Digital twinborn method for forecasting real-time bearing capacity of composite bearing cylinder structure based on field data inversion

Abstract

The invention relates to a digital twinning method for forecasting the real-time bearing capacity of a composite material bearing cylinder structure based on field data inversion, and belongs to the technical field of composite material structure health monitoring and digital twinning crossing. The method is directly oriented to health monitoring and safety evaluation of the structure during the in-service period. The core function is to sense the current damage state of the structure in real time and dynamically forecast the future bearing capacity of the structure. According to the invention, the urgent demand of instantly judging the structural safety in the key stages of launching preparation, task execution and the like of spacecrafts such as rockets is met. According to the invention, the application scene of digital twinning is changed from post-maintenance to pre-warning and monitoring in the event, and the conversion from simulation maintenance to guarantee operation is realized.

Description

Technical Field

[0001] This invention relates to a digital twin method for real-time load-bearing capacity prediction of composite load-bearing tube structures based on field data inversion, belonging to the interdisciplinary field of composite material structure health monitoring and digital twin technology. Background Technology

[0002] Overview of Existing Technologies: Composite materials, with their excellent specific strength, specific stiffness, and superior material designability, have become key materials for weight reduction and efficiency improvement in aerospace structures. However, under extreme service conditions, such as alternating loads, high-speed impacts, and harsh environments, the composite material structures of rockets and other space launch vehicles are highly susceptible to various forms of damage, including internal microcracks, delamination, and macroscopic penetrating damage. These damages, especially internal damage, are highly concealed and complex, making them difficult to detect in a timely manner through routine visual inspections, seriously threatening the service safety and lifespan of the structure.

[0003] Currently, structural health monitoring technology can detect and preliminarily locate damage in composite materials. However, existing methods have significant limitations: First, most monitoring technologies stop at damage identification and fail to establish a quantitative inversion method from the sensed signal to the damage geometry and then to the macroscopic mechanical properties of the structure, such as stiffness and strength decay. Second, traditional mechanical models are inefficient in updating, making it difficult to meet the real-time evaluation needs of in-service structures. This results in monitoring data not being effectively converted into an accurate assessment of the structure's remaining load-bearing capacity, thus failing to provide direct and quantitative basis for operation and maintenance decisions, such as whether to continue service or whether maintenance is needed.

[0004] Therefore, there is an urgent need for an intelligent method that can integrate multi-source monitoring information, invert damage status in real time, and dynamically predict the remaining performance of structures, in order to achieve rapid and accurate evaluation of the strength and functional reliability of rocket composite structures. This would provide support for the strength and functional reliability evaluation of composite structures in long-life-cycle rocket launch vehicles.

[0005] Existing Solution: A digital twin and intelligent repair method, system, and equipment for aerospace composite material repair. The repair elements of the aerospace composite material repair entity, acquired in real time, are transmitted to the aerospace composite material repair digital twin. The digital twin establishes an initial visualized digital model of the repair entity based on its structural and damage parameters. The digital twin performs a design space search based on repair process parameters, predicts the key performance characteristics of different combinations of repair process parameters, and obtains the target repair process parameters required for the repaired entity to achieve the key performance characteristics. Based on the repair process parameters and the initial visualized digital model, the digital twin predicts, displays, and analyzes the full-field distribution information of the key performance characteristics of the repaired entity. The received target repair process parameters are then transmitted to the repair tool. The repair tool sets the repair process according to the target repair process parameters, implementing intelligent repair of the repaired entity.

[0006] Existing digital twin solutions, such as "digital twins for the repair of aerospace composite materials and intelligent emergency repair methods and systems," still have the following major shortcomings when applied to in-service health monitoring and performance evaluation of rocket composite material structures:

[0007] Damage sensing methods are often limited and lack comprehensive quantification of multi-mode damage. Existing technologies largely rely on single types of sensors (such as strain gauge networks or ultrasonic probe arrays) for condition monitoring. These methods may only sense surface strain distribution, making it difficult to detect internal damage (such as delamination or diffuse cracks); or while sensitive to internal damage, they may struggle to accurately invert the size and shape of macroscopic geometric damage (such as through-holes or macroscopic cracks). The inability to simultaneously and quantitatively obtain a complete damage profile, from macroscopic geometric defects to microscopic material property degradation, is the primary bottleneck for accurately evaluating the remaining performance of structures.

[0008] The "sensing-inversion" chain is broken, lacking a quantitative bridge from signal to physical parameters. Most existing health monitoring solutions stop at damage identification or localization, failing to establish a quantitative model and method for directly and automatically inverting key damage physical parameters (such as damage geometry, location, and material equivalent stiffness) from raw monitoring signals (such as image grayscale, ultrasonic velocity, and point strain). For example, when updating existing digital twins, manual measurement or pre-setting of damage parameters is often required, making it impossible to achieve automated, high-fidelity model mapping based on real-time monitoring data.

