Large transformer multi-transport unit collaborative bearing device and synchronous control method thereof

Abstract

The application provides a large transformer multi-transport unit cooperative bearing device and a synchronous control method thereof, and belongs to the technical field of transformers. Static bearing components and transient impact components are separated from sensor signals through zero-point calibration and Gaussian mixture model dynamic window filtering. A bearing geometric model is established according to displacement sensors and pressure sensors, and preliminary equivalent support reaction forces of each tank are calculated. The weighted least squares method is used to fuse strain vectors of multiple measuring points and static balance estimation results to obtain corrected equivalent support reaction forces. A comprehensive bearing risk index is obtained by weighted summation of multiple bearing state indexes. Artificial intelligence residual error checking and secondary correction are performed on the gateneuron dynamic network model through adaptive feedback iteration. The synchronous compensation control amount of each tank is dynamically adjusted according to the corrected risk index, and the technical problem of insufficient identification accuracy of equivalent support reaction forces of the multi-transport unit under non-uniform transient start-stop working conditions is solved.

Description

Technical Field

[0001] This invention belongs to the field of transformer technology, and more specifically, relates to a large transformer multi-transportation unit collaborative bearing device and its synchronous control method. Background Technology

[0002] In the field of large-scale power equipment handling, the in-plant transfer of heavy equipment such as transformers typically employs a multi-tank collaborative load-bearing method. The equipment load is distributed to the support points of various tanks via a main load-bearing beam, and the control system synchronously compensates for the speed, acceleration, and steering of each tank. Traditional multi-tank synchronous control methods rely on static equilibrium models to estimate the equivalent support reaction force at each support point, calculate the load imbalance coefficient, and then output compensation control commands. However, in the actual working conditions of multi-tank collaborative transformer handling, due to differences in the drive response of each tank, strong nonlinear shear strain occurs between the main load-bearing beam and the trolley's load-bearing components during non-uniform speed transient start-stop processes. This shear strain is highly coupled with the bending strain generated by static load, making it impossible for the strain measurement signals in the traditional static model to accurately reflect the true equivalent support reaction force distribution. This leads to severe distortion in the support reaction force identification results based on strain signals, resulting in deviations in the synchronous compensation control commands and a significant decrease in the accuracy of the shear risk assessment of the limiting structure. In other words, existing technologies suffer from insufficient accuracy in identifying the equivalent support reaction force under non-uniform speed transient start-stop conditions involving multiple handling units. Summary of the Invention

[0003] In view of this, the present invention provides a large transformer multi-transport unit collaborative bearing device and its synchronous control method, which can solve the technical problem of insufficient accuracy of equivalent support reaction force identification in the existing multi-transport unit collaborative bearing system under non-uniform speed transient start-stop conditions.

[0004] This invention is implemented as follows: This invention provides a large transformer multi-transportation unit cooperative bearing device and its synchronous control method, including the following steps: Zero-point calibration, temperature compensation parameter setting, and calibration conversion are performed on pressure sensors, strain gauges, displacement sensors, and posture sensors. The sampling period is set to collect the output signals of each sensor and perform zero-point correction, filtering, and outlier removal. Based on the actual installation position of the transition plate detected by the displacement sensor and the distance between the left and right ground tank supports, a bearing geometric model between the equipment load input point and the ground tank support point is established. Based on the equipment input load collected by the pressure sensor, the coordinates of the equipment load center and the preliminary equivalent support reaction force of the ground tanks are calculated. After temperature compensation, the strain values ​​of each strain gauge on the main load-bearing crossbeam and the trolley limit plate are extracted. Based on the linear mapping relationship between the strain vector of multiple measuring points, the equivalent support reaction force of the ground tank, and the input load of the equipment, the weighted least squares method is used to obtain the strain-corrected support reaction force. Then, the strain is superimposed with the static equilibrium estimation result according to the fusion weight to obtain the final equivalent support reaction force of the ground tank. The load imbalance coefficient of tanks in various regions, the load difference coefficient of equipment feet input load, the maximum structural stress of the main load-bearing beam, the risk coefficient of the limiting structure of the trolley limit plate, the connection slip, the low-frequency slope angle and the maximum impact acceleration are weighted and summed according to their respective weights to calculate the comprehensive bearing risk index. The comprehensive bearing risk index is then input into the adaptive feedback iterative gated graph neural dynamic network model for equivalent support reaction force residual verification and secondary correction of the comprehensive bearing risk index. Based on the comparison results of the revised comprehensive risk index and the classification threshold, the neuron memory allocation ratio function value is dynamically adjusted to determine the current risk level, and the corresponding speed, acceleration, steering or braking force synchronous compensation control quantities are output to tanks in various locations to drive tanks in various locations to execute commands to maintain the original motion, reduce the overall acceleration and issue prompts, limit speed and correct deviation or coordinate braking. After the synchronous compensation control is completed, the sensor data collected is used as the input for the next round. The above steps are repeated to form a closed-loop control process, which includes sensor data acquisition, bearing geometric parameter identification, equivalent support reaction force calculation, strain fusion correction, comprehensive risk level judgment, multi-location tank synchronous compensation control, and bearing status feedback.

[0005] The aforementioned sampling period refers to the process of gradually shortening the sampling period and analyzing the dynamic distortion rate and processor computing load during unloaded and different weight transformer handling tests, and selecting the minimum sampling period where the dynamic distortion rate is lower than the allowable value and the computing load does not exceed the processor's rated load setting ratio.

[0006] Specifically, the zero-point correction refers to the correction made at the first... The zero-point corrected output value is obtained by subtracting the zero-point output value from the current output value of each sensor in the initial static state.

[0007] Specifically, the load-bearing geometric model is established with the center line of the main load-bearing beam as the reference to establish a coordinate system, calculates the vertical load with the calibration value of the pressure sensor, corrects the coordinates of the equipment load input point and the actual support spacing of the ground tank with the detection results of the displacement sensor, and calculates the coordinates of the equipment load center by weighted average.

[0008] Specifically, the preliminary equivalent support reaction force of the tanks in various locations is obtained based on the static equilibrium relationship of the equipment input load, the coordinates of the equipment load input point, the coordinates of the equivalent support point of the tank, and the equivalent self-weight of the bearing connection mechanism.

[0009] Specifically, the linear mapping relationship is obtained through finite element analysis or static load calibration experiments to obtain the influence matrix. The weight matrix is ​​determined by the signal-to-noise ratio and calibration reliability experimental data of each strain measurement point, and measurement points with higher signal-to-noise ratios are assigned larger weights.

[0010] The fusion weight is determined by repeatedly conducting calibration experiments under various load conditions and ground flatness conditions. The errors of the static balance estimation results and strain correction results relative to the direct weighing benchmark are compared, and the fusion weight value that minimizes the error of the equivalent support reaction force after fusion is selected.

[0011] Specifically, the connection slippage amount is measured directly by a displacement sensor to determine the relative displacement between the main load-bearing beam and the trolley load-bearing component; or an equivalent slippage risk index is constructed based on the abrupt change signal of the strain gauge on the trolley limit plate, the abrupt change signal of the strain gauge on the main load-bearing beam, and the impact signal detected by the posture sensor, which serves as a substitute state quantity for the connection slippage amount.

[0012] Specifically, the adaptive feedback iterative gated graph neural dynamic network model maps multiple tanks, main load-bearing beams, and trolley limit plates as nodes in a topological graph. Edges are constructed based on physical connections and load transfer paths. A gated graph neural network architecture is used to update the node states, and a neural ordinary differential equation module is introduced to achieve continuous-time dynamic modeling.

[0013] Specifically, when the residual between the predicted equivalent support reaction force output by the deep nodes of the network and the physical calibration value exceeds a set limit, the current output layer feature vector is weighted by learnable weights and then fed back to the input layer node features to perform secondary correction inference until the residual is lower than the set limit or the maximum number of iterations is reached.

[0014] Specifically, the neuron memory allocation ratio function is obtained by weighting and summing the current comprehensive bearing risk index correction amount, the number of adaptive feedback iterations, and the ratio of the maximum structural stress to the allowable stress of the main bearing beam. The memory ratio of each gate weight matrix of the gated loop unit is dynamically allocated according to the memory adjustment coefficient range.

[0015] The classification thresholds include a first risk threshold, a second risk threshold, and a third risk threshold, which are determined by the statistical lower bound of the indicator when a minor load abnormality actually occurs, the statistical lower bound of the indicator when manual intervention is required to correct the deviation, and the statistical lower bound of the indicator when the equipment safety is directly threatened, respectively.

[0016] Specifically, the coordinated braking refers to the control system simultaneously sending braking commands to all tanks, causing them to stop in coordination at a preset deceleration. The preset deceleration is selected with the constraint that the maximum structural stress during braking does not exceed a set proportion of the allowable stress.

[0017] The filtering method employs a dynamic window filtering algorithm based on a Gaussian mixture model. The Gaussian component parameters corresponding to the static load component and the transient oscillation impact component are estimated iteratively using the expectation-maximization algorithm. The dynamic window adaptively adjusts the sliding window length according to the local variance of the signal, and the two types of components are separated and output at each sampling time.

[0018] Specifically, the sampling period is set to a range of 1–5 ms; the fusion weight is set to a range of 0.3–0.7; and the preset deceleration is set to a range of 0.02. Up to 0.1 The processor's rated load is set at 80%; the allowable stress is set at 80%; and the maximum number of adaptive feedback iterations is 3 to 5.

