Special-shaped pipe forming compensation method and system based on digital twinning

Abstract

The invention discloses a special-shaped pipe forming compensation method and system based on digital twinning, and belongs to the technical field of computer-aided manufacturing and intelligent control. The method comprises the steps that multi-source sensing data in the special-shaped pipe forming process is obtained, and time-space fusion processing is executed; generating a time-space aligned multi-source sensing data set, carrying out dynamic calibration on the parameterized virtual digital twin, generating a calibrated virtual digital twin, carrying out simulation calculation on the forming process of the special pipe in a future time window, outputting predicted deviation data, and combining with a preset compensation strategy to generate a compensation instruction through logical judgment and calculation; and obtaining an adjustment result after the compensation instruction is applied, and updating the compensation strategy by using a reinforcement learning algorithm. According to the method, prospective simulation prediction is carried out by constructing the dynamically calibrated virtual digital twins, closed-loop optimization is carried out on a compensation strategy by utilizing a reinforcement learning algorithm, and active, high-precision and self-adaptive compensation control on the special pipe forming process can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer-aided manufacturing and intelligent control technology, in particular to a special-shaped tube forming compensation method and system based on digital twinning. BACKGROUND

[0002] At present, as a component with complex cross-section and spatial bending shape, the special-shaped tube plays a key role in high-end industrial fields such as aerospace, automobile manufacturing and energy equipment. Its forming precision directly affects the assembly quality and service performance of the final product. Digital twinning technology, as a data processing and simulation technology that realizes the deep integration of the physical world and the information world, provides a new way of thinking for the monitoring, analysis and optimization control of complex manufacturing processes by establishing a hyper-real dynamic virtual model of the physical entity.

[0003] In the practice of special-shaped tube forming, in order to deal with process defects such as material springback and cross-section distortion, the existing technical solutions usually rely on offline programming and trial-and-error method. The operator presets a set of compensation parameters based on experience or offline finite element analysis results, and performs trial production. After the first product is made, offline detection is performed by a three-coordinate measuring machine or other equipment. If the tolerance is found to be exceeded, the process parameters are manually adjusted again, and the trial production process is repeated. Some production lines with high automation may use a static compensation table based on historical data to query the preset compensation value according to the target shape, but this compensation model cannot be changed once it is set.

[0004] The existing technical solutions have obvious limitations. The trial-and-error method relying on human experience not only has low production efficiency and long debugging period, but also cannot guarantee the consistency and reliability of the compensation results. The method based on offline simulation or static compensation table cannot effectively deal with the dynamic uncertainties that exist in actual production, such as fluctuations in mechanical properties between batches of raw materials, equipment wear, and changes in environmental temperature, etc. These factors can cause the preset compensation value to fail. In addition, the traditional post-processing measurement method cannot capture the key stress evolution information during the forming process, and the mechanism of the defect is not well understood, resulting in a lack of physical basis for the compensation strategy, which often compromises and is difficult to achieve coordinated control of complex deviations. SUMMARY

[0005] To solve the above problems, the present application provides a special-shaped tube forming compensation method and system based on digital twinning, which uses a dynamic calibration virtual digital twin to perform forward-looking simulation prediction, and uses a reinforcement learning algorithm to optimize the compensation strategy in a closed loop, thereby achieving active, high-precision and adaptive compensation control of the special-shaped tube forming process.

[0006] The above object can be achieved by the following scheme: A method and system for compensating for the forming of a special-shaped tube based on digital twinning, comprising: acquiring real-time multi-source sensing data in the forming process of the special-shaped tube, and performing spatio-temporal fusion processing on the multi-source sensing data to generate a spatio-temporal aligned multi-source sensing data set; based on the multi-source sensing data set, dynamically calibrating a preset parameterized virtual digital twin to generate a calibrated virtual digital twin; using the calibrated virtual digital twin to simulate the forming process of the special-shaped tube in a preset future time window, and outputting prediction deviation data; according to the prediction deviation data and a preset compensation strategy, generating a compensation instruction through logical judgment and calculation; acquiring an adjustment result after applying the compensation instruction, and updating the compensation strategy based on the adjustment result using a preset reinforcement learning algorithm.

[0007] Optionally, the generating of the multi-source sensing data set comprises: acquiring geometric shape data of the forming process of the special-shaped tube through a laser scanning process; acquiring stress distribution data of the forming process of the special-shaped tube through a strain measurement process; and fusing the geometric shape data and the stress distribution data in time and space to generate a multi-source sensing data set.

[0008] Optionally, the generating of the calibrated virtual digital twin comprises: comparing simulation output data of the parameterized virtual digital twin with the multi-source sensing data set to calculate an error vector; inversely mapping and analyzing the error vector to generate a parameter adjustment value; updating internal parameters of the parameterized virtual digital twin based on the parameter adjustment value to generate the calibrated virtual digital twin.

[0009] Optionally, the outputting of the prediction deviation data comprises: acquiring current process parameters of the forming process of the special-shaped tube; inputting the process parameters as boundary conditions to the calibrated virtual digital twin; performing numerical solving in the future time window, and comparing the solving result with a preset design target geometry to output the prediction deviation data.

[0010] Optionally, the generating of the compensation instruction comprises: identifying a deviation type based on analysis of the prediction deviation data; automatically locating a key deviation area according to the deviation type; assigning a weight value to the key deviation area based on the compensation strategy, and calculating a compensation amount in combination with the prediction deviation data to generate the compensation instruction.

[0011] Optionally, the locating of the key deviation area according to the deviation type comprises: classifying based on the deviation type to distinguish between bending springback deviation and cross-section distortion deviation; performing size quantification based on the bending springback deviation and the cross-section distortion deviation to obtain a deviation quantification value; comparing the deviation quantification value with a corresponding preset deviation threshold to locate the key deviation area.

[0012] Optionally, the compensation instruction is converted into an adjustment amount of a physical process parameter through preset conversion logic; and adjustment result data caused by the application of the process parameter adjustment amount is obtained; Optionally, the updating of the compensation strategy comprises: measuring actual shape data of the adjustment result, comparing the actual shape data with the design target geometry data to obtain residual deviation data; determining a relevant reward value based on the size of the residual deviation data; calculating a strategy optimization value by using the reward value and the reinforcement learning algorithm; and updating the compensation strategy by using the strategy optimization value.