[0009] Model updates are disconnected from performance evaluation, resulting in insufficient real-time performance and forecasting capabilities. For example, the digital twin repair scheme cited in the background section focuses on the optimization and simulation of repair processes, rather than real-time status assessment of in-service structures. Even when some schemes attempt to update the model, the update process often relies on offline, manual methods, which are inefficient and cannot meet the urgent need for "real-time" or "near-real-time" evaluation of structural safety during rocket launch preparation or service. Furthermore, existing methods lack rapid structural mechanical response simulation and residual load-bearing capacity prediction modules that are linked to the updated model, preventing monitoring data from being directly converted into quantitative performance indicators (such as residual ultimate load and buckling critical load) to guide operational decisions. Summary of the Invention

[0010] The core technical problem this invention aims to solve is how to overcome the shortcomings of existing technologies, such as single monitoring methods, inaccurate damage parameter inversion, low model update efficiency, and disconnect from performance evaluation. This invention provides an integrated method that can fuse multi-source sensing data, automatically invert and quantify multi-mode damage in real time, and dynamically update the digital twin model to quickly predict the remaining structural load-bearing capacity. Specifically, this invention provides a real-time evaluation method for the remaining performance of rocket composite material structures based on multi-source sensing. Its core lies in constructing a closed-loop technical system of "physical sensing - damage inversion - model update - performance prediction".

[0011] The technical solution of this invention is: A digital twin method for real-time prediction of the bearing capacity of composite load-bearing tube structures based on field data inversion, the method comprising the following steps: Step 1: Construct a physical model of the composite load-bearing cylinder; Step 2: Perform static loading on the physical model of the composite load-bearing cylinder built in Step 1; Step 3: Apply initial damage to the physical model of the composite load-bearing cylinder under static loading in Step 2; use a 3D-DIC dual camera to acquire real-time strain field data of the physical model of the composite load-bearing cylinder; use a Lamb wave detection system to acquire guided wave data of the physical model of the composite load-bearing cylinder; and use strain gauges to acquire strain data of the measuring points of the physical model of the composite load-bearing cylinder. Step 4: Based on the real-time strain field data, guided wave data, and measuring point strain data obtained in Step 3, invert the damage evolution of the physical model of the composite load-bearing cylinder. Step 5: Evaluate the load-bearing capacity of the composite load-bearing tube based on the damage evolution inversion in Step 4, and complete the twin virtual evaluation of the composite load-bearing tube structure based on real-time field data inversion.

[0012] In step three, the initial damage types applied include penetrating damage and non-penetrating damage, with non-penetrating damage including delamination damage and diffuse crack damage.

[0013] The method for inverting the damage evolution of the physical model of the composite load-bearing cylinder in step four is as follows: Step 41: Based on the deep learning network, the location and shape of the penetrating damage are inverted from the real-time strain field data obtained in Step 3. Specifically, the real-time strain field data is input into the pre-trained deep neural network, and a pixel-level mask of the penetrating damage area is output. The output pixel-level mask is image processed and the scale is calibrated to obtain the geometric shape, size and position coordinates of the penetrating damage area. Step 42: Based on the Lamb wave empirical inversion model, the material stiffness of the composite load-bearing cylinder is derived from the guided wave data obtained in Step 3. Specifically, the measured wave velocity C and excitation frequency f of the guided wave are substituted into a pre-calibrated Lamb wave empirical inversion model to calculate the non-penetration damage factor. Based on the factor, the equivalent elastic modulus of the composite material in the non-penetration damage region is obtained. The calculation formula is as follows: E pre = E0× (1 - D), where E pre E0 is the equivalent elastic modulus of the composite material in the non-penetrating damage region, D is the initial elastic modulus of the composite material, and D is the non-penetrating damage factor. Step 4.3: Input the strain data from the measurement points into the trained neural network to locate the non-penetrating damage location using a data-driven approach.

[0014] In step 41, the loss function in the deep learning network is:

[0015] Among them, L cls For classification loss, L box For bounding box regression loss, L mask This is due to masking loss.

[0016] In step 41, the inversion result S of the deep learning network real for:

[0017]

[0018] Where true size represents the actual size in millimeters, pixel represents the number of corresponding image pixels, and S img This represents the area of ​​the penetrating damage.

[0019] In step 42, the Lamb wave empirical inversion model is as follows:

[0020] in, E predenoted as the equivalent elastic modulus of the composite material in the non-penetrating damage region, C is the measured wave velocity, and f is the excitation frequency.

[0021] In step 43, the neural network generator is trained using a composite loss function; the composite loss function for:

[0022] in, For counter-loss, To rebuild the losses, In order to perceive loss, This is a loss of style.