[0019] A second aspect of the present invention provides a large transformer multi-transport unit collaborative bearing device, including an adjustable bearing connection mechanism, a bearing status sensing component, a signal acquisition and processing component, a bearing status calculation component, a synchronous compensation control component, and a safety protection component.

[0020] This invention constructs a dynamic window filtering algorithm based on a Gaussian mixture model to separate the static load component and the transient oscillation impact component in the original sensor signal at each sampling time. This ensures that subsequent calculations of the equivalent support reaction force use only the static load component, thereby eliminating the interference of strong nonlinear shear strain generated by non-uniform transient start-stop on the identification of the static load state. Furthermore, this invention employs a weighted least squares method to fuse the strain vectors from multiple measurement points with the static equilibrium estimation results, and introduces an adaptive feedback iterative gated graph neural dynamic network model to perform secondary verification of the equivalent support reaction force residual. This ensures that the support reaction force identification result can still converge to the true mechanical state under transient start-stop disturbances. In summary, this invention solves the technical problem mentioned in the background art of insufficient accuracy in identifying the equivalent support reaction force of multiple transport units under non-uniform transient start-stop conditions. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the overall structure of the multi-location tank collaborative transport device of the present invention.

[0023] Figure 3 This is a front view of the supporting connecting device.

[0024] Figure 4 This is a side view of the load-bearing connecting device.

[0025] Figure 5 Top view of the supporting connection mechanism.

[0026] The attached diagram is labeled as follows: 1-Equipment foot; 2-Transition plate; 3-Annular gasket; 4-Main load-bearing beam; 5-Trolley load-bearing component; 6-Trolley limiting plate; 7-Ground tank; 8-Equal length double-ended stud; 9-Hex nut; 10-Flat washer; 11-M30×50 socket head cap screw; 12-M30×70 socket head cap screw; 13-M24×40 socket head cap screw; 14-M30×55 socket head cap screw; 15-Lifting lug; 16-Large washer; 17-Pressure sensor; 18-Strain gauge; 19-Displacement sensor; 20-Position-Attitude sensor. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0028] like Figure 1 The diagram shown is a flowchart of a large transformer multi-transportation unit collaborative bearing device and its synchronous control method provided by the present invention. The method includes the following steps: S01. Perform zero-point calibration, temperature compensation parameter setting and calibration conversion for pressure sensor, strain gauge, displacement sensor and posture sensor, so as to set the sampling period to collect the output signal of each sensor and perform zero-point correction, filtering and outlier removal. S02. Based on the actual installation position of the transition plate detected by the displacement sensor and the distance between the left and right ground tank supports, establish a bearing geometric model between the equipment load input point and the ground tank support point, and calculate the coordinates of the equipment load center and the preliminary equivalent support reaction force of the ground tanks based on the equipment input load collected by the pressure sensor. S03. Extract the strain values ​​of each strain gauge after temperature compensation on the main load-bearing crossbeam and the trolley limit plate. Based on the linear mapping relationship between the multi-point strain vector, the equivalent support reaction force of the ground tank, and the equipment input load, obtain the strain-corrected support reaction force using the weighted least squares method. Then, combine the results with the static equilibrium estimation results according to the fusion weight. By superimposing these forces, the final equivalent support reaction forces of tanks in various locations can be obtained. S04. The uneven load coefficient of tanks in various regions, the input load difference coefficient of equipment feet, the maximum structural stress of the main load-bearing beam, the risk coefficient of the limiting structure of the trolley limit plate, the connection slip, the low-frequency slope angle and the maximum impact acceleration are weighted and summed according to their respective weights to calculate the comprehensive bearing risk index. The comprehensive bearing risk index is then input into the adaptive feedback iterative gated graph neural dynamic network model for equivalent support reaction force residual verification and secondary correction of the comprehensive bearing risk index. S05. Based on the comparison results of the revised comprehensive risk index and the classification threshold, dynamically adjust the neuron memory allocation ratio function value, determine the current risk level, and output the corresponding speed, acceleration, steering or braking force synchronous compensation control quantity to tanks in various locations, drive tanks in various locations to execute the command to maintain the original motion, reduce the overall acceleration and issue a prompt, limit speed correction or coordinated braking. S06. After the synchronous compensation control is completed, the sensor data collected is used as the input for the next round. Steps S02 to S05 are repeated to form a closed-loop control process of sensor data acquisition, bearing geometric parameter identification, equivalent support reaction force calculation, strain fusion correction, comprehensive risk level judgment, multi-location tank synchronous compensation control and bearing status feedback.

[0029] The set sampling period is obtained experimentally: in unloaded and transformer handling tests of different weights, the sampling period is gradually shortened at intervals of 0.5ms to 10ms and multiple start-stop impact data are collected. The dynamic distortion rate of sensor data and processor computing load under each sampling period are analyzed. The minimum sampling period with a dynamic distortion rate lower than the set allowable value and a computing load not exceeding 80% of the processor's rated load is selected as the final set sampling period. The value range is generally 1 to 5ms.

[0030] The zero-point correction method is as follows: using the first The zero-point output value is obtained by subtracting the current output value of the sensor from the zero-point output value under the initial static state, thus eliminating the influence of initial installation error and sensor zero-point drift on subsequent calculations.

[0031] The load-bearing geometric model establishes a coordinate system based on the center line of the main load-bearing beam, and records the direction of the line connecting the left and right ground tanks as the transverse axis. The vertical load of the load input points of the left and right equipment is calculated using the calibration value of the pressure sensor. The coordinates of the load input points of the left and right equipment and the actual support spacing of the left and right ground tanks are corrected using the detection results of the displacement sensor. Then, the equivalent action point coordinates of the equipment input load in the transverse direction, i.e., the coordinates of the equipment load center, are calculated by weighted average.

[0032] The preliminary equivalent support reaction force of the tanks in each region is obtained based on the static balance relationship of the equipment input load, the coordinates of the equipment load input point, the coordinates of the equivalent support point of the tank, and the equivalent self-weight of the bearing connection mechanism, without the need to directly place pressure sensors at the support point of the tank.

[0033] The linear mapping relationship between the multi-point strain vector, the equivalent support reaction force of the ground tank, and the input load of the equipment is obtained by finite element analysis or static load calibration experiment to obtain the influence matrix, which is solved by the weighted least squares method. The weight matrix is ​​determined by the signal-to-noise ratio and calibration reliability experimental data of each strain measuring point. The measuring points with higher signal-to-noise ratio are given larger weights, and the measuring points with lower signal-to-noise ratio are given smaller weights.

[0034] Wherein, the fusion weight The range of values ​​is By repeatedly conducting calibration experiments under various load conditions and ground flatness conditions, comparing the errors of the static equilibrium estimation results and strain correction results relative to the direct weighing reference, the method that minimizes the error of the equivalent support reaction force after fusion was selected. The value generally ranges from 0.3 to 0.7.

[0035] The load imbalance coefficient is the ratio of the absolute value of the difference between the final equivalent support reaction forces of the left and right ground tanks to their sum, used to evaluate the load balance state of the left and right ground tanks; the equipment foot input load difference coefficient is the ratio of the absolute value of the difference between the vertical loads of the left and right equipment feet to their sum, used to evaluate the balance of the load input from the equipment feet to the load-bearing connection mechanism; the maximum structural stress is the maximum value of the stress obtained by multiplying the strain values ​​after temperature compensation at all strain measurement points by the elastic modulus; and the limiting structure risk coefficient is the ratio of the equivalent stress of the trolley limiting plate to its allowable stress.

[0036] The connection slip amount is obtained through one of the following two methods: the displacement sensor directly measures the relative displacement between the main load-bearing beam and the trolley load-bearing component; when no direct slip displacement sensor is arranged, an equivalent slip risk index is constructed based on the abrupt change signal of the strain gauge on the trolley limit plate, the abrupt change signal of the strain gauge on the main load-bearing beam, and the impact signal detected by the posture sensor, which serves as a substitute state quantity for the connection slip amount.

[0037] The low-frequency slope angle is the slow tilt component obtained by low-pass filtering the original tilt angle signal collected by the posture sensor, reflecting the slope change of the transport path; the maximum impact acceleration is the maximum absolute value of the acceleration signal collected by the posture sensor during the transport process, reflecting the unevenness of the road surface or the impact intensity of start-stop.

[0038] Among them, each weight coefficient to The following methods were used to obtain the data: multiple rounds of handling experiments were conducted, covering transformers of different weights, different road conditions, and different start-up and shutdown conditions. The correspondence between the comprehensive bearing risk index and actual bearing accidents or early warning events under each condition was collected. The analytic hierarchy process (AHP) was used in combination with the experimental data regression analysis to determine a set of weight coefficients that maximized the classification accuracy of the comprehensive bearing risk index for various risk events. The stability of the coefficients was verified through multiple iterative experiments.

[0039] The classification thresholds include a first risk threshold, a second risk threshold, and a third risk threshold, which are obtained by: artificially simulating off-center loading, tilting, slippage, and impact conditions in the handling experiment, recording the distribution range of the comprehensive bearing risk index under each condition, taking the statistical lower bound of the index when a slight bearing abnormality actually occurs as the first risk threshold, the statistical lower bound of the index when manual intervention is required to correct the deviation as the second risk threshold, and the statistical lower bound of the index when the equipment safety is directly threatened as the third risk threshold. All three thresholds are determined through statistical analysis of no less than 30 valid experimental data.