[0013] Optionally, the method further comprises: establishing a data transmission link between a physical device and an upper computing platform running the calibration virtual digital twin through a high-speed data communication network; uploading the multi-source sensing data in real time based on the data transmission link; and issuing the compensation instruction in real time based on the data transmission link to realize the synchronization of the calculation state of the calibration virtual digital twin and the real-time data of the profiled pipe forming process.

[0014] Based on the same inventive concept, the present application also provides a profiled pipe forming compensation system based on digital twinning, which comprises: a multi-source data acquisition module for acquiring real-time multi-source sensing data in a profiled pipe forming process and performing spatio-temporal fusion processing on the multi-source sensing data to generate a spatio-temporally aligned multi-source sensing data set; a virtual twin calibration module for dynamically calibrating a preset parameterized virtual digital twin based on the multi-source sensing data set to generate a calibrated virtual digital twin; a process simulation and prediction module for simulating the forming process of the profiled pipe in a preset future time window by using the calibrated virtual digital twin and outputting prediction deviation data; a compensation instruction generation module for generating a compensation instruction by logical judgment and calculation according to the prediction deviation data and a preset compensation strategy; and a strategy optimization and updating module for obtaining an adjustment result after the application of the compensation instruction and updating the compensation strategy by using a preset reinforcement learning algorithm based on the adjustment result.

[0015] Compared with the prior art, the present application has the following advantages: 1. The present application realizes the prospective prediction and active control of the forming process by constructing a closed-loop system for real-time interaction between a physical entity and a virtual model. The system can predict the shape and magnitude of manufacturing deviation before it actually occurs by using a dynamically calibrated digital twin to deduce the future forming result, thereby changing the logic of error correction and improving the initiative and timeliness of control by transforming the traditional passive compensation based on post-measurement into active intervention based on pre-prediction.

[0016] 2. The present application establishes a high-fidelity virtual digital twin through spatiotemporal fusion processing and reverse mapping analysis of multi-source sensing data. The twin not only maintains high consistency with the physical entity in macro geometry, but also in internal physical fields such as stress distribution. This deep data fusion and model calibration method ensures that the system's understanding of the current state of the heterogeneous tube is comprehensive and accurate, providing a solid and reliable data foundation for subsequent simulation prediction and compensation decisions, thereby improving the accuracy of the entire compensation system.

[0017] 3. The present application introduces a reinforcement learning algorithm, giving the compensation system the ability to learn and continuously optimize itself. The system can autonomously adjust and optimize its internal compensation strategy based on the actual effect of each compensation action, i.e., residual deviation. This adaptive closed-loop learning mechanism enables the system to continuously accumulate experience from interactions with the physical world, dynamically adapt to material property fluctuations, equipment operating conditions, and other uncertainty factors, achieving self-evolution of compensation ability and enhancing the system's robustness and intelligence level.

[0018] 4. The present application proposes a fine deviation diagnosis and targeted compensation method. By classifying and identifying predicted deviations and locating key areas, the system can decompose complex mixed errors and assign weights and calculate compensation amounts for different types and areas of deviations. This intelligent decision-making process makes the compensation behavior more targeted and efficient, allowing limited adjustment resources to be accurately applied to the most critical defects, avoiding negative effects of global adjustment, and improving the accuracy and resource utilization efficiency of compensation.

[0019] Other features and advantages of the present application will be set forth in the following description, and in part will be apparent from the description, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and attained by the structure particularly pointed out in the description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0021] Figure 1 is a flowchart of a heterogeneous tube forming compensation method based on digital twin according to an embodiment of the present application.

[0022] Figure 2 is a dynamic perception diagram of geometric deviation and stress level according to an embodiment of the present application.

[0023] Figure 3 is a prediction deviation trend graph in a future time window of an embodiment of the present application.

[0024] Figure 4 is a difference graph before and after the profiled pipe surface stress distribution compensation of an embodiment of the present application.

[0025] Figure 5 is a structural schematic diagram of a profiled pipe forming compensation system based on digital twinning of an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0027] With reference to Figure 1 , one embodiment of the present application proposes a profiled pipe forming compensation method based on digital twinning. The method adopts a dynamic calibration virtual digital twin to perform forward-looking simulation prediction, and utilizes a reinforcement learning algorithm to perform closed-loop optimization on a compensation strategy, so as to realize active, high-precision and adaptive compensation control of the profiled pipe forming process.

[0028] The method of the embodiment specifically includes: acquiring real-time multi-source sensing data in a profiled pipe forming process, and performing spatiotemporal fusion processing on the multi-source sensing data, thereby generating a spatiotemporally aligned multi-source sensing data set; based on the multi-source sensing data set, performing dynamic calibration on a preset parameterized virtual digital twin, thereby generating a calibrated virtual digital twin; utilizing the calibrated virtual digital twin to perform simulation calculation on the profiled pipe forming process in a preset future time window, and outputting prediction deviation data; according to the prediction deviation data and a preset compensation strategy, generating a compensation instruction through logical judgment and calculation; acquiring an adjustment result after application of the compensation instruction, and updating the compensation strategy by utilizing a preset reinforcement learning algorithm based on the adjustment result.

[0029] Specifically, first, a parameterized virtual digital twin is dynamically calibrated by fusing the geometry data and stress distribution data obtained from multiple sensors, so that it can accurately map the real state of the physical entity. Then, this calibrated virtual digital twin is used to conduct a forward-looking physical simulation of the forming process in the future short time window, and potential forming deviations are predicted. According to the prediction results and the preset compensation strategy, compensation instructions are generated. Finally, through a reinforcement learning framework, the actual effect of the compensation instructions is continuously evaluated, and the results are used as feedback to continuously optimize the compensation strategy, thereby realizing a complete intelligent closed loop from perception, prediction, decision-making to learning. It changes the traditional passive and lagging error compensation mode to an active and feedforward deviation prediction and correction mode, fundamentally improving the control ability of uncertain factors in the complex profiled pipe forming process. Through real-time calibration and accurate prediction of the digital twin, intervention measures can be taken before manufacturing defects are formed, effectively suppressing the generation and accumulation of errors.

[0030] Optionally, the generating a multi-source sensing data set comprises: obtaining geometry data of the profiled pipe forming process through a laser scanning process; obtaining stress distribution data of the profiled pipe forming process through a strain measurement process; fusing the geometry data and the stress distribution data in time and space to generate a multi-source sensing data set.