[0023] In step five, the method for evaluating the load-bearing capacity of the composite load-bearing cylinder based on the inverted damage evolution is as follows: Step 5.1: Establish a finite element mesh model based on the CAD model of the composite load-bearing cylinder as the initial digital twin; Step 5.2: Update the initial digital twin established in step 5.1 using real-time strain field data, guided wave data, and strain data at measurement points; Step 5.3: Perform finite element analysis based on the updated digital twin from Step 5.2 to obtain the remaining load-bearing capacity of the composite load-bearing cylinder, and evaluate its load-bearing performance after damage during service.

[0024] The load-bearing capacity includes ultimate load, buckling load, stress distribution, and risk of failure.

[0025] Beneficial effects What are the technical advantages of this application? 1. The functional leap from "repair process simulation" to "real-time evaluation of in-service status" solves the fundamental differences in model application scenarios.

[0026] Analysis of existing technology: The core objective of the cited patent solution is the optimization and simulation of the repair process. Its digital twin is mainly used to simulate the repair effects under different repair parameters before or during the repair process, in order to find the optimal repair solution. This is an offline design and analysis tool oriented towards the "repair process".

[0027] Advantages of this invention: This invention directly addresses the health monitoring and safety assessment of structures during their service life. Its core function is to perceive the current damage status of the structure in real time and dynamically predict its future load-bearing capacity. This solves the urgent need for real-time assessment of structural safety in critical stages such as launch preparation and mission execution for rockets and other spacecraft. This invention shifts the application scenario of digital twins from "post-event maintenance" to "pre-event early warning" and "in-event monitoring," realizing the transformation from simulated maintenance to ensuring operation.

[0028] 2. The model update mechanism has been innovated from "preset / manual input parameters" to "multi-source perception real-time automatic inversion", which solves the core bottlenecks of model fidelity and timeliness.

[0029] Analysis of existing technologies: In existing digital twin repair solutions, the initial state of the model (damage parameters) usually needs to be measured or preset manually, and model updates depend on preset repair process parameters. This approach cannot automatically acquire the intrinsic damage that evolves in the structure in real time in complex service environments. The model and the physical entity have poor synchronization, updates are delayed and dependent on manual intervention, and cannot meet real-time requirements.

[0030] Advantages of this invention: Automatic data acquisition: Through an integrated multi-source sensing system (DIC, Lamb wave, strain gauge), the original response data of the structure under load is automatically and synchronously collected without manual intervention.

[0031] Intelligent inversion process: Utilizing deep learning and physical mechanism models, raw, low-level sensory signals (images, wave velocity, point strain) are automatically and quantitatively inverted into high-order physical parameters (geometric dimensions, elastic modulus, crack location) that can be used for simulation. This establishes an automated bridge from monitoring data to updating model parameters.

[0032] Extremely high update efficiency: Through a programmatic interface, the inverted parameters are automatically read, written, and the finite element input file is modified, enabling one-click updates of the twin model. This process is fully automated and highly accurate, reducing model update time from the traditional manual hours or even days to minutes or even seconds, truly meeting the requirements of real-time or near-real-time evaluation.

[0033] 3. The improved perception capability, from "single or fuzzy damage characterization" to "precise quantification of multi-scale damage," solves the problem of accuracy in performance evaluation.

[0034] Analysis of existing technologies: Existing technologies focus on repair solutions, and the characterization of the damage itself may be relatively simple or vague. It is difficult to comprehensively and accurately quantify the multiple damage modes that coexist in composite materials (such as macroscopic through holes, internal invisible delamination, diffuse matrix microcracks, and macroscopic cracks), which leads to doubts about the accuracy of performance evaluation based on this model.

[0035] Advantages of this invention: Multi-mode collaborative sensing: The DIC system accurately captures macroscopic geometric damage on the surface (such as through holes and cracks) and the resulting full-field strain singularities; the Lamb wave system is extremely sensitive to internal material property degradation (stiffness reduction caused by delamination and diffuse cracks); sparse strain gauges provide accurate local measurements at key locations and assist in crack inversion. The integration of these three systems enables collaborative capture of multi-scale, multi-mode damage.

[0036] Quantitative parameter output: The inversion output does not provide a qualitative conclusion of "damage present / absent," but rather the actual size of the damage (millimeter level), precise coordinates, and specific values ​​of the material's equivalent modulus (GPa level). This quantitative output provides accurate input for subsequent high-fidelity finite element simulations, fundamentally ensuring the reliability and accuracy of the residual performance prediction results. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 The physical model loading and damage sensing system for digital twin systems loads the damaged composite load-bearing cylinder structure while simultaneously detecting and sensing the surface strain field and feeding it back to the twin model. Figure 3 This is a composite material internal damage inversion system based on lamb wave. It measures information such as wave velocity through a lamb wave generator and receiver and inverts the residual properties of the material using empirical formulas. Figure 4 This diagram illustrates the process of updating the perceived damage geometry information to the twin model, where the nodes and meshes in the twin model are eliminated using the perceived damage geometry information. Figure 5 This is a schematic diagram of the real-time process of the twin virtual evaluation method for composite load-bearing tube structures based on real-time field data inversion disclosed in this invention. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0039] The following combination Figure 1 The overall flowchart shown provides a detailed explanation of each step.