[0040] The adaptive feedback iterative gated graph neural dynamic network model is based on deep learning. Its specific structure is as follows: multiple ground tanks, main load-bearing beams, and trolley limit plates in the collaborative transport system are mapped as nodes in a topology graph. Edges are constructed using physical connections and load transfer paths to form a dynamic topology graph input structure. The graph neural network layer adopts a gated graph neural network architecture. Each node receives the hidden state vectors of its neighboring nodes, aggregates them through message passing, and inputs them to a gated loop unit for node state updates. The gated loop unit contains an input gate, a forget gate, and an output gate, which control the weight allocation of current-time feature input, historical state forgetting, and output features, respectively. A neural ordinary differential equation module is introduced between the gated loop unit layers. The hidden state of the nodes is considered as the trajectory of the solution to an ordinary differential equation in the time dimension. A vector field parameterized by the neural network replaces the discrete-time step, realizing the dynamic modeling of the ground tank's movement speed and turning angle, the trolley limit plate's shear strain, relative slip, and overall low-frequency slope angle in continuous time. An adaptive feedback iterative model is set between the gated loop unit layers within the network. In the feedback iterative loop, when the residual between the predicted equivalent support reaction force output by the deep nodes of the network and the physical calibration value obtained in step S03 exceeds the set limit, the internal iterator is activated. The current output layer feature vector is weighted according to the learnable weights and then fed back to the input layer node features to perform secondary correction inference. The iteration is performed up to 3 to 5 times until the residual is lower than the set limit or the maximum number of iterations is reached. The output layer receives the global mean pooling result of the final hidden state of all nodes and maps it to the comprehensive bearing risk index correction amount and the residual verification value of the equivalent support reaction force of tanks in various regions through the fully connected layer. The network dynamically adjusts the memory ratio of each gate weight matrix of the gated recurrent unit using the neuron memory allocation ratio function. The graph message passing computation and the integral solution of the neuronal ordinary differential equation are allocated to different CUDA streams for parallel execution using the CUDA stream allocation mechanism. The dynamic topology graph adjacency matrix and node feature cache are stored in the cache and main memory in layers using the memory allocation strategy. The inter-layer activation value adopts the gradient checkpoint mechanism to reduce the memory occupation of backpropagation, thereby realizing efficient inference of multiple nodes and multiple time steps under limited computing resources.

[0041] The technical advantages of the adaptive feedback iterative gated graph neural dynamic network model are as follows: By modeling multiple transport units and structural components as nodes in a topological graph, the model can naturally express the transmission relationship of loads in a multi-rigid-body non-uniform constraint system. The gating mechanism of the gated loop unit significantly enhances the model's ability to distinguish between transient starting impacts and steady-state load imbalances. The neural ordinary differential equation module makes continuous-time modeling of the load dynamic process possible, overcoming the distortion problem of discrete-time step modeling under non-uniform starting conditions. The adaptive feedback iterative loop enables the model to actively perform internal secondary corrections when the physical calibration residuals are too large, introducing physical constraints into the network inference process, making the prediction results of equivalent support reaction forces closer to the actual mechanical state. This effectively improves the problem of insufficient accuracy in assessing the shear risk of the constraint structure caused by inconsistent velocity responses of multiple transport units, and reduces the model's sensitivity to single-sensor noise through a closed-loop verification mechanism.

[0042] The steps for establishing the training dataset for the adaptive feedback iterative gated graph neural dynamic network model specifically include: collecting multiple sets of sensor time-series data under laboratory and field handling conditions, covering different transformer weights, different base spacings, different tank support distances, different road surface smoothness, and different start-stop conditions. These data include the vertical load of pressure sensors, strain vectors after strain gauge temperature compensation, installation positions and support spacing of displacement sensors, tilt angles and impact accelerations of pose sensors, and simultaneously recording the actual support reaction force benchmark values ​​of tanks in various locations obtained through static load weighing or calibration devices, as well as manually labeled risk level tags. The collected time-series data are aligned according to the sampling period to construct a dynamic topology graph node feature sequence, using the actual support reaction force benchmark values ​​and risk level tags as supervision signals. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio, and the features of each channel are normalized. The normalization parameters are determined by the statistics of the training set.

[0043] The specific steps for training the adaptive feedback iterative gated graph neural dynamic network model include: using the weighted sum of the mean square error loss of the equivalent support reaction force prediction error and the risk level classification cross-entropy loss as the total loss function, updating the parameters using the AdamW optimizer, and setting the initial learning rate to... to The batch size is 16–64, and the maximum number of training rounds is 100–200. In each training round, an adaptive feedback iterative loop is executed for each sample. The residual between the actual support reaction force benchmark value and the network output value is used to determine whether to trigger internal iteration. The trigger threshold is determined by statistical analysis of the residual distribution on the validation set. A learning rate cosine annealing strategy is adopted to fine-tune the model parameters with a smaller learning rate in the later stages of training. The model parameters with the smallest equivalent support reaction force prediction error on the validation set are used as the final model weights and saved. After training, the model performance is evaluated on the test set. If the test error exceeds the set allowable value, data is added and retraining is performed.

[0044] The neuron memory allocation ratio function is used to dynamically adjust the memory ratio of each gate weight matrix in the gated recurrent unit of the adaptive feedback iterative gated graph neural dynamic network model. Its inputs are the current comprehensive bearing risk index correction amount, the number of adaptive feedback iterations, and the ratio of the maximum structural stress to the allowable stress of the main bearing beam at the current moment. The memory adjustment coefficient is obtained by weighted summation of the three. ;when At that time, the input gate weight matrix was set to 40% of the video memory, and the forget gate and output gate each accounted for 30%; when At this time, the input gate weight matrix's memory usage is increased to 55%, the forget gate to 25%, and the output gate to 20%, to enhance the processing capability of new features at the current time step; when At this time, the input gate weight matrix's memory usage is further increased to 65%, the forget gate to 20%, and the output gate to 15%, prioritizing the inference accuracy of new input features under high-risk conditions; when When the time comes, the emergency memory reallocation mechanism is triggered, which transfers part of the output gate weight matrix to main memory, increases the memory usage of the input gate to 70%, the memory usage of the forget gate to 20%, and the memory usage of the output gate to 10%. At the same time, the gradient checkpoint is activated to release the memory of intermediate activation values. The above thresholds were determined by analyzing multiple experimental data on inference accuracy and memory usage under different risk levels.

[0045] The principle and technical effect of the dynamic window filtering algorithm based on the Gaussian mixture model are as follows: In the signal acquisition stage of step S01, the algorithm establishes a Gaussian mixture model for the original output sequence of the sensor, and estimates the mean, variance, and mixing weight of the Gaussian components corresponding to the static load component and the transient oscillation impact component through the expectation-maximization algorithm iteratively. The dynamic window mechanism adaptively adjusts the sliding window length according to the local variance of the current signal, expands the window to suppress noise in the signal stable stage, and shrinks the window to retain dynamic details in the transient impact stage. The static load component and the transient oscillation impact component are separated and output at each sampling time, and used for load-bearing state calculation and impact risk judgment, respectively. The algorithm solves the interference of strong nonlinear shear strain on the identification of static load-bearing state when multiple transport units start non-uniformly and transiently, so that the subsequent equivalent support reaction force calculation results are not contaminated by start-stop impact distortion, and improves the accuracy and stability of load-bearing state calculation throughout the transport process.

[0046] The principle and technical effect of the adaptive adjustment of the moving boundary conditions digital twin agent model are as follows: When establishing the bearing geometry model in step S02, the tilt angle and impact acceleration data fed back by the posture sensor in real time are used as input to correct the equivalent boundary stiffness coefficients at various tank support points in the bearing geometry model in real time; when the posture sensor detects that the slope or impact acceleration exceeds the set threshold, the digital twin agent model automatically calls the moving boundary condition update program, and re-estimates the dynamic constraint stiffness of each support point under the current road surface condition based on the mapping relationship between the sensor feedback and the pre-calibrated boundary stiffness matrix, and inputs the corrected boundary stiffness into the bearing geometry model; this method overcomes the problem that the traditional static model has increased estimation error of equivalent support reaction force due to the mismatch of constraint boundary conditions under the condition of small road surface undulations and potholes, so that the bearing geometry model can still maintain high support reaction force identification accuracy under the condition of transporting on uneven ground, and improves the adaptability and reliability of the overall closed-loop control; the set threshold and boundary stiffness matrix are obtained by joint calibration of multiple transport experiments and finite element simulation data under different road surface unevenness conditions.

[0047] The coordinated braking command is a braking command simultaneously sent by the control system to all tanks, causing them to stop in unison at a preset deceleration. The deceleration value is generally in the range of 0.02. Up to 0.1 The specific value was determined by testing the sudden change in strain of the main load-bearing beam during the braking process under different load conditions, and was selected with the constraint that the maximum structural stress during the braking process does not exceed 80% of the allowable stress.

[0048] In the closed-loop control process, the sensor data after step S06 is completed will enter step S02, replacing the sensor data collected in the previous round, and participate in the establishment of the bearing geometric model and the calculation of the equivalent support reaction force in the next round, thereby realizing continuous dynamic tracking of the bearing state throughout the entire handling process.

[0049] The present invention also provides a collaborative bearing device for multiple handling units of large power equipment, including an adjustable bearing connection mechanism, a bearing status sensing component, a signal acquisition and processing component, a bearing status calculation component, a synchronous compensation control component, and a safety protection component.