[0031] Specifically, to generate the spatio-temporally aligned multi-source sensing dataset, two parallel data acquisition tasks need to be performed first. The first task is to obtain the geometry data of the profiled tube forming process. This process uses a line laser scanner to perform high-speed scanning on the outer surface of the profiled tube in forming, ensuring that the scanning operation is triggered at the key time nodes of the forming process or at a set time frequency, so as to capture the dynamically changing profile of the tube. The scanning result is a series of discrete three-dimensional coordinate point sets, i.e. point cloud data, which accurately describe the actual external morphology of the profiled tube at a specific time. The second task is to obtain the stress distribution data of the profiled tube forming process. This process is also obtained by a strain measurement process, usually using digital image correlation method technology. A random speckle pattern is prepared on the surface of the profiled tube, and a synchronous triggered binocular or multi-view high-speed camera group is used to shoot the speckle image on the surface of the tube at the same time when the geometry data is collected. By continuously tracking the displacement and deformation of these speckles in the image, combined with the constitutive relationship model of the material, the full-field strain distribution on the surface of the tube can be calculated, and then the stress distribution data can be calculated. The stress distribution data describes the internal stress state of each point on the surface of the tube at the time, such as in the form of von Mises equivalent stress field. After obtaining the geometry data and stress distribution data at the same time, the two need to be fused in time and space. Time fusion triggers the three-dimensional measurement device and the high-speed camera group synchronously by a unified master clock signal, ensuring that the two groups of data have completely consistent time stamps, eliminating the error caused by time delay. Spatial fusion is to register the data in different sensor coordinate systems to the same global coordinate system. Usually, the coordinate system established by the three-dimensional measurement device is taken as the reference, and the rigid transformation matrix between the camera device and the scanner device is obtained by pre-calibration, so as to accurately map the calculated stress field data to the corresponding position of the three-dimensional point cloud model. The final result of fusion, i.e. the multi-source sensing dataset, is a structured data set. For the multi-source sensing dataset representing the profiled tube at the time at the spatial point has: wherein, Figure 2 ​​​​As shown, the gray solid line represents the difference between the actual shape of the profiled pipe and the design target shape. During the forming process, the geometry will change in real time due to material deformation, springback, and other factors, resulting in deviation, and this curve simulates the fluctuation of such real-time deviation; the gray dashed line represents the mechanical state of the material inside the profiled pipe during the forming process, and the stress distribution is the internal physical cause of defects such as springback and distortion. This curve simulates the real-time evolution of the stress level during the forming process. The technical solution realizes comprehensive and high-fidelity perception of the profiled pipe forming physical process by spatiotemporal fusion of the geometry data and the stress distribution data. Compared with single geometry measurement, this method additionally obtains the stress distribution information which is the key internal factor leading to defects such as springback and distortion. The comprehensive data of multiple physical fields provides a complete data basis for the accurate calibration of the subsequent digital twin, and helps to improve the forming precision and quality stability of the profiled pipe.

[0032] Optionally, the generating the calibrated virtual digital twin comprises: Comparing the simulation output data of the parameterized virtual digital twin with the multi-source sensing data set to calculate an error vector; Reverse mapping and analyzing the error vector to generate a parameter adjustment value; Updating the internal parameters of the parameterized virtual digital twin based on the parameter adjustment value to generate the calibrated virtual digital twin.

[0033] Specifically, the simulation output data of the pre-set parametric virtual digital twin is first precisely compared with the aforementioned multi-source sensor dataset to quantify the differences between them. The parametric virtual digital twin is a high-fidelity finite element model. First, based on the CAD design model of the shaped tube, a three-dimensional geometric model of the tube is constructed in professional finite element software. Shell elements or solid elements are typically used for meshing, and the mesh density is defined to balance computational accuracy and efficiency. To simulate the complex shape and forming characteristics of the shaped tube, locally refined meshes can be used for critical areas. Second, based on the structural characteristics of the shaped tube forming equipment, a three-dimensional geometric model of the forming mold, such as a bending mold or a clamping mold, is constructed. The shaped tube is usually set as a rigid body to simplify calculations. Then, the contact conditions between the tube and the mold are set, including contact types such as surface-to-surface or point-to-surface contact, normal contact behavior, and tangential friction. The initial material constitutive model parameters, such as elastic modulus, yield strength, hardening model parameters, and Poisson's ratio, are preset as adjustable internal parameters. The initial values ​​of these internal parameters can be determined based on material supplier data, empirical data, or offline experimental data, providing a starting point for subsequent dynamic calibration. At the same time point in the forming process, the parameterized virtual digital twin is run to obtain a set of simulation output data, including simulated geometric data and stress distribution data. This set of simulation data is spatially mapped point-by-point to a multi-source sensor dataset acquired from the physical world, and an error vector is calculated. For the error objective function... ,have: , in, It is a set of adjustable internal parameters for a virtual digital twin, including the material's hardening index, strength coefficient, and friction coefficient. and These are parameterized virtual digital twins in terms of parameters The simulation results of the first Geometric coordinates and stress of each node; The weighting coefficients for geometric errors are adjusted based on offline experiments to evaluate the model calibration effect under different weighting configurations. The specific values ​​are chosen to balance the degree of matching between the simulation results and the intrinsic physical fields such as stress distribution. The weighting coefficient for stress error is used to adjust its contribution to the overall objective function, and is set similarly to the weighting coefficient for geometric error; summation sign. This indicates that all comparison points are summed. represents the square of the Euclidean norm, i.e. the sum of squares of errors, for quantifying the difference. After the error vector is obtained by tabular quantization of the error objective function, it is input as the objective function of the optimizer, such as advanced optimization algorithms like gradient descent, Newton method, particle swarm optimization or genetic algorithm, etc. to perform inverse mapping analysis to generate parameter adjustment values for calibration. In the inverse mapping analysis process, the internal adjustable parameters of the parameterized virtual digital twin are regarded as the decision variables of the optimization algorithm. For example, the hardening index and friction coefficient of the material can be directly adjusted as input parameters of the optimization algorithm. The optimization algorithm iteratively adjusts these internal parameters, runs the virtual digital twin for simulation, and calculates the new error objective function value. This process is repeated until the error vector converges to an acceptable minimum value, at which point the output parameter set is the optimal parameter, and the change in the initial parameter is the parameter adjustment value. Finally, based on the calculated parameter adjustment value, the corresponding parameters in the parameterized virtual digital twin are updated, thereby generating a calibrated virtual digital twin. By simultaneously calibrating the geometric and stress dimensions of physical quantities, it is ensured that the twin not only maintains consistency with the real process in macroscopic morphology, but also in internal physical mechanisms, greatly improving the physical fidelity and prediction reliability of the model, enabling it to more accurately deduce complex behaviors in future forming processes, such as springback deformation, etc.