[0040] like Figure 1 As shown, the present invention discloses a digital twin method for real-time prediction of the bearing capacity of composite load-bearing tube structures based on field data inversion. In this embodiment, it is specifically implemented through the following steps: Step 1: Physical Entity Construction and Damage Introduction. A scaled-down rocket composite material structure, specifically a composite material cylinder, is used as the physical entity object for building the digital twin system. The selected carbon fiber epoxy resin-based composite material cylinder is shown below. Figure 2The cylindrical body of the loading device has a height of 300 mm and a radius of 150 mm. The surface of the composite material cylinder needs to be prepared with speckle patterns to meet the surface texture requirements of optical measurements. The process includes: cleaning with ethanol, uniformly spraying a matte white primer, and creating high-contrast black random speckles using a template or brush. This process ensures the texture features required by digital image correlation algorithms. Simultaneously, a piezoelectric sensor array and sparse strain measurement points are arranged on or inside the cylinder to form the basis for subsequent collaborative sensing. Step 2: Damage introduction and multi-source sensing system deployment. The cylinder is installed on a WDW-100C universal testing machine and loaded with 40kN at a rate of 0.05kN / s using force control to simulate quasi-static service load. Step 2.1: Introduce initial damage to the physical entity. To verify the system's damage identification capability, after preloading, use a handheld cutter to introduce artificial square through-hole damage on the cylinder surface to simulate penetrating damage caused by impact, etc. Step 2.2: Setup of the Optical Full-Field Strain Sensing System: Erect a 3D-DIC dual-camera system and complete high-precision calibration: Use a high-precision calibration board to move in multiple poses within the measurement space, simultaneously capturing at least 20 images with both cameras. After importing into VIC-3D software, calculate the camera's internal parameters (focal length, principal point, lens distortion coefficient) and external parameters (relative position and attitude) using an optimization algorithm. This system requires an overall reprojection error of less than 0.05 pixels (calibration score < 0.05) to ensure strain measurement accuracy reaches the micro-strain (µε) level. Once calibration is set, the camera cannot be moved or adjusted.

[0041] Step 2.3: Layout of the Internal State Ultrasonic Sensing System: Activate the Lamb wave detection system, which consists of a function generator, power amplifier, oscilloscope, and piezoelectric element. The internal state detection system comprises an excitation end and a receiving end. The RIGOL DG952 function generator produces a 5-cycle sine wave modulated by a Hanning window. After amplification by an Aigtek ATA-2021B power amplifier, it drives the PZT-5H piezoelectric element attached to the cylinder to excite the Lamb wave. The PZT sensor at the other end receives the signal, which is then digitized and stored by a Tektronix TDS2024B oscilloscope.

[0042] Step 2.4: Sparse point strain sensing system setup: Arrange piezoelectric sensor arrays and sparse strain measuring points on the surface / inside of the structure, start data acquisition of resistance strain gauges arranged in key parts, the sparse point system collects local strain time history data, and can input it into a pre-trained deep learning framework (e.g., SG-damage framework containing LSTM, CGAN and FPN modules) to reconstruct the full-field strain field and assist in crack identification. Step 3: Multi-source collaborative sensing and data acquisition. Three sets of sensing systems deployed on the physical entity are used to collaboratively acquire the structure's response signals under load. Step 3.1: The 3D-DIC system synchronously acquires a sequence of images of the structural surface at a specific frame rate (e.g., 1Hz), and calculates the full-field displacement and strain field distribution of the structural surface using a digital image correlation algorithm. Preferably, the processing includes delineating the region of interest, setting a calculation subset, removing rigid body displacements, and calculating and smoothing the engineering strain field.

[0043] Optical full-field strain sensing, using a 3D-DIC system such as Figure 2 As shown, images of the structural surface were synchronously acquired at a specific frame rate, and the full-field displacement and strain distribution of the structural surface were calculated using a digital image correlation algorithm. The entire experimental process was recorded using a DIC device with a frame rate of 1Hz. The acquired sequence of images was then processed and analyzed using Vic-3D 7.0 software: 1) The region of interest (ROI) of the measured object in the image was selected; 2) The subset size for calculation was set. A subset size of approximately 0.01 was considered appropriate and acceptable; 3) An averaging transformation method was selected to remove rigid body motion, eliminating the interference of overall motion on the results. This helps to more accurately understand the data's changing trends and improve the accuracy of data processing and analysis; 4) The engineering strain was calculated. A filter size of 15 was set to appropriately smooth the data, preserving its characteristics while maintaining visibility, ultimately obtaining the full-field strain results. Step 3.2: The exciter in the Lamb wave detection system emits a sinusoidal signal modulated by a Hanning window, and the sensor receives the propagated signal. The propagation time difference is calculated by intercepting the direct wave packet. t, combined with the sensor spacing d, according to the formula C = d / t Calculate the Lamb wave velocity C.