[0050] The adjustable load-bearing connection mechanism includes equipment feet, a transition plate, an annular gasket, a main load-bearing crossbeam, a trolley load-bearing component, a trolley limiting plate, fastening connectors, and a ground tank. The equipment feet transfer the weight of the equipment being transported to the transition plate; the transition plate and annular gasket adapt to different equipment bases and compensate for installation height errors; the main load-bearing crossbeam enables span load bearing and load distribution; the trolley load-bearing component is positioned below the main load-bearing crossbeam to support it and transfer the load to the corresponding ground tank; the trolley limiting plate restricts abnormal lateral, longitudinal, and vertical displacements of the main load-bearing crossbeam relative to the trolley load-bearing component; and the fastening connectors ensure reliable connection between the equipment feet, transition plate, main load-bearing crossbeam, and trolley load-bearing component.

[0051] The main load-bearing crossbeam is equipped with adjustment holes and reserved connection holes distributed along its length. The transition plate can be adjusted in position on the main load-bearing crossbeam according to the spacing of the feet of the equipment being transported, so that the equipment feet are aligned and connected with the transition plate to accommodate the base spacing of large power equipment such as circuit breakers and transformers of different specifications. After adjustment, the equipment feet, transition plate and main load-bearing crossbeam are locked and fixed by fastening fasteners, thereby forming a stable upper load-bearing connection structure.

[0052] The trolley-mounted component can be adjusted in a set direction below the main load-bearing crossbeam to change the support distance between the left and right ground tanks or multiple ground tanks, allowing the ground tank support points to be arranged according to the equipment weight, center of gravity, base size, and on-site handling conditions. After adjustment, the trolley-mounted component is locked and fixed to the main load-bearing crossbeam, and the main load-bearing crossbeam is limited and prevented from detaching by the trolley limiting plate. Through the above structure, the present invention forms an adjustable connection structure on the upper part of the main load-bearing crossbeam for adapting to different equipment bases, and an adjustable support structure on the lower part of the main load-bearing crossbeam for adjusting the ground tank support distance, solving the problems of fixed dimensions, insufficient on-site adaptability, difficult installation and positioning, and difficulty in optimizing the fulcrum spacing of existing load-bearing connection mechanisms.

[0053] The load-bearing state sensing component includes a pressure sensor, strain gauges, a displacement sensor, and a posture sensor. The pressure sensor is positioned between the equipment feet and the transition plate, or in the main pressure-bearing area of ​​the equipment base, to detect the vertical load input from the transported equipment to the load-bearing connection mechanism. The strain gauge is positioned in the equipment loading area of ​​the main load-bearing beam, the ground tank support area, and the force-bearing area of ​​the trolley limiting plate, to detect the bending deformation of the main load-bearing beam, the off-center load of the support area, and the stress state of the limiting structure. The displacement sensor is positioned in the adjustment area of ​​the transition plate, the adjustment area of ​​the ground tank support, or the relative connection area between the main load-bearing beam and the trolley load-bearing component, to detect the actual installation position of the equipment connection point, the actual support distance between the ground tanks, and the relative slippage of the connection mechanism. The posture sensor is positioned in the rigid connection area near the main load-bearing beam, the transition plate, or the equipment base, to detect the overall tilt angle, posture changes, vibration, and impact state during transport. If necessary, the control system can incorporate a temperature compensation module to correct the detection results of the pressure sensor and strain gauges to improve the accuracy of load and strain identification. The aforementioned synchronous compensation control component is used to execute the synchronous control method of the aforementioned large transformer multi-transportation unit collaborative bearing device.

[0054] The specific implementation of step S01 is as follows: Before the formal operation of the handling system, the pressure sensor, strain gauge, displacement sensor, and posture sensor are first zero-point calibrated. The output value of each sensor in the initial static no-load state is recorded as the zero-point reference. The corresponding zero-point reference is subtracted from each subsequent raw output value to eliminate the influence of installation error and sensor zero drift. The temperature compensation parameter is obtained through calibration experiments. The output deviation of each sensor under different temperature environments is fitted as a temperature coefficient. The sensor output is linearly corrected after the ambient temperature is collected in real time. The sampling period is determined experimentally. In no-load and transformer handling tests of different weights, the sampling period is gradually shortened in increments of 0.5ms. The dynamic distortion rate of sensor data and processor computing load under each period are analyzed. The minimum sampling period with a dynamic distortion rate lower than the allowable value and a computing load not exceeding 80% of the processor's rated load is selected. The actual value range is generally 1 to 5ms. Within each sampling period, a dynamic window filtering algorithm based on a Gaussian mixture model is used to process the original signal. The mean, variance, and mixing weights of the Gaussian components corresponding to the static load component and the transient oscillating impact component are estimated iteratively using the expectation-maximization algorithm. The dynamic window adaptively adjusts its sliding window length based on the local variance of the current signal, expanding the window during stable signal phases to suppress noise and shrinking it during transient impact phases to preserve dynamic details. The two types of components are separated and output at each sampling time for subsequent load-bearing state calculations and impact risk assessments, respectively. After filtering, outliers are statistically tested and removed to ensure data quality for subsequent calculations.

[0055] The specific implementation of step S02 is as follows: A plane coordinate system is established with the center line of the main load-bearing beam as the reference, and the direction of the line connecting the left and right ground tanks is defined as the transverse axis. The detection results of the displacement sensor are used to correct the actual installation position coordinates of the transition plate and the actual support spacing of the left and right ground tanks. The corrected coordinate values ​​replace the nominal design values ​​in subsequent calculations to eliminate the influence of installation deviations. The vertical loads of the left and right equipment feet are calculated using the output values ​​of the pressure sensor after zero-point correction and temperature compensation in step S01. When the posture sensor detects that the slope or impact acceleration exceeds the set threshold, the digital twin agent model for adaptive adjustment of moving boundary conditions automatically calls the moving boundary condition update program. Based on the mapping relationship between the sensor feedback and the pre-calibrated boundary stiffness matrix, the dynamic constraint stiffness of each support point under the current road surface condition is re-estimated, and the corrected boundary stiffness is input into the load-bearing geometric model. On this basis, the equivalent action point coordinates of the equipment input load in the transverse direction, i.e., the equipment load center coordinates, are calculated by weighted averaging. Based on the equipment input load, the coordinates of the equipment load input point, the coordinates of the equivalent support point of the ground tank, and the equivalent self-weight of the bearing connection mechanism, the preliminary equivalent support reaction force of the ground tank is obtained through the static equilibrium equation, without the need to directly place pressure sensors at the support point of the ground tank.

[0056] The specific implementation of step S03 is as follows: Extract the temperature-compensated strain values ​​from each strain measuring point in the loading area of ​​the main load-bearing beam, the support area of ​​the ground tank, and the stress area of ​​the trolley limit plate from the static load components separated in step S01, forming a multi-point strain vector. The linear mapping relationship between the multi-point strain vector, the equivalent support reaction force of the ground tank, and the input load of the equipment is determined by obtaining the influence matrix through finite element analysis or static load calibration experiments. The weight matrix of the weighted least squares method is determined by the signal-to-noise ratio and calibration reliability experimental data of each strain measuring point. Points with higher signal-to-noise ratios are assigned larger weights, and points with lower signal-to-noise ratios are assigned smaller weights, thereby solving for the strain-corrected support reaction force. The strain-corrected support reaction force is then combined with the static equilibrium estimation result obtained in step S02 according to the fusion weights. Perform weighted summation. The value of was determined through calibration experiments under various load conditions and ground flatness conditions, and the value that minimizes the error of the equivalent support reaction force after fusion was selected. The value is generally taken in the range of 0.3 to 0.7, from which the final equivalent support reaction force of tanks in various regions is obtained.

[0057] The specific implementation of step S04 is as follows: Seven load-bearing state indicators are calculated: load imbalance coefficient, equipment foot input load difference coefficient, maximum structural stress of the main load-bearing beam, risk coefficient of the trolley limit plate limiting structure, connection slippage, low-frequency slope angle, and maximum impact acceleration. Each indicator is assigned a weight coefficient determined through the analytic hierarchy process combined with regression analysis of experimental data. to Weighted summation yields the comprehensive bearing capacity risk index. This index is input into an adaptive feedback iterative gated graph neural network model. This model maps multiple tanks, main bearing beams, and trolley limit plates as nodes in a dynamic topology graph. A gated graph neural network architecture is used for message passing and node state updates, and a neural ordinary differential equation module is introduced to achieve continuous-time dynamic modeling. When the residual between the network output's predicted equivalent bearing reaction force and the physical calibration value obtained in step S03 exceeds a set limit, the adaptive feedback iterative loop is activated, performing a secondary correction inference, iterating a maximum of 3-5 times until the residual is below the set limit or the maximum number of iterations is reached. The corrected comprehensive bearing capacity risk index and the residual verification values ​​of the equivalent bearing reaction forces of tanks in various locations are then output.