[0034] Optionally, the output prediction deviation data includes: Obtaining the current process parameters of the profiled pipe forming process; Inputting the process parameters as boundary conditions to the calibrated virtual digital twin; Numerical solving within the future time window and comparing the solving result with the preset design target geometry to output prediction deviation data.

[0035] Specifically, first, the current process parameters of the profiled tube forming process need to be obtained. These process parameters are a direct set of instructions that drive the physical equipment to perform the forming action, such as the current bending angle setting value of the numerical control pipe bender, the feeding speed, the mold pressure, or the laser power and scanning path in laser scanning forming, etc. These data are usually read in real time from the controller of the equipment, such as programmable logic controller (PLC) or computer numerical control (CNC), through industrial bus protocols. After obtaining the current process parameters, they need to be input as boundary conditions into the calibrated virtual digital twin generated in the previous stage. This step is to map the control intention of the physical world into the driving condition of the virtual simulation environment. For example, the bending angle setting value is converted into an accurate rotational displacement applied to the corresponding tool model; the feeding speed is defined as the translational speed boundary of the pipe model in the specified direction. Since the calibrated virtual digital twin already contains the geometric, stress, and material properties synchronized with the current physical entity state, taking the current process parameters as its initial and boundary conditions ensures that the starting point of the simulation calculation is completely consistent with the true state of the physical process. Subsequently, in this calibrated virtual digital twin, numerical solving is performed for a preset future time window. The future time window is a short, forward-looking time period, such as the forming time representing the next bending segment, which can be set through offline trial and error, sensitivity analysis, and in combination with the minimum time unit of process operation. For example, through sensitivity analysis of a specific pipe, TC4 titanium alloy pipe with bending radius R = 100 mm and wall thickness T = 2 mm, it can be found that setting the time window to 1.2 times the predicted forming time of the next key forming action, a value between 0.5 seconds and 5 seconds, can achieve the best balance between prediction accuracy and real-time calculation burden. Within this future time window, the finite element solver gradually calculates the deformation process, stress evolution, and final unloading springback shape of the pipe in the future short period of time based on the material constitutive model, contact algorithm, and set boundary conditions through explicit or implicit dynamic analysis. This solving result is a high-fidelity physical prediction of the future forming result based on the current state and current process parameters. Finally, the predicted forming result obtained by numerical solving, especially the geometry after unloading springback, is compared with the preset design target geometry to generate the prediction deviation data. The design target geometry is the ideal three-dimensional model defined in the product design stage. The comparison process is computationally implemented by comparing the node coordinates of the prediction result with the corresponding points on the surface of the design target geometry to calculate the spatial deviation vector of each node. For the calculation of the prediction deviation vector of the i-th node , , , , , , , wherein is the predicted three-dimensional coordinate of the i-th node on the surface of the pipe obtained after numerical solving; is the three-dimensional coordinate of the i-th node on the surface of the design target geometry; and is the spatial deviation vector of the i-th node. is the ideal three-dimensional coordinate corresponding to the design target geometry at the first node. The set of these deviation vectors constitutes the predicted deviation data, which visually presents in the form of a three-dimensional data field that in the current process parameters, future forming results will deviate from the design requirements in which areas and in what form. As shown in Figure 3 , the horizontal axis is the future time window, representing the simulation prediction of the digital twin for the future short period of time, such as the next bending segment, before the actual forming action is performed; the vertical axis is the predicted deviation, in which the predicted springback angle deviation is represented by a gray solid line, showing the extent to which the bending angle of the pipe deviates from the design target due to the springback effect after the forming action is completed; the predicted cross-sectional distortion deviation is represented by a dark gray dashed line, showing the extent to which the cross-sectional shape of the pipe, such as the ovality, deviates from the design specification; the trend of the curve simulates the possible trend of these deviations in the future forming process. By extrapolating the future in the virtually calibrated digital twin, this method can change passive error measurement to active deviation prediction. This feedforward prediction mechanism based on physical simulation provides key decision-making basis for formulating accurate and timely online compensation strategies, thereby more effectively suppressing the accumulation of forming errors and improving the success rate of one-time forming of complex profiled pipes and the geometric accuracy of the final product.

[0036] Optionally, the generating compensation instructions comprises: identifying the deviation type based on analyzing the predicted deviation data; automatically locating the key deviation area according to the deviation type; assigning a weight value to the key deviation area based on the compensation strategy, and calculating the compensation amount in combination with the predicted deviation data to generate the compensation instructions.