[0044] Internal damage condition is sensed by ultrasound, using an array of piezoelectric sensors such as... Figure 3As shown, the exciter emits a sinusoidal signal modulated by a Hanning window, and the sensor receives the propagated Lamb wave signal, records the waveform, and calculates the signal propagation time difference to obtain wave speed information. The Lamb wave detection system consists of a RIGOL DG952 function generator, an Aigtek ATA-2021B power amplifier, a Tektronix TDS 2024B digital storage oscilloscope, a PZT-5H ceramic piezoelectric element, and a PC control terminal. After connecting the experimental equipment and completing self-test, a five-cycle sinusoidal function excitation waveform modulated by a Hanning window is applied to the excitation sensor using the PC terminal. The acquired time-domain waveform can then be observed on the oscilloscope. After the waveform stabilizes, the excitation waveform and the received waveform are simultaneously saved as CSV files. The saved received waveforms include wave packets that directly reach the receiver after passing through the detection area and wave packets that arrive at the receiver after boundary reflection. Here, it's necessary to remove interference from reflected waves, retaining only the wave packets that directly reach the receiver after passing through the detection area. Since the excitation and received waveforms are saved simultaneously, the time difference between them is the propagation time of the Lamb wave in the detection area. Combined with the distance between the excitation and receiver, the Lamb wave velocity along the detection path can be calculated. The wave velocity calculation formula is:

[0045] In the formula: d is the distance between the sensors. t is the propagation time of the Lamb wave.

[0046] Step 3.3: The sparse measurement points acquire strain time history data at local points. Further, these sparse measurements can be input into a pre-trained deep learning framework (e.g., a framework containing an LSTM preprocessing network, a Conditional Generative Adversarial Network (CGAN), and a Feature Pyramid Network (FPN)) to reconstruct the full-field strain field and assist in identifying damage such as cracks.

[0047] Sparse-point strain field sensing: By using traditional resistance strain gauges placed at key locations, strain time history data of local points are acquired at a high sampling frequency. This serves two purposes: firstly, it supplements, verifies, and calibrates the full-field information of the DIC (Diverterless Computational) system; secondly, it allows for the reconstruction of the full-field strain field using sparse-point strain data through data-driven methods. Crack sensing employs the SG-damage framework, a deep learning-based approach designed to identify structural damage from sparse sensor measurements. The original sparse sensor measurements are interpolated into a dense, fixed-length sequence. Then, the SG-damage model is used to handle the complex full-field strain reconstruction task, learning a robust sparse-to-dense representation. Finally, a crack identification model is used to extract crack damage information from the reconstructed field. The SparseGenDamage framework (SG-damage) consists of three main modules: an LSTM preprocessing network, a Conditional Generative Adversarial Network (CGAN), and a Lightweight Feature Pyramid Network (FPN). First, the LSTM processes different numbers of measurement points and converts them into a fixed-length sequence, performing initial high-resolution reconstruction of the sensor data. The core of this framework is the CGAN, composed of a generator (G) and an exciter (D). The generator enhances feature extraction and generates full-field response data by incorporating a multi-scale extrusion and excitation (SE) module and an aggregation context transition (AOT) module, while the FPN evaluates the authenticity of the generated data. Finally, the FPN extracts key features from the full-field response and improves damage identification through multi-scale feature fusion, ultimately outputting a key point heatmap representing the crack tip.

[0048] Step 4: Multi-mode quantitative damage inversion. Based on the multi-source data collected in Step 3, a method combining mechanism and data-driven approaches is used to perform qualitative and quantitative damage inversion. Step 4.1: Macroscopic Damage Geometry Inversion Based on Deep Learning Network: The full-field strain contour map is input into a pre-trained instance segmentation neural network (e.g., Mask R-CNN or a variant thereof). The network outputs a pixel-level mask of the damage region, and through image processing and scale calibration, the precise geometric shape, size, and location coordinates of the macroscopic damage (such as penetration damage, delamination) are inverted.