[0058] The specific implementation of step S05 is as follows: The corrected comprehensive bearing risk index output in step S04 is compared sequentially with the first risk threshold, the second risk threshold, and the third risk threshold. All three thresholds are determined through statistical analysis of no less than 30 valid experimental data. When the index is below the first risk threshold, all tanks maintain their original movement commands; when the index is between the first and second risk thresholds, the control system reduces the overall acceleration and issues a warning; when the index is between the second and third risk thresholds, speed limiting correction is implemented; when the index is above the third risk threshold, a coordinated braking command is sent to all tanks, and all tanks stop coordinatedly at a preset deceleration. The preset deceleration is selected with the constraint that the maximum structural stress during braking does not exceed 80% of the allowable stress, and its general range is 0.02. Up to 0.1 The calculation of the synchronous compensation control quantity is based on the current risk level. The proportion of memory occupied by the weight matrix of each gate of the gated recurrent unit is dynamically adjusted through the neuron memory allocation ratio function to ensure inference accuracy under high-risk conditions.

[0059] The specific implementation of step S06 is as follows: After each synchronous compensation control is executed, the system immediately re-acquires the output signals of the pressure sensor, strain gauge, displacement sensor and posture sensor at a set sampling period, replaces the previous data with the newly acquired data, enters step S02 to re-establish the bearing geometric model and calculate the equivalent support reaction force, and executes all operations from step S02 to step S05 in sequence, forming a closed-loop control process of sensor data acquisition, bearing geometric parameter identification, equivalent support reaction force calculation, strain fusion correction, comprehensive risk level judgment, multi-location tank synchronous compensation control and bearing status feedback, so as to realize continuous dynamic tracking of the bearing status throughout the entire transportation process.

[0060] It should be noted that the key technologies of this invention include: a dynamic window filtering algorithm based on a Gaussian mixture model achieves the separation of static bearing components and transient impact components, solving the problem of strain signals being contaminated by start-stop impacts in traditional methods. This is because the Gaussian mixture model can parameterize the multimodal probability distribution of the signal, the expectation-maximization algorithm makes the mean and variance of the two types of components gradually converge during the iteration process, and the dynamic window mechanism further ensures the time resolution of the impact stage; the weighted least squares multi-source fusion method establishes a linear mapping between the strain vector of multiple measurement points and the equivalent support reaction force through the influence matrix, and the signal-to-noise ratio weighting mechanism amplifies the contribution of high-quality measurement points, effectively suppressing the influence of single-point anomalies on the overall identification results; the adaptive feedback iterative gated graph neural dynamic network model introduces physical topological relationships into the graph structure, the gated loop unit distinguishes between transient and steady-state characteristics, the neural ordinary differential equation realizes continuous-time modeling, and the adaptive iterative loop introduces physical residual constraints into the reasoning process. The three work together to make the support reaction force prediction continuously approach the real mechanical state under dynamic working conditions. The three key technologies are progressively implemented at the signal layer, computation layer, and model layer, forming a complete physical constraint closed loop. This enables the system to maintain high accuracy in identifying equivalent support reaction forces and synchronous compensation control throughout the entire non-uniform transient start-up and shutdown process.

[0061] It should be noted that during the coordinated transport of transformers by multiple tanks, when the speed responses of the drive units in different tanks vary due to differences in mechanical characteristics or control responses, a relative displacement trend will occur between the main load-bearing beam and the trolley's load-bearing components, leading to local shear stress concentration at the trolley's limiting plate. Traditional methods typically rely on static models to estimate the force on the limiting plate or place a small number of strain gauges at key sections for monitoring. However, because static models ignore the dynamic internal forces caused by differences in the speed responses of tanks in different locations, and the placement of a small number of static strain gauges cannot cover the spatial distribution of transient internal forces, the shear risk assessment results of the limiting structure are underestimated, making it difficult to issue early warnings. The reason for the above technical problem is that when the speed responses of tanks in different locations are inconsistent, the main load-bearing beam no longer bears pure static bending moments and support reactions, but is superimposed with dynamic shear internal forces caused by speed differences. These internal forces are highly coupled with the bending strain generated by static bearing, and traditional static models cannot separate the two. Therefore, the true value of the shear risk of the limiting plate is systematically underestimated by the static estimate. The usual solution to the above technical problems is to increase the density of strain gauges or to place multi-axial force sensors at the limiting plate. However, since the relative motion of tanks in different locations is a dynamic process during transportation, static multi-point monitoring still cannot capture the time-varying characteristics of transient internal forces. Furthermore, the calibration of a large number of sensors and the synchronous processing of signals pose significant difficulties for on-site implementation. This invention effectively solves this technical problem. On the one hand, it separates the transient oscillation impact component through Gaussian mixture model dynamic window filtering, enabling the transient component of the limit plate shear strain to be independently extracted and input into risk calculation. On the other hand, the adaptive feedback iterative gated graph neural dynamic network model incorporates the tank's movement speed and steering angle, the vehicle's limit plate shear strain, and relative slip into the dynamic topology graph node features. It models the dynamic internal force evolution process over continuous time using neural ordinary differential equations, enabling the model to perceive the dynamic internal force trend caused by inconsistent speed response. In the comprehensive bearing risk index, the risk coefficient weight of the limit structure reflects the true magnitude of the limit plate shear risk in a timely manner, thus outputting speed limit correction or coordinated braking commands in a timely manner before the risk reaches the warning threshold. This effectively avoids the problem of delayed warning of limit structure failure caused by insufficient dynamic internal force estimation in traditional methods.

[0062] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems lies in the fact that its technical solution works synergistically from three levels: signal separation, multi-source fusion, and closed-loop verification. Each level is logically progressive, forming a complete physical constraint closed loop. At the signal separation level, the Gaussian mixture model establishes a probability distribution model for the original sensor output sequence using the expectation-maximization algorithm. It assigns different Gaussian components to the static load component and the transient oscillation impact component, respectively. A dynamic window mechanism adaptively adjusts the sliding window length based on the signal's local variance, expanding the window to suppress noise during the steady-state phase and shrinking it to retain dynamic details during the impact phase. This mechanism ensures from the source that the strain signal entering the support reaction calculation stage does not contain start-stop impact distortion, making subsequent calculations physically valid. At the multi-source fusion level, the weighted least squares method, based on the linear mapping relationship between the multi-point strain vector and the equivalent support reaction force obtained through finite element analysis or static load calibration experiments, determines the weight matrix using the signal-to-noise ratio and calibration reliability of each measuring point, achieving a weighted fusion of the strain correction result and the static equilibrium estimation result. This fusion mechanism fully utilizes strain information from measurement points with high signal-to-noise ratios, while suppressing the propagation of errors from single sources through fusion weights, thus improving the overall robustness of support reaction force identification under various load conditions. At the closed-loop verification level, the adaptive feedback iterative gated graph neural dynamic network model models multiple platform tanks, main load-bearing beams, and trolley limit plates as dynamic topology graph nodes. The gating mechanism of the gated loop unit distinguishes between transient starting impacts and steady-state load imbalances, and the neural constant differential equation module achieves continuous-time dynamic modeling, overcoming the distortion problem of discrete-time step modeling under non-uniform starting conditions. When the residual between the network output's equivalent support reaction force prediction value and the physical calibration value exceeds a set limit, the adaptive feedback iterative loop is activated, weighting the current output features and feeding them back to the input node for secondary correction inference. This introduces physical constraints into the network inference process, making the final support reaction force identification result closer to the actual mechanical state. The synergistic effect of the above three levels enables the present invention to maintain a high accuracy of equivalent support reaction force identification throughout the entire process of non-uniform transient start-stop, thereby ensuring the accuracy of synchronous compensation control commands and solving the problem of insufficient identification accuracy in the prior art from a technical principle perspective.

[0063] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0064] The specific implementation of step S01 is as follows: zero-point calibration is performed on the pressure sensor, strain gauge, displacement sensor, and posture sensor, and the output value of each sensor in the initial static state is recorded as the zero-point reference value. , No. The zero-point correction output values ​​of each sensor are: ; In the formula, For the first The current raw output value of each sensor. This is the output value after zero-point correction to eliminate the effects of initial installation errors and sensor zero-point drift. The sampling period was determined experimentally: in unloaded and transformer handling tests of different weights, the sampling period was gradually shortened at intervals of 0.5ms to 10ms, and the dynamic distortion rate of the sensor data under each sampling period was analyzed. and processor computing load Select Below the set allowable value and The minimum sampling period, not exceeding 80% of the processor's rated load, shall be used as the final set sampling period. The value range is typically 1–5 ms. Zero-point correction, filtering, and outlier removal are sequentially performed on the acquired signal. The filtering uses a dynamic window filtering algorithm based on a Gaussian mixture model: for the first… Output sequence after zero-point correction of each sensor Establish a Gaussian mixture model, with the mixture probability density function as follows: ; In the formula, The number of Gaussian components. For the first The mixed weights of the components, and satisfying , For the first The mean of each component, For the first The variance of each component, For For the mean, The variance is represented by a Gaussian probability density function. These parameters are estimated iteratively using an expectation-maximization algorithm, separating the static load component from the transient oscillatory impact component for use in load-bearing state calculations and impact risk assessments, respectively. Dynamic window length. Based on the current signal local variance Adaptive adjustment: when When the value is below a set stability threshold, the window is expanded to suppress noise. When the impact threshold is exceeded, the window is reduced to preserve dynamic details. This represents the current sampling time.