[0037] Specifically, first, the morphological and topological features of the predicted deviation data, which is a three-dimensional vector field, are analyzed to identify the deviation types. By calculating the components of the deviation vector in the normal and tangential directions of the pipe cross-section, the cross-sectional distortion deviation, which is mainly radial contraction or expansion, and the overall displacement mainly along the normal of the bending plane, i.e. the bending springback deviation, can be distinguished. This classification process can decompose complex mixed deviations into several basic modes with clear physical meaning. After identifying the deviation types, the key deviation regions are automatically located according to the deviation types. This step is achieved by quantifying different types of deviations and comparing them with the pre-set engineering tolerance threshold, which is a static value pre-set by engineers or designers based on the product function requirements, assembly accuracy requirements, industry standards and the actual ability of existing manufacturing processes defined in the product design stage. For example, for the bending springback deviation, the predicted springback angle or the maximum deviation of the centerline is calculated; for the cross-sectional distortion deviation, the cross-sectional ovality or the profile error at a specific position is calculated. Compare these deviation quantification values with the corresponding pre-set engineering tolerance threshold. Any region that exceeds the engineering tolerance threshold, such as a specific bending section or a straight line section, is marked as a key deviation region. Next, based on the pre-set compensation strategy, each identified key deviation region is assigned a corresponding weight value. The compensation strategy can be initialized in multiple ways: it can be an initial static lookup table established according to the results of offline finite element simulation, which is established from offline finite element simulation data or historical production experience data, as shown in Table 1, Table 1 Initial static lookup table As when the predicted rebound angle deviation is 1.5°, the corresponding preset over-bending compensation amount is +1.6°; when the cross-section ovality deviation is 1.0%, the corresponding boost force adjustment amount is +3%. These initial strategies provide a reasonable starting point for subsequent continuous optimization based on reinforcement learning, avoiding the inefficiency of completely random exploration from scratch. Further, the compensation strategy can be modeled as a complex mapping function, such as a polynomial regression model or a small feedforward neural network. Its input is the identified deviation types such as bending rebound, cross-section distortion and their quantitative values, as well as the current process parameters, and the output is the preliminary compensation parameter adjustment direction and amplitude, such as over-bending amount increase value, boost force adjustment percentage, etc. Even in the absence of a large amount of historical data, preliminary training can be carried out through small-scale offline finite element simulation data. The weight values are allocated based on multiple factors, including the severity of the region deviation, its importance in the overall part function, and its sensitivity to different process parameter adjustments. Regions with high weight values, 0.7-1.0, mean that their deviations need to be corrected first and given priority, while regions with low weight values, 0.1-0.3, can be given secondary or slight compensation; the weight value is positively correlated with the deviation quantitative value, and a piecewise linear function can be defined, with the weight set to 0.5 when the deviation quantitative value reaches 50% of the threshold, 0.8 when it reaches 100% of the threshold, and 1.0 when it reaches 150% of the threshold. Then, combined with the allocated weight values and the predicted deviation data in the region, the specific compensation amount is calculated. The calculation process is driven by the core algorithm defined in the compensation strategy, which aims to solve one or a set of process parameter adjustment amounts to effectively offset the predicted deviations. For the calculation of compensation amount , there is: , wherein, represents a complex mapping function defined by the compensation strategy; is the weight value allocated for the key deviation region ; is the predicted deviation data vector set in the region. Finally, the calculated compensation amount is packaged into a specific format instruction, i.e. compensation instruction, ready to be issued to the physical device for execution. For example, Figure 4As shown, (a) shows the stress distribution without compensation, which reflects the stress concentration that may occur in the key forming area of the profiled pipe without compensation or insufficient compensation, the darker part, or uneven distribution; (b) shows that after applying the compensation method of the present application, the stress distribution on the surface of the profiled pipe becomes more uniform, and the overall stress level, that is, the color distribution, tends to be mild. This shows that the compensation strategy effectively relieves the internal mechanical state that leads to defects, thereby helping to achieve more accurate shape control. The technical solution can obtain the best correction effect with the least adjustment cost, avoiding the introduction of new manufacturing errors due to excessive compensation or improper compensation, thereby more stably improving the forming accuracy of the profiled pipe.

[0038] Optionally, the locating the key deviation area according to the deviation type comprises: classifying and processing based on the deviation type to distinguish the bending springback deviation and the cross-section distortion deviation; performing size quantification based on the bending springback deviation and the cross-section distortion deviation to obtain a deviation quantification value; comparing the deviation quantification value with a corresponding preset deviation threshold value to locate the key deviation area.

[0039] Specifically, first, classification processing is performed based on the deviation type to decompose the complex predicted deviation data into two independent components with clear physical meaning, i.e., the bending springback deviation and the cross-section distortion deviation. The bending springback deviation mainly describes the deviation of the overall axis shape of the profiled pipe after bending forming from the designed target axis due to elastic recovery, which is a macroscopic and global deviation. The cross-section distortion deviation describes the geometric shape of the pipe cross-section, such as roundness, wall thickness uniformity, etc., which deviates from the designed specification, and is a local and morphological deviation. The deviation vector can be decomposed into the local coordinate system of the pipe by geometric analysis of the predicted deviation data vector field. The component along the normal direction of the bending plane mainly contributes to the bending springback, while the radial and tangential components mainly constitute the cross-section distortion. After classification, size quantification is performed on the bending springback deviation and the cross-section distortion deviation respectively to obtain deviation quantification values for comparison. For the bending springback deviation, the deviation quantification value can be the predicted springback angle change at a specific bending section, or the maximum spatial distance between the predicted pipe axis and the designed target axis. For the cross-section distortion deviation, the deviation quantification value can be the ovality, profile error or maximum inside and outside diameter deviation of the key cross-section, etc. For example, for calculating the ovality value , wherein, and are the maximum and minimum diameters of the distorted cross-section calculated according to the predicted deviation data; ​is the design nominal diameter of the section. In this way, the complex deviation patterns are transformed into a series of specific, measurable scalar values. Finally, these deviation quantification values are compared one by one with the corresponding preset deviation threshold values. The deviation threshold value is the upper limit of the acceptable error preset according to the functional requirements of the product and the process capability. The calculated springback angle change value can be compared with the maximum allowed springback angle deviation threshold value, and the ovality value can be compared with the maximum allowed ovality threshold value. Any area whose deviation quantification value exceeds its corresponding deviation threshold value, whether it is a certain bending section or a certain section of the pipe body, will be identified and marked. The collection of all marked areas constitutes the final identified key deviation area, providing an accurate target for subsequent targeted compensation. The technical solution classifies and quantifies the predicted deviation, transforming vague and general geometric deviations into clear and specific engineering problems, and can distinguish the priority and severity of the problem, providing more targeted and physically meaningful input for the development of subsequent compensation strategies.