[0049] Macroscopic damage geometry inversion based on deep learning networks involves inputting the full-field strain cloud map obtained from the DIC system into a pre-trained instance segmentation neural network, such as Mask R-CNN. This network directly identifies and segments the precise geometric shape, size, and location of macroscopic damage, such as penetration damage and delamination damage, at the pixel level. By building the network model Res-Mask R-CNN and an image-based damage quantification algorithm, accurate monitoring and classification of various forms of damage in composite laminates are achieved, and damage areas are segmented to obtain damage information. The loss of Res-Mask R-CNN during training mainly consists of three parts, and the loss function is shown in the formula:

[0050] Among them, L cls Classification loss is used to guide the model in learning the classification features of different targets; L box Bounding Box Regression Loss is used to optimize the positional difference between the predicted bounding box and the ground truth bounding box, enabling the model to accurately locate the target object; L mask Mask loss is used to supervise the consistency between the mask generated by the model and the real mask. By integrating the pixels of the damaged closed contour in the binary image, the pixel-based damaged area and contour length are calculated. The distance and actual length of the image are calibrated, and the ratio of a unit pixel to the actual size is defined as P. By transforming the statistical location information, geometric contour information, and area information according to the ratio P, the true damage location and size of the structural component can be obtained. The specific formula is:

[0051] Where true size represents the actual dimensions in millimeters, and pixel represents the number of corresponding image pixels. Using a scaling factor P, the center point and area Simg of the image-based damage region can be converted into actual coordinates and the actual area S of the damage region. real The specific process is as follows:

[0052]

[0053] Using the above method, the actual location and size of damage in composite cylindrical shell structures can be accurately obtained from image data, enabling quantitative inversion of the damage. The system reads the .inp file of the non-destructive model, automatically deletes the units in the damaged area according to the damage spatial domain output by Mask R-CNN, and renumbers them to generate a new geometric model with defects.

[0054] Step 4.2: Material stiffness inversion based on physical mechanism model: Substitute the measured wave velocity C and excitation frequency f into a pre-calibrated empirical inversion model (e.g., E_pre = F(C, f) or calculate E_pre = E_0 ×(1 - D) through damage factor D) to invert the equivalent elastic modulus of the composite material in the detection path region, thereby quantifying stiffness degradation.

[0055] Material stiffness inversion based on a physical mechanism model: A definite physical relationship exists between the propagation velocity of Lamb waves and the elastic modulus of the propagation medium. Through systematic "mechanical experiment-guided wave testing" calibration experiments, an empirical inversion model for this specific composite material was constructed. The model formula is as follows:

[0056] Where Epredicted is the predicted modulus after damage, C is the measured wave velocity, and f is the excitation frequency. This model encapsulates the intrinsic relationship between material stiffness and guided wave propagation characteristics, or, alternatively, by first fitting an expression relating guided wave velocity and damage value: ×

[0057] The predicted modulus of the cylindrical structure after damage can be calculated based on the damage value:

[0058] By substituting the real-time detected wave velocity C and the known excitation frequency f into the model, the equivalent elastic modulus of the composite material along the detection path can be quickly inverted. This allows for the quantification of material stiffness degradation caused by delamination, dispersed microcracks, etc. In the .inp file, the system automatically locates the set of elements covered by the Lamb wave detection path and modifies its material modulus to the inverted E.

[0059] Step 4.3: Crack location inversion based on sparse measurement point data. Using the full-field strain field reconstructed by S3.3, key points at the crack tip are extracted through a crack identification model (e.g., a network trained by a generator loss function), and their coordinates are mapped to the real coordinate system to locate the crack position.

[0060] Sparse strain data from measurement points are input into a trained neural network to invert the strain field. Crack damage locations are then located using a data-driven approach. The network generator is trained using a composite loss function, and the total generator loss is defined as follows:

[0061] This function incorporates adversarial loss. Reconstruction losses Perceived loss and style loss Based on the trained network, the strain field containing the crack can be inverted using the strain values ​​of the coefficient measurement points. The coordinates of the key points at the crack tip can be extracted from this strain field, and the heat map coordinates can be mapped back to the real-world sensor coordinate system using a known scale to update the crack location in the twin model.

[0062] Step 5: Construction and dynamic updating of the digital twin model. A twin virtual model matching the physical model is established in the computer system, and the damage information inverted in Step 4 is transmitted to the twin model for updating. Step 5.1: Initial high-fidelity model construction. Based on the original CAD model of the structure and digital image / point cloud data, a high-fidelity finite element mesh model is established as the initial digital twin. Step 5.2: Automatically update the digital twin model using the damage parameters obtained from the S4 inversion through a programmed interface: Automated damage implantation and attribute update. The damage parameters obtained from the inversion in Step 3 are automatically read, written, and modified in the finite element input file (.inp) through a programmed interface, updating the digital twin model in real time, such as... Figure 4 As shown.

[0063] Geometric Update: Based on the damage geometry inverted in steps 4.1 and 4.3, elements and nodes within the damage region are automatically deleted from the finite element model, achieving accurate mapping of the macroscopic damage geometry. The Inp file of the undamaged model is read, and the locations of *Node and *ELEMENT are located programmatically. Node and element information from the Inp file are then read. All nodes are traversed, nodes located within the defect spatial domain are marked, elements associated with defect nodes are deleted, and the remaining elements are renumbered.