[0065] The specific implementation of step S02 is as follows: establish a coordinate system with the center line of the main load-bearing crossbeam as the reference, and denote the direction of the line connecting the left and right tanks as the transverse direction. The vertical load at the load input points of the left and right equipment is calculated based on the pressure sensor calibration value. and The lateral coordinates of the load input point of the left device are corrected based on the detection results of the displacement sensor. Horizontal coordinates of the load input point on the right device and the actual support spacing between the left and right tanks Lateral coordinates of equipment load center The weighted average is calculated as follows: ; Initial support reaction force of the left-side tank and the initial support reaction force of the right-side tank Based on the static equilibrium relationship, the formula is expressed as follows: ; ; In the formula, To bear the equivalent self-weight of the connecting mechanism, To support the lateral coordinate of the center of gravity of the connecting mechanism, and The combined forces form a preliminary equivalent support reaction vector. This method eliminates the need to directly place pressure sensors at the tank's support points. When the attitude sensor detects that the slope or impact acceleration exceeds a set threshold, the digital twin agent model automatically invokes the moving boundary condition update program, based on the sensor feedback and the pre-calibrated boundary stiffness matrix. Based on the mapping relationship, the dynamic constraint stiffness of each support point is re-estimated. ( (Numbering the support points), and inputting the corrected boundary stiffness into the bearing geometry model to improve the accuracy of support reaction force identification under uneven ground conditions, where... The data was obtained through multiple handling experiments under different road surface unevenness conditions and joint calibration with finite element simulation data.

[0066] The specific implementation of step S03 is as follows: Extract the strain values ​​of each strain gauge after temperature compensation from the main load-bearing crossbeam and the trolley limiting plate to form a multi-point strain vector: ; In the formula, The total number of strain measurement points. ( ) is the first Strain values ​​at individual measuring points after temperature compensation. The strain vector at multiple measuring points and the equivalent support reaction force vector of the tank. and device input load vector The linear mapping relationship between them is: ; In the formula, for The column submatrix corresponding to the support reaction force in the dimensional influence matrix. for The column submatrix corresponding to the device input load in the dimensional influence matrix. For the number of ground tanks, The number of equipment load input points. for The measurement error vector and its influence matrix are obtained through finite element analysis or static load calibration experiments. A weight matrix is ​​defined. , For the first The weighted least squares weights for each measuring point are determined by the signal-to-noise ratio and calibration reliability experimental data for each point. Points with higher signal-to-noise ratios are assigned larger weights. The weighted least squares method is used to solve for the strain-corrected support reaction vector. for: ; The final equivalent support reaction force vector of tanks in various regions By fusion weights The result of superposition is: ; In the formula, Through repeated calibration experiments under various load conditions and ground flatness conditions, comparisons were made. and Compared to the error of the direct weighing datum, the one that minimizes the error of the equivalent support reaction force after fusion is selected. value.

[0067] The specific implementation method of step S04 is: comprehensive risk-bearing index Defined as a weighted sum of all sub-indicators according to their respective weights, the formula is as follows: ; In the formula, is the load imbalance coefficient, where and The left and right tanks represent the final equivalent support reaction forces, calculated as follows: ; Input the load difference coefficient for the equipment feet, calculated as follows: ; In the formula, , These represent the vertical loads on the left and right equipment feet, respectively. The maximum structural stress at all strain measurement points is calculated by multiplying the temperature-compensated strain value at each measurement point by the elastic modulus. get, The allowable stress of the main load-bearing beam This is the dimensionless stress ratio, used to participate in the weighted summation after unifying the dimensions with other dimensionless sub-indices. The weighting coefficients are... Absorb the corresponding dimension conversion factor. The risk coefficient of the trolley limit plate limiting structure is calculated as follows: ; In the formula, For the equivalent stress of the car's limiting plate, The allowable stress of the trolley limit plate. To measure the slippage, the relative displacement between the main load-bearing beam and the trolley load-bearing component is directly measured using a displacement sensor. When no direct slippage displacement sensor is used, an equivalent slippage risk index is constructed as a substitute state quantity based on the abrupt changes in strain gauge signals on the trolley limit plate, the abrupt changes in strain gauge signals on the main load-bearing beam, and the impact signals detected by the posture sensor. Its dimensions are similar to... Consistency, weighting coefficient Absorb the corresponding dimension conversion factor. The low-frequency slope angle is represented by the slow tilt component obtained after low-pass filtering of the original tilt angle signal from the pose sensor. This represents the maximum absolute value of the acceleration signal collected by the posture sensor during the handling process. This is a dimensionless comprehensive risk indicator with weighting coefficients. , , Absorb separately , , Dimension conversion factor to All were determined using the analytic hierarchy process (AHP) combined with regression analysis of data from multiple rounds of handling experiments. The input is an adaptive feedback iterative gated graph neural network model. This model maps multiple tank platforms, main load-bearing beams, and trolley limit plates as nodes in a topological graph. Edges are constructed based on physical connections and load transfer paths. The graph neural network layer adopts a gated graph neural network architecture. In the Layer message aggregation vector Obtained by aggregating the hidden states of neighboring nodes: ; In the formula, For nodes The set of neighboring nodes, For nodes In the The hidden state vector of the layer. The update formulas for each gate and state within the gated loop unit are: ; ; ; ; ; In the formula, For input gate output, Output for the forget gate. For output gate output, In the candidate hidden state, For nodes In the The hidden state vector after layer update, for function, For element-wise multiplication, , , , For learnable weight matrix, This represents the vector concatenation operation. A neural frequent differential equation module is introduced between the gated recurrent unit layers to model the hidden state of the nodes in the time dimension as follows: ; In the formula, For a vector field parameterized by a neural network, Using its learnable parameters, a numerical integrator is used to solve the trajectory, enabling dynamic modeling of the ground-attack tank's velocity, steering angle, strain, and tilt angle over continuous time. In the adaptive feedback iterative loop, when the deep node outputs the predicted equivalent support reaction force... Compared with the physical calibration value obtained in step S03 The residual satisfies At this time, the internal iterator is activated, and the current output layer feature vector is adjusted according to the learnable weights. The weighted features are backpropagated to the input layer nodes for secondary correction inference, iterating a maximum of 3-5 times until the residual is lower than the threshold value. Or reach the maximum number of iterations, The value is determined statistically through the residual distribution of the validation set. The output layer, after global mean pooling, is mapped to a comprehensive risk-bearing index correction value via a fully connected layer. And the equivalent support reaction force residual values ​​of tanks in various regions, the final corrected comprehensive bearing risk index is as follows: ; Neuron memory allocation scaling function Adaptive feedback iteration count and the ratio of the maximum structural stress to the allowable stress of the current main load-bearing beam. The weighted summation yields the memory adjustment coefficient: ; In the formula, , , The weighting coefficients were determined through analysis of experimental data on inference accuracy and memory usage under various risk levels. This is a dimensionless memory adjustment coefficient. absorb Dimension conversion factor absorb Dimension conversion factor. When At that time, the input gate weight matrix was set to 40% of the video memory, and the forget gate and output gate each accounted for 30%; when At that time, the input gate's memory usage increased to 55%, the forget gate to 25%, and the output gate to 20%; when At that time, the input gate's memory usage increased to 65%, the forget gate to 20%, and the output gate to 15%; When this happens, an emergency memory reallocation mechanism is triggered, increasing the memory allocation of the input gate to 70%, the memory allocation of the forget gate to 20%, and the memory allocation of the output gate to 10%. At the same time, gradient checkpoints are activated to release intermediate activation value memory.

[0068] The specific implementation of step S05 is as follows: ... Compared with the first risk threshold Second risk threshold and the third risk threshold The system compares and determines the current risk level based on the comparison results, then outputs corresponding speed, acceleration, steering, or braking force synchronization compensation control values ​​to tanks in various locations, driving them to perform the corresponding actions. The deceleration value range of the coordinated braking command is typically [missing value]. to The specific value is selected based on the constraint that the maximum structural stress during braking does not exceed 80% of the allowable stress.

[0069] The specific implementation of step S06 is as follows: After the synchronous compensation control is completed, the sensor data collected is used as the input for the next round. Steps S02 to S05 are repeated to form a closed-loop control process of sensor data acquisition, bearing geometric parameter identification, equivalent support reaction force calculation, strain fusion correction, comprehensive risk level judgment, multi-location tank synchronous compensation control and bearing status feedback, so as to realize continuous dynamic tracking of the bearing status throughout the entire transportation process.

[0070] To better understand and implement this invention, the following is a specific application scenario of the invention, embodiment 2: This embodiment takes the in-plant transfer and handling operation of a three-phase power transformer with a rated weight of 220t as the object, and adopts a method of four ground tanks cooperating to carry the load, with two on the left and two on the right symmetrically arranged. The main load-bearing crossbeam and the trolley load-bearing component form a load-bearing connection mechanism to implement the synchronous control method of the invention.

[0071] Before handling, technicians first calibrated and zero-point-calibrated the load-bearing status sensing components. Pressure sensors were installed at four points between the equipment feet and the transition plate, and strain gauges were installed at 14 measuring points, covering the loading area of ​​the main load-bearing beam, the tank support area, and the force-bearing area of ​​the trolley limit plate. Displacement sensors were installed in the adjustment area of ​​the transition plate and the relative connection area between the main load-bearing beam and the trolley load-bearing components. A posture sensor was installed in the rigid connection area in the middle of the main load-bearing beam. The output values ​​of each sensor in the initial static no-load state were recorded as the zero-point reference. The temperature compensation coefficient was obtained through an environmental temperature calibration experiment. The actual handling environment temperature was approximately 32℃, and the temperature correction coefficient was entered into the signal acquisition and processing components according to the calibration value. Through experiments of gradually shortening the sampling period, the set sampling period for this handling was determined to be 2ms. The dynamic distortion rate was lower than the set allowable value, and the processor's computational load was approximately 73% of the rated load, meeting the requirement of not exceeding 80%. The overall structure of the device is as follows: Figure 2 As shown, the front view of the load-bearing connection device is as follows: Figure 3 As shown.