[0040] Optionally, the obtaining the adjustment result after the application of the compensation instruction comprises: transforming the compensation instruction into an adjustment amount of a control physical process parameter through a preset conversion logic; obtaining adjustment result data caused by the actual application of the process parameter adjustment amount; Specifically, the first step is to translate the abstract compensation instructions into physical, executable actions. The compensation instructions generated in the previous stage, such as "reduce the bending springback angle by 0.5 degrees" or "reduce the cross-sectional ellipticity by 0.2%", are converted into adjustments to control physical process parameters through a pre-defined conversion logic. This conversion logic is not a simple lookup table, but a real-time reverse process model built based on advanced regression analysis or machine learning models such as neural networks and support vector regression. This model is trained using historical simulation data, such as offline FEM simulation results generated when establishing the "pre-defined compensation strategy," and actual production data, establishing a precise mapping relationship between the "expected deviation correction amount" and the "required process parameter adjustment amount." It can calculate the specific process parameter adjustments required to implement the compensation instructions based on information such as the material grade of the current pipe fitting (e.g., TC4 titanium alloy, 304 stainless steel), pipe diameter, wall thickness, forming temperature, and the current mold status. If the compensation instruction requires reducing the bending springback angle by 0.5 degrees in a certain bending section, the conversion logic model will calculate, based on known material mechanics parameters and the current process state, the necessary adjustment amount for the bending die of the tube bending machine to be increased by 0.8 degrees, or the boosting force to be increased by 200N. If the compensation instruction is to reduce the ellipticity of a certain area of ​​the cross-section, the conversion logic may convert it into specific adjustments to parameters such as die pressure, boosting speed, or even forming path. After generating specific process parameter adjustments, these adjustments will be applied to the forming process of the shaped tube. This means that the updated set of process parameters, including these adjustments, will be sent to the physical forming equipment, such as a CNC tube bending machine or a laser forming unit. The equipment will then execute the next or ongoing forming step according to the new instructions. For example, the tube bending machine will complete the next bending action with the adjusted bending angle. Once the process parameters containing the adjustments are applied and produce results, the resulting adjustment data needs to be acquired immediately. This step is similar to the process of generating a multi-source sensor dataset, accomplished through a sensor network deployed around the physical equipment. Specifically, after the forming action following the application of the compensation command is completed, a 3D measurement device and a DIC camera group are immediately triggered to perform high-precision measurements on the corresponding areas of the formed pipe fitting. The adjusted actual geometric shape data is obtained through 3D scanning, and the adjusted surface stress distribution data is obtained through DIC technology. These two sets of data together constitute the adjustment result data, which completely and realistically reflects the actual effect of the compensation command in the physical world. This data records the direct output of the compensation behavior and is a key basis for evaluating the compensation effect and optimizing subsequent strategies. This technical solution achieves objective and real-time evaluation of the compensation effect by constructing a complete "perception-execution-re-perception" closed loop. It ensures that the effect of each compensation action can be quantified and recorded, avoiding the disconnect between the compensation strategy and the actual process effect, and possessing an adaptive capability that continuously approaches the optimal solution.

[0041] Optionally, the updating the compensation strategy comprises: measuring actual shape data of the adjustment result, and comparing with the design target geometry data to obtain residual deviation data; calculating a relevant reward value based on the size of the residual deviation data; using the reward value and the reinforcement learning algorithm to calculate a strategy optimization value, and then using the strategy optimization value to update the compensation strategy.

[0042] Specifically, first, the actual shape data of the adjustment result obtained in the last stage is measured, and is compared with the design target geometry data as a reference. This comparison process is completed by calculating the spatial distance vector between each measurement point on the actual shaped pipe surface and its corresponding point on the design target geometry model, thereby obtaining a set of residual deviation data. This set of residual deviation data accurately quantifies the geometric gap between the pipe and the ideal shape after applying a compensation. Next, based on the size of the residual deviation data, a relevant reward value is calculated and determined. The reward value is the core feedback signal in the reinforcement learning algorithm, which is used to evaluate the pros and cons of the last compensation decision. An ideal compensation will make the residual deviation data tend to zero, and should obtain a high reward value; conversely, if the residual deviation data is large, a low reward value or a negative reward value should be obtained. The calculation function of the reward value can be designed as the inverse or inverse proportional function of the residual deviation, to ensure that the smaller the residual deviation data, the larger the reward value. For the calculation of the reward value after performing the compensation action at time t , there is: , wherein, is a large positive number, representing the maximum reward when the deviation is zero; is a positive scaling coefficient, used to adjust the sensitivity of the reward value to the error; is a function of An aggregate error scalar is calculated from the residual deviation data of the moment, which is the root mean square of the residual deviation vector module length of all measurement points. The calculated return value and the preset reinforcement learning algorithm, such as policy gradient or deep deterministic policy gradient, are used to update the compensation strategy. The reinforcement learning algorithm regards a complete operation, i.e. from the observation of the predicted deviation data, i.e. the state, to the generation of the compensation instruction, i.e. the action, and then to the acquisition of the return value, i.e. the reward, as a learning sample. A policy optimization value is calculated according to the sample. The policy optimization value is not a single numerical value, but a set of correction amounts for adjusting the internal parameters of the compensation strategy. The compensation strategy itself can be modeled as a function, such as a neural network, which inputs the predicted deviation data and outputs the compensation instruction. The policy optimization value is the adjustment amount of the neural network weight and bias. Then the policy optimization value is used to update the compensation strategy, so that when it encounters similar predicted deviation states in the future, it is more inclined to generate compensation instructions that can obtain higher return values, i.e. compensation actions that can lead to smaller residual deviations. The technical solution can learn and master the complex and nonlinear internal laws in the special pipe forming process through continuous interaction and trial and error with the physical world.

[0043] Optionally, the method further comprises: establishing a data transmission link between the physical device and the upper computing platform running the calibration virtual digital twin through a high-speed data communication network; uploading the multi-source sensing data in real time based on the data transmission link; based on the data transmission link, real-time issuing of the compensation instruction to realize the real-time data synchronization of the calculation state of the calibration virtual digital twin and the special pipe forming process.

[0044] Specifically, first, a high-bandwidth, low-latency data transmission link is established between the physical devices on site and the remote or local upper computing platform through a high-speed data communication network. The above-mentioned physical devices include the numerical control system of the profile tube forming machine tool, as well as the three-dimensional measurement device and the data acquisition unit of the stress monitoring sensor deployed thereon. The upper computing platform refers to a high-performance computer or server cluster running the calibrated virtual digital twin finite element simulation and reinforcement learning algorithm. The data transmission link can use industrial Ethernet technology, such as Ethernet control automation technology, and be supplemented by a unified communication protocol such as OPC unified architecture to ensure interoperability between heterogeneous devices and reliability of data transmission. Based on this established data transmission link, real-time bidirectional data flow is performed. The uplink data flow is responsible for real-time uploading of the multi-source sensor data. During the forming process, the massive geometric point cloud data and stress field data collected by the sensors are immediately transmitted without delay to the upper computing platform through the link after preliminary processing and packaging by the on-site data acquisition unit. This continuous data injection is a prerequisite for ensuring that the calibrated virtual digital twin can perceive and synchronize the current state of the physical world in real time. In parallel, the downlink data flow is responsible for real-time issuance of the compensation instructions. When the parameterized virtual digital twin in the upper computing platform completes the prediction and the latest compensation instructions are calculated by the compensation instruction generation module and the strategy optimization update module, the instructions will be encoded into a format that the physical device controller can parse. Subsequently, the instructions are quickly issued to the numerical control system or PLC of the forming equipment through the data transmission link. After the controller receives the instructions, it will immediately adjust the motion parameters or process settings of the actuator, such as modifying the bending angle of the next bending section or adjusting the size of the boost force, so as to seamlessly convert the optimized decisions in the virtual world into actual compensation actions in the physical world. This continuous "perception-computation-control" cycle, supported by the high-speed data transmission link, iterates at a very high time frequency, ultimately achieving synchronization of the calculated state of the corrected virtual digital twin with the real-time data of the profile tube forming process. This technical solution breaks down the barriers between information flow and physical flow in traditional manufacturing, tightly integrates perception, decision-making, and execution in a high-speed closed loop. This enables the entire compensation method to respond quickly to any minor deviations or disturbances in the forming process, improving the dynamic performance of online compensation and the final control accuracy.