[0064] Attribute Update: Based on the equivalent elastic modulus obtained from step 4.2, the material properties of the corresponding element set are automatically modified in the finite element model to achieve a quantitative mapping of material stiffness degradation. According to the placement of the piezoelectric elements on the structure in the experiment, elements within the detection area are selected by coordinates in the Inp file and established as a new element set. The detection direction modulus of the elements within this set is modified to match the modulus obtained from the prediction model inversion. The modified Inp file is then saved.

[0065] Step Six: Using the updated digital twin model from Step Five, apply the expected service load conditions, perform rapid finite element calculations, and predict in real time the remaining load-bearing capacity (such as ultimate load, buckling load), stress distribution, and failure risk of the structure. Modify the ODB data extraction configuration file: Open the .cfg file in the software path, modify the path parameters under input and outputpath, including odb_path (ODB path), odb_name (ODB name), vtk_path (output VTK file path), and variables under Output variables, which are divided into scalars, vectors, and tensors. Click the settings button in the digital twin software and modify the Time interval value. This value represents the interval for extracting the ODB file. It is recommended to set this value according to the model's calculation speed; it should be slightly less than the interval between each ODB field output. In the digital twin main control software interface, click the "Start Prediction" button. The system will automatically submit the updated .inp file to the finite element solver for calculation and monitor the ODB specified in the above steps. When the model's field output results are updated, the specified variables can be redrawn in the interface. Click the settings button in the software and modify the Time interval. This value should be slightly smaller than the time interval between outputs in the finite element analysis to ensure that each result update is captured. If you need to modify the output variables, you can edit the configuration file app\odb\ODB2VTK+.cfg in the software path to specify the scalar, vector, or tensor to be extracted. After the calculation starts, the software interface will display the predicted results in real time. Users can select to view the contour plots of different variables (such as stress and strain) and their different components through the drop-down menu on the interface to intuitively observe the structural response and potential hazardous areas under load.

[0066] Step 7: Based on the forecast results from Step 6, generate load control recommendations to prevent further structural damage during actual service, completing the closed loop from perception to decision-making. The overall rocket tube structure digital twin flowchart is shown below. Figure 5 As shown.

[0067] Key points and areas to be protected in this application 1. Multi-source collaborative sensing and fusion architecture: The protection system integrates three sensing systems—optical full-field strain sensing (3D-DIC), internal state ultrasonic sensing (Lamb wave), and sparse point strain sensing (resistance strain gauge)—into a collaborative deployment and synchronous data acquisition solution, forming a comprehensive monitoring capability for composite material structures from surface to interior, from full field to local, and from geometric damage to material degradation.

[0068] 2. Quantitative inversion method for multimodal damage: Macroscopic damage geometry inversion based on deep learning: This method utilizes instance segmentation neural networks (such as Mask R-CNN) to automatically analyze the full-field strain cloud map obtained from DIC, directly outputs a pixel-level mask of the damage area, and inverts the actual geometric size and location of the damage by converting the scale.

[0069] Material stiffness inversion based on physical mechanism model: Protective measures the conversion of real-time wave velocity C and excitation frequency f detected by Lamb wave detection into the material's equivalent elastic modulus by using a pre-established empirical inversion model (E_pre = F(C, f)) to quantify stiffness degradation.

[0070] Crack location inversion based on sparse measurement points: This is a technical solution that utilizes deep learning frameworks (such as SG-Damage) to reconstruct the full-field strain field from sparse strain data and locate the crack tip.

[0071] 3. Automated Dynamic Update Mechanism for Digital Twin Models: The protection system automatically reads the damage parameters obtained from the inversion through a programmatic interface (script) and directly modifies the finite element model input file (e.g., .inp file) to achieve the following operations: Geometric Update: Based on the inverted damage geometry, automatically delete the corresponding elements and nodes in the finite element mesh to achieve precise implantation of damage morphology.

[0072] Property Update: Automatically modify the material properties of a specified set of elements in the finite element model based on the inverted material equivalent modulus.

[0073] 4. Real-time performance forecasting and decision-making closed loop based on the updated model: The protection utilizes the dynamically updated digital twin model to apply the expected service load for rapid finite element simulation and visualize the forecast results in real time (such as stress cloud map and failure risk), thereby generating load adjustment or maintenance decision suggestions, forming a complete technical closed loop of "sensing-inversion-update-forecasting-decision".

[0074] 5. Integrated system and equipment for implementing the above methods: protect the complete system including a multi-source collaborative sensing module, a data processing and inversion module, a digital twin model update module, and a performance prediction module, as well as electronic equipment and computer-readable storage media for performing the above methods.