[0072] During the signal acquisition phase, a dynamic window filtering algorithm based on a Gaussian mixture model processes the raw output sequences of each sensor. The Gaussian component parameters of the static load component and the transient oscillating impact component are estimated iteratively using an expectation-maximization algorithm. The dynamic window adaptively adjusts the sliding window length based on the signal's local variance, separating the two types of components at each sampling time. During the transformer handling and startup phase, due to differences in drive response, the measured speed response time difference among the four ground tanks was approximately 80ms. During start-up and shutdown, strain gauges at the trolley's limit plate detected significant transient shear strain abrupt changes. The dynamic window filtering algorithm successfully separated this transient component from the static bending strain component, ensuring that the strain signal used in subsequent calculations of the equivalent support reaction force was free from impact distortion.

[0073] During the geometric model establishment phase, displacement sensor readings showed that the installation position of the left transition plate deviated by 3mm from its nominal position. After displacement correction, the coordinates of the left and right equipment feet and the actual support spacing of the four ground tanks were updated. After zero-point correction and temperature compensation, the vertical load distribution at the four installation points is shown in Table 1. Table 1 Vertical Load Distribution of Equipment Legs

[0074] Table 1 shows that the total vertical load at the four installation points is 2195 kN, which matches well with the gravity value corresponding to the rated weight of 220 t. Through weighted average calculation, the load center coordinates of the equipment are slightly offset to the right by about 12 mm in the lateral direction, a slight deviation from the design center, which has been incorporated into the load-bearing geometry model correction. The posture sensor detected a slope of approximately 1.2° in the middle of the transport path. The digital twin agent model automatically invoked the moving boundary condition update program, re-estimated the dynamic constraint stiffness of each support point based on the pre-calibrated boundary stiffness matrix, and output the preliminary equivalent support reaction force of the tank at each location after correction. The side view of the load-bearing connection device is shown below. Figure 4 As shown.

[0075] In the equivalent support reaction force calculation stage, the strain values ​​of the static load components after temperature compensation at 14 strain measurement points are extracted to form a multi-point strain vector. The strain-corrected support reaction force is solved using the influence matrix obtained from finite element analysis and the weighted least squares method. The weight matrix is ​​determined by the signal-to-noise ratio of each measurement point. (Weight fusion) The value was determined to be 0.45 through calibration experiments. The strain-corrected support reaction force was then weighted and superimposed with the static equilibrium estimation results to obtain the final equivalent support reaction force of the four platform tanks, as shown in Table 2. Table 2. Final equivalent support reaction force of tanks in various regions

[0076] As shown in Table 2, the total support reaction force of the two ground tanks on the left is 1080kN, and the total support reaction force of the two ground tanks on the right is 1115kN. The load imbalance coefficient is 0.032, which is at a low level, indicating that the current load condition is basically balanced, but there is a slight imbalance on the right side, which is consistent with the calculation result of the equipment load center being offset to the right.

[0077] In the comprehensive bearing risk index calculation stage, the following factors are weighted: load imbalance coefficient, equipment foot input load difference coefficient, maximum structural stress of the main bearing beam (the maximum value obtained by multiplying the strain values ​​of 14 measuring points by the elastic modulus is 186MPa, the allowable stress is 240MPa, and the ratio is 0.775), risk coefficient of the trolley limit plate limit structure, connection slippage (directly measured by displacement sensor), low-frequency slope angle (1.2°), and maximum impact acceleration. to Weighted summation yields the current comprehensive bearing capacity risk index value. This index is input into an adaptive feedback iterative gated graph neural dynamic network model. The model maps six nodes—four platform tanks, the main bearing beam, and the trolley limit plate—to a dynamic topology graph. Continuous-time dynamic modeling is completed through the gated graph neural network architecture and the neural constant differential equation module, outputting the equivalent support reaction force residual verification value and the comprehensive bearing capacity risk index correction amount. In this inference process, the adaptive feedback iterative loop is triggered once, and the residual drops below the set limit after the second iteration. A top view of the bearing connection mechanism is shown below. Figure 5 As shown.

[0078] Based on the comparison between the revised comprehensive bearing capacity risk index and the classification threshold, the comprehensive bearing capacity risk index of this transport slope section is between the first and second risk thresholds. The control system outputs a synchronous compensation control quantity to the four ground tanks to reduce the overall acceleration and issue a warning. The overall acceleration limit is automatically calculated by the system based on the current risk level. After execution, a new round of sensor data immediately enters step S02, and the closed-loop control continues to run. The risk thresholds and the distribution of this index are shown in Table 3. Table 3. Thresholds for the Classification of Comprehensive Bearing Risk Indicators and the Actual Measured Values.

[0079] As shown in Table 3, the measured value of the comprehensive bearing risk index of the slope section during this transport was 0.38, which is within the slightly abnormal range. The control system promptly output a speed reduction warning, and the transport process passed through the slope section smoothly without any limit structure over-limit or coordinated braking events.

[0080] The components to be described with reference to the attached drawings include: equipment feet (1), transition plate (2), annular gasket (3), main load-bearing crossbeam (4), trolley load-bearing component (5), trolley limiting plate (6), ground tank (7), equal-length double-ended studs (8), hexagonal nuts (9), flat washers (10), internal hexagonal head screws, lifting lugs (15), and large washers (16). When the weight of a large transformer is large, the span of the base is large, or the on-site handling path requires high support stability, multiple sets of load-bearing connection mechanisms can be used in parallel, and the number of sensors and control output objects should be increased accordingly based on the number of sets of load-bearing connection mechanisms.

[0081] In this embodiment, the ground tank (7) serves as a power-bearing unit, with multiple ground tanks (7) arranged below the main load-bearing beam (4) to jointly support and transport the large transformer. The main load-bearing beam (4) spans the tops of multiple ground tanks (7) to form an integral load-bearing platform and distribute the load of the large transformer to each ground tank (7). The trolley load-bearing component (5) is located between the ground tank (7) and the main load-bearing beam (4) to receive the main load-bearing beam (4) and transfer the load to the corresponding ground tank (7). The trolley limiting plate (6) is located on the side or top of the trolley load-bearing component (5) to limit the lateral, longitudinal, and vertical abnormal displacement of the main load-bearing beam (4) relative to the trolley load-bearing component (5), thereby improving the anti-slip capability and anti-detachment safety of the load-bearing connection mechanism during the transport process.

[0082] The top surface of the ground tank (7) is provided with a positioning guide rail. The trolley carrier (5) is installed on the positioning guide rail by high-strength fasteners to achieve lateral and longitudinal positioning of the trolley carrier (5) relative to the ground tank (7). The trolley carrier (5) is provided with a bearing groove or settling groove that matches the main bearing beam (4). The main bearing beam (4) is embedded in the bearing groove or settling groove so that the load on the main bearing beam (4) can be stably transmitted to the trolley carrier (5), and then transmitted from the trolley carrier (5) to the ground tank (7). When the main bearing beam (4) is placed in place, the trolley limiting plate (6) covers or abuts against the side and top of the main bearing beam (4) so ​​that the main bearing beam (4) remains in a restricted state under conditions of starting and stopping, turning, uneven road surface or local impact, to avoid separation, movement or impact between the main bearing beam (4) and the trolley carrier (5).

[0083] The main load-bearing crossbeam (4) is provided with a set of adjustment holes distributed along the length direction and reserved connection holes. The transition plate (2) and the trolley load-bearing component (5) can be connected to the main load-bearing crossbeam (4) by fasteners. During the placement or installation adjustment of the large transformer, the transition plate (2) can be adjusted on the main load-bearing crossbeam (4) according to the actual spacing of the equipment feet (1) to adapt to the spacing of different large transformer equipment feet (1) and on-site hoisting errors; the trolley load-bearing component (5) can be adjusted on the main load-bearing crossbeam (4) according to the equipment weight, center of gravity position and on-site handling conditions to change the support distance between the left and right ground tanks or multiple ground tanks. After adjustment, the equipment feet (1), transition plate (2), main load-bearing beam (4) and trolley load-bearing components (5) are locked and fixed by equal-length double-ended studs (8), hexagonal nuts (9), flat washers (10), large washers (16) and corresponding internal hexagonal head screws, so that the load-bearing connection mechanism changes from the adjustable connection state in the installation and adjustment stage to the rigid load-bearing state in the handling stage.

[0084] The transition plate (2) is set between the large transformer equipment foot (1) and the main load-bearing crossbeam (4), and the annular gasket (3) is set between the transition plate (2) and the equipment foot (1), or between the transition plate (2) and the main load-bearing crossbeam (4), to compensate for the installation height error between the equipment foot (1) and the main load-bearing crossbeam (4), so that the self-weight of the large transformer can be smoothly transferred step by step through the equipment foot (1), transition plate (2), annular gasket (3), main load-bearing crossbeam (4), trolley load-bearing component (5) and ground tank (7). The end of the main load-bearing crossbeam (4) is provided with connection holes and lifting lugs (15). When it is necessary to move a large transformer with a larger weight or a larger span, adjacent main load-bearing crossbeams (4) can be connected by connecting rods, bolts or auxiliary connection structures to form multiple sets of parallel load-bearing structures to improve the overall load-bearing capacity and bending stiffness. Through the above structure, the present invention can adjust the connection position before the large transformer is placed and form a rigid load-bearing connection after locking. It can not only adapt to on-site installation deviations, but also ensure the load-bearing stability and anti-detachment safety during the collaborative transportation process of multiple ground tanks (7).

