagnostic classification, including:

[0238] 1) Support Vector Machine Classifier (training accuracy 98.5%): Input feature vector; Output: Class = 1 (optimal anchor), confidence = 0.96;

[0239] 2) Diagnostic results of 12 anchor points: 11 optimal anchors (confidence level > 0.9); 1 acceptable anchor (confidence level = 0.85); no failure points.

[0240] S4: Force field reconstruction and adjustment, including:

[0241] 4.1 Force data acquisition, including:

[0242] 1) Force sensor data during vehicle movement (example time):

[0243] Sensor 1: [0.15, 0.12, 52.3, 25.6, 22.1, 0.45];

[0244] Sensor 2: [0.11, -0.08, 48.9, 24.3, -20.8, 0.38];

[0245] Sensor 3: [-0.13, 0.10, 51.8, -26.7, 23.5, -0.42];

[0246] Sensor 4: [0.08, -0.09, 47.5, -25.2, -21.3, 0.41].

[0247] 2) Global force calculation:

[0248] F_global = [0.21, 0.05, 200.5] N;

[0249] M_global = [3.2, -2.5, 0.82] N·mm.

[0250] 4.2 Force field reconstruction, including:

[0251] 1) Construction of the geometric matrix G (6×36): The position of each anchor point is r_i = (x_i, y_i, 0); the elements of the G matrix are determined by r_i;

[0252] 2) Weight matrix W (36×36 diagonal matrix): w_i = 1 / confidence level; for example: w1 = 1 / 0.96 = 1.042;

[0253] 3) Regularization solution:

[0254] Minimize: ||G·f - F_global|| 2 + λ·||W·f||2 λ = 0.1;

[0255] Solution: f = (G T G + λW T W) -1 G T F_global yields the force distribution at each point (example):

[0256] Point 1: f1 = [0.02, 0.01, 16.35] N;

[0257] Point 2: f2 = [0.01, 0.03, 16.42] N; ...

[0259] Point 12: f 12 = [0.03, -0.02, 16.28] N;

[0260] 4) Force field statistics: mean normal force: 16.38 N; standard deviation: 0.18 N; coefficient of variation: 0.011.

[0261] 4.3 Adaptive adjustment, including:

[0262] 1) Risk Identification

[0263] Safety factor calculation: S_i = f_max_allowed / |f_i|;

[0264] All points S_i > 15, no high-risk points;

[0265] 2) Homogenization adjustment:

[0266] The normal force at point 5 is 14.2 N, which is less than the average of 16.38 N.

[0267] Adjustment command: Reduce the holding current by 10% (1.2A → 1.08A);

[0268] The temperature rose from -35°C to -33°C;

[0269] After redistribution, point 5 experiences a force of 15.8.

[0270] 3) Feedforward adjustment:

[0271] Received deceleration command: Decelerate by 0.2 m / s after 0.5 s. 2 ;

[0272] The shear force is predicted to increase at the points (1,2,3,4) ahead.

[0273] Pre-enhancement: Current increased by 5% for 0.5s.

[0274] S5: Controlled release and self-evolution, including:

[0275] 5.1 Controlled Release

[0276] Release sequence:

[0277] Phase 1 (0-60s): Heating rate: 0.5°C / s; Temperature change: -35°C → -5°C; Adsorption force change: 200.5N → 42.3N;

[0278] Phase 2 (60-65s): Rapid heating: -5°C → +3°C; Adsorption force change: 42.3N → 0.5N;

[0279] Phase 3 (65-67s): Retract the anchor pin and completely separate.

[0280] 5.2 Performance Evaluation

[0281] Evaluation metrics: Mission success: 1.0 (no slippage); Total energy consumption: 1180J; Average anchoring strength: 16.38N (safety factor 18.2); Force field uniformity: coefficient of variation 0.011; Release time: 65s; Dynamic tracking error: 0.12; Weighted score: Score = 0.3×1.0 + 0.15×(1-1180 / 2000) + 0.2×(18.2 / 20) + 0.15×(1-0.011) +0.1×(1-65 / 100) + 0.1×(1-0.12 / 0.5)= 0.300 + 0.123 + 0.182 + 0.148 + 0.035 +0.076= 0.864.

[0282] 5.3 Parameter Self-Evolution

[0283] 1) Training dataset (last 100 tasks): Input feature dimension: 120 (operating condition features + control parameters); Output: performance score [0,1];

[0284] 2) Gaussian process model: Kernel function: RBF + linear kernel; Hyperparameter optimization: Maximum likelihood estimation;

[0285] 3) Bayesian optimization:

[0286] The current optimal parameter is θ_current, with an average score of 0.82.