[0045] Based on the same inventive concept, as Figure 5 shown, the present application also provides a profile tube forming compensation system based on digital twinning, which comprises: a multi-source data acquisition module for acquiring real-time multi-source sensor data in the profile tube forming process and performing spatio-temporal fusion processing on the multi-source sensor data to generate a spatio-temporally aligned multi-source sensor data set; a virtual twin calibration module configured to dynamically calibrate a preset parameterized virtual digital twin based on the multi-source sensing data set, and generate a calibrated virtual digital twin; a process simulation prediction module configured to simulate the forming process of the profiled pipe in a preset future time window by using the calibrated virtual digital twin, and output prediction deviation data; a compensation instruction generation module configured to generate compensation instructions by logical judgment and calculation according to the prediction deviation data and a preset compensation strategy; a strategy optimization update module configured to obtain an adjustment result after applying the compensation instructions, and update the compensation strategy by using a preset reinforcement learning algorithm based on the adjustment result.

[0046] To verify the feasibility of the application in implementation, the application is applied to a numerical control bending forming production line of a profiled pipe of a manufacturing enterprise. The profiled pipe is a key component of an engine, has a profiled structure with multiple bending and variable cross section, and has extremely strict requirements on forming precision. The traditional process often causes high scrap rate and long debugging period due to material springback and cross section distortion. The application aims to realize online and closed-loop compensation control of the forming process by a method based on digital twinning.

[0047] In this embodiment, the enterprise deploys the system described in the application beside the numerical control pipe bender. By a line laser scanner triggered synchronously and a multi-view high-speed camera group based on digital image correlation method (DIC), geometric shape data and surface stress distribution data of the pipe are obtained in real time during the forming process of each bending section, and a multi-source sensing data set aligned in time and space is generated by time and space fusion processing .

[0048] The multi-source sensing data set is received, and compared with a preset parameterized virtual digital twin containing parameters such as material hardening index 0.25 and friction coefficient 0.1. Simulation output data and are obtained by running, and an error vector is obtained by calculating the error objective function between , wherein the geometric error weight coefficient is set to 0.6, and the stress error weight coefficient is set to 0.4, to ensure the balance of the model in geometric precision and physical field fidelity. The error vector is inversely mapped and analyzed by gradient descent method or particle swarm optimization algorithm, and parameter adjustment values are iteratively generated. The material hardening index is dynamically adjusted +0.02, i.e. to 0.27, and the friction coefficient is adjusted -0.01, i.e. to 0.09, to generate a calibrated virtual digital twin highly synchronized with the physical forming process.

[0049] Subsequently, the current process parameters of the pipe bending machine, the bending angle of 88° and the feeding speed of 50 mm / s, are input as boundary conditions into the calibration virtual digital twin to quickly simulate the forming and unloading process of the next bending section, and output the predicted deviation data including the predicted springback angle and the cross-sectional ellipticity, etc. The predicted springback angle is 2.15°, and the predicted cross-sectional ellipticity is 1.8%.

[0050] The predicted deviation data is analyzed and divided into bending springback deviation and cross-sectional distortion deviation, and compared with the preset deviation threshold, such as springback angle deviation <0.1° and ellipticity <1%, to locate the key deviation area. According to the built-in compensation strategy, the key deviation area is assigned a weight value . The predicted springback angle deviation 2.15 is 21.5 times the allowed threshold 0.1°, so its springback is set to 0.95; the predicted cross-sectional ellipticity deviation 1.8% is 1.8 times the allowed threshold 1.0%, so its distortion is set to 0.75, ensuring that more serious deviations are given priority. Combined with the assigned weight value and the predicted deviation data in the key area , the compensation amount is calculated. When the predicted springback angle is 2.15° and the assigned weight value is 0.95, the over-bending compensation amount is calculated to be +2.20°; when the predicted ellipticity is 1.8% and the assigned weight value is 0.75, the boost speed adjustment amount is calculated to be +5%. Finally, the compensation instructions containing these specific adjustment amounts are generated.

[0051] The instructions are issued to the pipe bending machine PLC through high-speed industrial Ethernet for execution. The compensation instructions such as over-bending amount +2.20° are converted into adjustment amounts for controlling physical process parameters through a conversion logic based on the inverse process model, i.e. the bending angle of the pipe bending machine is adjusted from 88° to 90.20°, and the boost force is adjusted from 10kN to 10.50kN. After forming, the adjustment results are collected again and compared with the design target geometry to obtain residual deviation data. The actual measured final bending angle is 90.05°, and the final ellipticity is 1.1%, so the residual angle deviation is 0.05°, and the residual distortion deviation is +0.1%. According to the size of the residual deviation, the reward value is calculated, and the root mean square value of the aggregated residual error of the residual deviation data is 0.06. The is set to 100, and the scaling coefficient is 5, so the reward value of this round is calculated, and the result is rounded to one decimal place. The deep deterministic policy gradient reinforcement learning algorithm is used to iteratively optimize the compensation strategy.

[0052] To verify the beneficial effects of the present application, five consecutive batches of profiled pipes were selected, one piece per batch, for tracking tests, and compared with the results of the process adjusted using the traditional trial-and-error method. The key test point is the third bending segment of the profiled pipe, which is designed to have a bending angle of 90°.

[0053] Table 2 Profiled pipe forming deviation prediction and compensation instruction generation data table (taking the third bending segment as an example) Table 3 Effect and residual deviation data comparison table after compensation implementation (taking the third bending segment as an example) Table 4 Compensation strategy reinforcement learning iteration optimization data table The above Tables 2-4 record the actual application data of the present application in profiled pipe forming, and details the excellent performance of the system in deviation prediction, compensation execution and strategy self-optimization.

[0054] As can be seen from Table 2, the system can accurately predict the magnitudes of the two main deviations, springback and distortion, based on the calibrated digital twin before forming, and generate targeted and quantified compensation instructions accordingly.