[0085] In this embodiment, a set of load-bearing connection mechanisms corresponds to two ground tanks, denoted as left ground tank T1 and right ground tank T2 respectively; the two ends of the main load-bearing crossbeam (4) are supported by the left trolley load-bearing component and the right trolley load-bearing component respectively, forming the left equivalent support point A1 and the right equivalent support point A2. It should be noted that the number of sensors does not correspond one-to-one with the number of ground tanks. Multiple sensors can be arranged at different stress points of the same load-bearing connection mechanism to obtain information on equipment input load, support position, crossbeam strain, connection slippage, attitude change and vibration impact from multiple measuring points. The controller first integrates the data from multiple sensor measuring points into the overall state quantity of the load-bearing connection mechanism, and then generates the corresponding synchronous compensation control quantity for left ground tank T1 and right ground tank T2 based on the overall state quantity.

[0086] (17) A pressure sensor is set between the equipment foot (1) and the transition plate (2), or in the main pressure-bearing area of ​​the equipment foot (1), to detect the vertical load input from the large transformer to the bearing connection mechanism. (18) Strain gauges are set in the equipment loading area of ​​the main bearing beam (4), the left and right ground tank support areas, and the force-bearing area of ​​the trolley limit plate (6), to detect the bending deformation of the main bearing beam (4), the off-center load of the support area, and the stress state of the limit structure. (19) A displacement sensor is set in the adjustment area of ​​the transition plate (2), the adjustment area of ​​the ground tank support, or the relative connection area between the main bearing beam (4) and the trolley bearing component (5), to detect the actual installation position of the equipment foot (1) or the transition plate (2) relative to the main bearing beam (4), the actual support distance between the left and right ground tanks, and the relative slippage of the bearing connection mechanism during the handling process. (20) The position sensor is set in the rigid connection area near the main load-bearing crossbeam (4), transition plate (2) or equipment foot (1) to detect the overall tilt angle, posture change, vibration and impact state during the handling process. If necessary, the controller can combine the temperature compensation module to perform temperature correction on the detection results of the (17) pressure sensor and (18) strain gauge to reduce the impact of ambient temperature changes on the load and strain identification results.

[0087] The technological advancements of this invention compared to traditional synchronous control methods are reflected in the following aspects: Traditional static models cannot distinguish between the strain components generated by transient start-stop impacts and static loads, leading to distortion in the identification results of support reaction forces under non-uniform start-up conditions. This invention, however, uses a Gaussian mixture model with dynamic window filtering to separate the two types of components at the probability distribution level, ensuring the identification results are physically valid throughout the start-stop process. Traditional methods rely on a single information source for support reaction force estimation, making it difficult to suppress the influence of single-point sensor anomalies. This invention, through a weighted least squares multi-source fusion mechanism, fuses the strain vectors from multiple measurement points with the static equilibrium estimation results according to the signal-to-noise ratio, improving the robustness of the estimation results from the information fusion level. Traditional risk assessment models cannot perceive the dynamic internal force trends caused by inconsistent tank speed responses in different locations. This invention's gated graph neural dynamic network, through graph topology modeling and continuous-time modeling of neural ordinary differential equations, explicitly expresses the time-varying process of dynamic internal forces. The adaptive iterative loop further ensures that the model output closely approximates the real mechanical state through physical residual constraints, fundamentally improving the timeliness and accuracy of shear risk assessment for restrained structures.

[0088] It should be noted that the variables involved in this invention are explained in detail in Tables 4 and 5.

[0089] Table 4. Variable Explanation Table (Part 1)

[0090] Table 5. Variable Explanation Table (Part Two)

[0091] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A synchronous control method for a large transformer multi-transportation unit collaborative bearing device, characterized in that, Includes the following steps: Zero-point calibration, temperature compensation parameter setting, and calibration conversion are performed on pressure sensors, strain gauges, displacement sensors, and posture sensors. The sampling period is set to collect the output signals of each sensor and perform zero-point correction, filtering, and outlier removal. Based on the actual installation position of the transition plate detected by the displacement sensor and the distance between the left and right ground tank supports, a bearing geometric model between the equipment load input point and the ground tank support point is established. Based on the equipment input load collected by the pressure sensor, the coordinates of the equipment load center and the preliminary equivalent support reaction force of the ground tanks are calculated. After temperature compensation, the strain values ​​of each strain gauge on the main load-bearing crossbeam and the trolley limit plate are extracted. Based on the linear mapping relationship between the strain vector of multiple measuring points, the equivalent support reaction force of the ground tank, and the input load of the equipment, the weighted least squares method is used to obtain the strain-corrected support reaction force. Then, the strain is superimposed with the static equilibrium estimation result according to the fusion weight to obtain the final equivalent support reaction force of the ground tank. The load imbalance coefficient of tanks in various regions, the load difference coefficient of equipment feet input load, the maximum structural stress of the main load-bearing beam, the risk coefficient of the limiting structure of the trolley limit plate, the connection slip, the low-frequency slope angle and the maximum impact acceleration are weighted and summed according to their respective weights to calculate the comprehensive bearing risk index. The comprehensive bearing risk index is then input into the adaptive feedback iterative gated graph neural dynamic network model for equivalent support reaction force residual verification and secondary correction of the comprehensive bearing risk index. Based on the comparison results of the revised comprehensive risk index and the classification threshold, the neuron memory allocation ratio function value is dynamically adjusted to determine the current risk level, and the corresponding speed, acceleration, steering or braking force synchronous compensation control quantities are output to tanks in various locations to drive tanks in various locations to execute commands to maintain the original motion, reduce the overall acceleration and issue prompts, limit speed and correct deviation or coordinate braking. After the synchronous compensation control is completed, the sensor data collected is used as the input for the next round. The above steps are repeated to form a closed-loop control process, which includes sensor data acquisition, bearing geometric parameter identification, equivalent support reaction force calculation, strain fusion correction, comprehensive risk level judgment, multi-location tank synchronous compensation control, and bearing status feedback.

2. The synchronization control method according to claim 1, characterized in that, The set sampling period refers to the process of gradually shortening the sampling period and analyzing the dynamic distortion rate and processor computing load during unloaded and different weight transformer handling tests, and selecting the minimum sampling period where the dynamic distortion rate is lower than the allowable value and the computing load does not exceed the set ratio of the processor's rated load.

3. The synchronization control method according to claim 2, characterized in that, The zero-point correction is specifically based on the first... The zero-point corrected output value is obtained by subtracting the zero-point output value from the current output value of each sensor in the initial static state.

4. The synchronization control method according to claim 3, characterized in that, The load-bearing geometric model is specifically established by using the center line of the main load-bearing beam as a reference to establish a coordinate system, calculating the vertical load using the calibration value of the pressure sensor, correcting the coordinates of the equipment load input point and the actual support spacing of the ground tank using the detection results of the displacement sensor, and calculating the coordinates of the equipment load center by weighted averaging.

5. The synchronization control method according to claim 4, characterized in that, The preliminary equivalent support reaction force of the tanks in various locations is specifically obtained based on the static equilibrium relationship of the equipment input load, the coordinates of the equipment load input point, the coordinates of the equivalent support point of the tank, and the equivalent self-weight of the load-bearing connection mechanism.

6. The synchronization control method according to claim 5, characterized in that, The linear mapping relationship is specifically obtained by finite element analysis or static load calibration experiments to obtain the influence matrix. The weight matrix is ​​determined by the signal-to-noise ratio of each strain measurement point and the experimental data of calibration reliability. Measurement points with higher signal-to-noise ratios are assigned larger weights.

7. The synchronization control method according to claim 6, characterized in that, The fusion weight is determined by repeatedly conducting calibration experiments under various load conditions and ground flatness conditions. The errors of the static balance estimation results and strain correction results relative to the direct weighing reference are compared, and the fusion weight value that minimizes the error of the equivalent support reaction force after fusion is selected.

8. The synchronization control method according to claim 7, characterized in that, The connection slippage amount is specifically measured by directly measuring the relative displacement between the main load-bearing beam and the trolley load-bearing component using a displacement sensor; or by constructing an equivalent slippage risk index based on the abrupt change signal of the strain gauge on the trolley limit plate, the abrupt change signal of the strain gauge on the main load-bearing beam, and the impact signal detected by the posture sensor, as a substitute state quantity for the connection slippage amount.

9. The synchronization control method according to claim 8, characterized in that, The adaptive feedback iterative gated graph neural dynamic network model specifically maps multiple tanks, main load-bearing beams, and trolley limit plates as nodes in a topological graph. Edges are constructed based on physical connections and load transfer paths. A gated graph neural network architecture is used to update the node states, and a neural ordinary differential equation module is introduced to achieve continuous-time dynamic modeling.

10. A multi-transportation unit collaborative bearing device for large transformers, characterized in that, It includes an adjustable load-bearing connection mechanism, a load-bearing status sensing component, a signal acquisition and processing component, a load-bearing status calculation component, a synchronous compensation control component, and a safety protection component.

Images