[0287] Acquisition function: Desired improvement (EI);

[0288] Recommended new parameter: θ_new;

[0289] Simulation evaluation: 50 random operating conditions, average score 0.84;

[0290] Online verification: 10 actual tasks, average score 0.847;

[0291] 4) Parameter update:

[0292] Genetic algorithm safety weight: 0.60 → 0.62;

[0293] PID proportional gain: 1.2 → 1.25;

[0294] Temperature maintenance setting: -35°C → -34.5°C.

[0295] This embodiment demonstrates a complete intelligent adsorption control process. Calculations based on actual data yielded the following results: Positioning accuracy: Anchor point positioning error <0.1mm; Anchoring success rate: 11 out of 12 points were optimally anchored, a success rate of 91.7%; Force field uniformity: Coefficient of variation was only 0.011, indicating uniform distribution; Safety factor: Minimum safety factor of 2.65, meeting engineering requirements; System efficiency: Comprehensive score of 0.864, indicating excellent performance; Learning and evolution: Performance improved by 3.3% after parameter optimization. This embodiment verifies the feasibility and effectiveness of the method, providing detailed technical reference for the development of practical systems.

[0296] Unlike traditional passive adsorption methods that rely on friction or magnetism, this invention actively generates mechanical anchoring points at the interface through phase change anchoring technology, fundamentally solving the problem of insufficient adhesion in low-temperature, frosty environments. This microscopic mechanical anchoring based on instantaneous phase change can reliably penetrate the surface frost layer, achieving direct mechanical interlocking with the metal substrate. The adhesion is more than three times stronger than traditional magnetic attraction methods, and it is unaffected by low-temperature environments.

[0297] Through a multi-dimensional force sensor array and dynamic force field reconstruction technology, the system can sense the force state of each anchoring point in real time and proactively identify instability risks based on model predictive control. When it is predicted that the load in a specific area will exceed the threshold, the system will activate the backup anchoring unit in advance to divert the load. This "predictive enhancement" mechanism enables the adsorption system to have active anti-interference capabilities, effectively preventing the shuttle from slipping or falling due to sudden stops or turns of the forklift.

[0298] The system constructs a high-fidelity digital twin model of contact mechanics using multimodal sensing data, including multispectral scanning, thermal imaging, and pressure distribution detection. This model can accurately predict the stress distribution at the interface, providing a scientific basis for anchor point planning. Compared to traditional methods that rely on empirical rules, the decision-making process of this invention is based on precise physical simulation and data-driven approaches, improving the efficiency of anchor point layout optimization by more than 60%.

[0299] The system establishes anchor unit health records and performance degradation models by recording end-to-end data for each operation, enabling predictive maintenance. Simultaneously, by comparing measured data with model predictions, the system automatically corrects model parameters, allowing the digital twin model to continuously evolve and become increasingly accurate with use. This self-learning capability significantly extends equipment lifespan, is expected to reduce maintenance costs by more than 50%, and simultaneously improves system performance over time.

[0300] In summary, this invention upgrades traditional passive anchoring devices into intelligent systems with sensing, decision-making, execution, and learning capabilities. This not only solves the special technical challenges in cold storage environments but also pioneers a new paradigm for intelligent anchoring technology, providing a complete technical solution for reliable equipment connection in extreme environments.

[0301] 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. A smart control method for automatically adsorbing forklift rulers on a cold storage shuttle, characterized in that, include: By collecting interface state information through multimodal sensing data, a digital twin model of contact mechanics is constructed based on the interface state information; Anchor point dynamic planning is performed based on a digital twin model to generate an optimal anchor point coordinate set resistant to dynamic disturbances. This includes: quantifying and generating a dynamic load input spectrum based on vehicle motion planning and environmental vibration data; using the digital twin model as a simulation engine and employing an evolutionary algorithm to optimize the anchor point layout with interface safety and force field uniformity as optimization objectives; verifying the engineering feasibility of the optimized solution and performing post-processing to output the final anchor point coordinate instruction set. The phase change anchoring operation is performed based on the anchoring point coordinate set, and the anchoring process status is monitored in real time, including: driving the actuator to accurately position the anchor pin and apply micro-contact pre-pressure; applying multi-stage phase change anchoring energy pulses to the target anchoring point; synchronously acquiring the temperature and resistance transient curves of the anchoring process; extracting dynamic feature parameters from the transient curves to form a process feature vector; and performing real-time diagnosis and classification of the anchoring quality based on the process feature vector to generate a panoramic status report of the anchoring process. Multidimensional force vector data is collected and dynamic force field reconstruction is performed. Based on the reconstruction results, adaptive and predictive adjustment of the adsorption force field is carried out, including: synchronously collecting raw signals from multiple force sensors, fusing and decoupling them to obtain the global resultant external force and resultant torque; constructing a weighted regularized model based on anchor point location and confidence score, inverting and solving the local force vector of each anchor point, and reconstructing the microscopic force field distribution map; identifying risky anchor points and idle points in real time, and performing graded intensity adjustment and force field homogenization adjustment based on feedback rules; predicting future load distribution based on vehicle motion commands, and performing feedforward pre-strengthening adjustment on key anchor points. After the task is completed, a controlled release operation is performed, and system performance is evaluated and parameters are self-evolved.