[0055] The data comparison in Table 3 clearly shows the technical advantages of the present application. Compared with the nearly 1.8° angle deviation and severe cross-sectional distortion produced by the traditional trial-and-error method, the test piece applying the present application has its residual angle deviation and distortion deviation controlled within a very small range when it is first formed, with an angle deviation <0.1° and a distortion deviation <0.2%, far superior to the engineering requirements, significantly improving the first-time pass rate of the product.

[0056] Table 4 shows the strong self-learning ability of the system. Through the evaluation of each compensation result, the system calculates the return value and drives the reinforcement learning algorithm to continuously iterate the compensation strategy. From test piece #1 to #5, it can be seen that the residual deviation continuously decreases and the return value steadily increases, indicating that the compensation strategy becomes increasingly accurate and robust as the production batch increases, achieving self-evolution from "qualified" to "excellent" and effectively dealing with uncertain factors such as material batch fluctuations.

[0057] It should be noted that the electrical connections between the various units described above do not necessarily represent direct connections, and indirect connection methods can also be used as long as the purpose of the present application is achieved. The above-described examples are only exemplary embodiments of the present application and cannot limit the scope of the present application.

[0058] intended to encompass any and all embodiments of the application with equivalents as would be ascertained by those skilled in the art to which the application pertains. Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the application being indicated by the following claims.

Claims

1. A method for compensating for irregularly shaped tubes based on digital twins, characterized in that, The method includes: Real-time multi-source sensor data during the forming process of irregular tubes is acquired, and spatiotemporal fusion processing is performed on the multi-source sensor data to generate a spatiotemporally aligned multi-source sensor dataset. Based on the multi-source sensor dataset, the preset parameterized virtual digital twin is dynamically calibrated to generate a calibrated virtual digital twin. The calibration virtual digital twin is used to simulate the forming process of the irregular tube within a preset future time window, and the predicted deviation data is output. Based on the predicted deviation data and the preset compensation strategy, a compensation instruction is generated through logical judgment and calculation; Obtain the adjustment result after applying the compensation instruction, and update the compensation strategy based on the adjustment result using a preset reinforcement learning algorithm.

2. The method for compensating for irregular tube forming based on digital twins according to claim 1, characterized in that, The generated multi-source sensing dataset includes: The geometric shape data of the shaped tube forming process is obtained through laser scanning. Stress distribution data during the forming process of the shaped tube are obtained through strain measurement. The geometric data and the stress distribution data are fused in time and space to generate a multi-source sensing dataset.

3. The method for compensating for irregular tube forming based on digital twins according to claim 1, characterized in that, The generation of the calibration virtual digital twin includes: The simulation output data of the parameterized virtual digital twin is compared with the multi-source sensor dataset to calculate the error vector; The error vector is reverse-mapped and parsed to generate parameter adjustment values; The internal parameters of the parameterized virtual digital twin are updated based on the parameter adjustment values ​​to generate the calibrated virtual digital twin.

4. The method for compensating for irregular tube forming based on digital twins according to claim 1, characterized in that, The output prediction deviation data includes: Obtain the current process parameters for the forming process of the irregular tube; The process parameters are input as boundary conditions into the calibration virtual digital twin; Numerical solutions are performed within the future time window, and the solution results are compared with the preset design target geometry to output the predicted deviation data.

5. The method for compensating for irregular tube forming based on digital twins according to claim 1, characterized in that, The generated compensation instruction includes: Based on the analysis of the predicted deviation data, the type of deviation is identified; Based on the deviation type, the key deviation area is automatically located; The compensation strategy is used to assign weight values ​​to the key deviation regions, and the compensation amount is calculated by combining the predicted deviation data to generate the compensation instruction.

6. The method for compensating for irregular tube forming based on digital twins according to claim 5, characterized in that, The step of locating the key deviation region based on the deviation type includes: Based on the aforementioned deviation type, the deviations are classified into bending springback deviation and cross-sectional distortion deviation. Based on the bending springback deviation and the cross-sectional distortion deviation, magnitude quantization is performed to obtain the deviation quantization value; The deviation quantification value is compared with the corresponding preset deviation threshold to locate the critical deviation region.

7. The method for compensating for irregular tube forming based on digital twins according to claim 1, characterized in that, The process of obtaining the adjustment result after applying the compensation instruction includes: The compensation command is converted into an adjustment amount for controlling the physical process parameters through a preset conversion logic; After the process parameter adjustment amount is actually applied, the resulting adjustment result data is obtained.

8. The method for compensating for irregular tube forming based on digital twins according to claim 1, characterized in that, The updated compensation strategy includes: The actual shape data of the adjustment result is measured and compared with the geometric data of the design target to obtain residual deviation data; Based on the magnitude of the residual deviation data, the relevant return value is calculated and determined; Using the reward value and the reinforcement learning algorithm, a policy optimization value is calculated, and then the compensation policy is updated using the policy optimization value.

9. The method for compensating for irregular tube forming based on digital twins according to claim 1, characterized in that, The method further includes: A data transmission link is established between the physical device and the host computing platform running the calibration virtual digital twin via a high-speed data communication network; Based on the data transmission link, the multi-source sensor data is uploaded in real time; Based on the data transmission link, the compensation command is issued in real time to achieve real-time data synchronization between the calculation status of the calibration virtual digital twin and the forming process of the irregular tube.

10. A digital twin-based compensation system for irregularly shaped tube forming, applied to the digital twin-based compensation method for irregularly shaped tube forming as described in any one of claims 1-9, characterized in that, The system includes: The multi-source data acquisition module is used to acquire real-time multi-source sensor data during the forming process of the irregular tube, and to perform spatiotemporal fusion processing on the multi-source sensor data to generate a spatiotemporally aligned multi-source sensor dataset. The virtual twin calibration module is used to dynamically calibrate a preset parameterized virtual digital twin based on the multi-source sensor dataset, and generate a calibrated virtual digital twin. The process simulation and prediction module is used to simulate and calculate the forming process of the irregular tube within a preset future time window using the calibration virtual digital twin, and output prediction deviation data. The compensation instruction generation module is used to generate compensation instructions through logical judgment and calculation based on the predicted deviation data and the preset compensation strategy. The strategy optimization and update module is used to obtain the adjustment result after applying the compensation instruction, and update the compensation strategy based on the adjustment result using a preset reinforcement learning algorithm.

Images