2. The method according to claim 1, characterized in that, By acquiring interface state information through multimodal sensing data, a digital twin model of contact mechanics is constructed based on the interface state information, including: Acquire multimodal raw data including visual images, surface temperature, initial contact force, and dielectric properties; Parallel processing of multimodal raw data is performed to extract multi-domain features such as micromorphology, temperature gradient, mechanical spectrum and liquid phase content, forming a multi-dimensional feature vector. Multi-dimensional feature vectors are input into a parameter identification network to dynamically output key time-varying material parameters of the digital twin model; By instantiating the basic finite element model using key time-varying material parameters, a high-fidelity digital twin model of contact mechanics is obtained.

3. The method according to claim 2, characterized in that, Parallel processing of multimodal raw data to extract multi-domain features, including: Polarization difference and 3D reconstruction are performed on visual images to extract ice crystal distribution parameters and frost thickness maps; Interpolation and gradient calculations are performed on surface temperature data to generate a temperature field distribution map and a two-dimensional temperature gradient vector field; Coordinate fusion and spectrum analysis were performed on the initial contact force data to obtain the global resultant external torque and the spectrum characteristics of the dominant vibration. Dielectric property data are solved, and the estimated value of liquid water content at the interface is obtained by inverting the mixed medium model; All extracted features are standardized and concatenated to form a multi-dimensional feature vector.

4. The method according to claim 1, characterized in that, An evolutionary algorithm is used to optimize the anchor point layout, including: An initial population representing different anchor point layouts is randomly generated within the feasible region of the adsorption surface. For each individual in the initial population, a dynamic load input spectrum is applied to the digital twin model to perform transient dynamic simulation; The peak interfacial stress, safety factor, and distribution uniformity indices are extracted from the simulation results to calculate individual fitness. Based on fitness, selection, crossover, and mutation operations are performed to iteratively update the population until the convergence condition is met, and the optimal layout scheme is output.

5. The method according to claim 1, characterized in that, Real-time diagnosis and classification of anchoring quality based on process feature vectors, including: The initial descent time and maximum cooling slope of the temperature curve were detected, as well as the steep descent start time and stable value of the resistance curve. The extracted dynamic feature parameters are matched and compared with the pre-stored feature library of the gold anchoring process. Based on the comparison results, output the anchoring quality diagnostic label and the relative anchoring strength confidence score.

6. The method according to claim 1, characterized in that, Reconstructing the microscopic force field distribution map, including: Construct a geometric matrix based on the locations of all valid anchor points, and establish the mechanical equilibrium equations for local forces and global loads; A weight matrix is ​​constructed based on the anchor point confidence score, and a Tikhonov regularization term is introduced to transform the problem into a least squares optimization problem. The least squares optimization problem is solved in real time to obtain the three-dimensional force estimate of each anchor point, and a two-dimensional force field cloud map is generated by interpolation.

7. The method according to claim 1, characterized in that, After the task is completed, a controlled release operation is performed, and system performance is evaluated and parameters are self-evolved, including: A multi-stage controlled temperature-controlled release sequence was executed, and the adsorption force decay was monitored to verify complete separation; Archive the data for the entire task lifecycle and calculate the overall performance score for this task based on a multi-dimensional indicator system; Based on a set of historical task archives, a Bayesian optimization framework is used to search for a better set of system control parameters. Candidate update parameter groups are validated online, and after passing the validation, they are progressively deployed to the fleet to achieve continuous evolution of system performance; The search using the Bayesian optimization framework includes: Extract task features, control parameter sets used, and corresponding performance scores from historical task archives to form a training dataset; A surrogate model from control parameters to performance scores is built on the training dataset using Gaussian process regression. Using the desired improvement as the acquisition function, the system iteratively recommends the parameter combination to be evaluated within the search space of the control parameters. The performance of the recommended parameter combinations is evaluated in batches within a digital twin simulation environment to select candidate parameter update groups.

Images