[0075] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A digital twin method for real-time prediction of the bearing capacity of composite load-bearing tube structures based on field data inversion, characterized in that... The steps of this method include: Step 1: Construct a physical model of the composite load-bearing cylinder; Step 2: Perform static loading on the physical model of the composite load-bearing cylinder built in Step 1; Step 3: Apply initial damage to the physical model of the composite load-bearing cylinder under static loading in Step 2; use a 3D-DIC dual camera to acquire real-time strain field data of the physical model of the composite load-bearing cylinder; use a Lamb wave detection system to acquire guided wave data of the physical model of the composite load-bearing cylinder; and use strain gauges to acquire strain data of the measuring points of the physical model of the composite load-bearing cylinder. Step 4: Based on the real-time strain field data, guided wave data, and measuring point strain data obtained in Step 3, invert the damage evolution of the physical model of the composite load-bearing cylinder. Step 5: Evaluate the load-bearing capacity of the composite load-bearing tube based on the damage evolution inversion in Step 4, and complete the twin virtual evaluation of the composite load-bearing tube structure based on real-time field data inversion.

2. The digital twin method for real-time bearing capacity prediction of composite load-bearing tube structures based on field data inversion according to claim 1, characterized in that: In step three, the initial damage types applied include penetrating damage and non-penetrating damage, with non-penetrating damage including delamination damage and diffuse crack damage.

3. The digital twin method for real-time bearing capacity prediction of composite load-bearing tube structures based on field data inversion according to claim 1, characterized in that: The method for inverting the damage evolution of the physical model of the composite load-bearing cylinder in step four is as follows: Step 41: Based on the deep learning network, the location and shape of the penetrating damage are inverted from the real-time strain field data obtained in Step 3. Specifically, the real-time strain field data is input into the pre-trained deep neural network, and a pixel-level mask of the penetrating damage area is output. The output pixel-level mask is image processed and the scale is calibrated to obtain the geometric shape, size and position coordinates of the penetrating damage area. Step 42: Based on the Lamb wave empirical inversion model, the material stiffness of the composite load-bearing cylinder is derived from the guided wave data obtained in Step 3. Specifically, the measured wave velocity C and excitation frequency f of the guided wave are substituted into a pre-calibrated Lamb wave empirical inversion model to calculate the non-penetration damage factor. Based on the factor, the equivalent elastic modulus of the composite material in the non-penetration damage region is obtained. The calculation formula is as follows: E pre = E0× (1 - D), where E pre E0 is the equivalent elastic modulus of the composite material in the non-penetrating damage region, D is the initial elastic modulus of the composite material, and D is the non-penetrating damage factor. Step 4.3: Input the strain data from the measurement points into the trained neural network to locate the non-penetrating damage location using a data-driven approach.

4. The digital twin method for real-time bearing capacity prediction of composite load-bearing tube structures based on field data inversion according to claim 3, characterized in that: In step 41, the loss function in the deep learning network is: Among them, L cls For classification loss, L box For bounding box regression loss, L mask This is due to masking loss.

5. The digital twin method for real-time bearing capacity prediction of composite load-bearing tube structures based on field data inversion according to claim 3, characterized in that: In step 41, the inversion result S of the deep learning network real for: Where true size represents the actual size in millimeters, pixel represents the number of corresponding image pixels, and S img This represents the area of ​​the penetrating damage.

6. The digital twin method for real-time bearing capacity prediction of composite load-bearing tube structures based on field data inversion according to claim 3, characterized in that: In step 42, the Lamb wave empirical inversion model is as follows: in, E pre denoted as the equivalent elastic modulus of the composite material in the non-penetrating damage region, C is the measured wave velocity, and f is the excitation frequency.

7. The digital twin method for real-time bearing capacity prediction of composite load-bearing tube structures based on field data inversion according to claim 3, characterized in that: In step 43, the neural network generator is trained using a composite loss function; the composite loss function for: in, For counter-loss, To rebuild the losses, In order to perceive loss, This is a loss of style.

8. The digital twin method for real-time bearing capacity prediction of composite load-bearing tube structures based on field data inversion according to claim 1, characterized in that: In step five, the method for evaluating the load-bearing capacity of the composite load-bearing cylinder based on the inverted damage evolution is as follows: Step 5.1: Establish a finite element mesh model based on the CAD model of the composite load-bearing cylinder as the initial digital twin; Step 5.2: Update the initial digital twin established in step 5.1 using real-time strain field data, guided wave data, and strain data at measurement points; Step 5.3: Perform finite element analysis based on the updated digital twin from Step 5.2 to obtain the remaining load-bearing capacity of the composite load-bearing cylinder, and evaluate its load-bearing performance after damage during service.

9. A digital twin method for real-time bearing capacity prediction of composite load-bearing tube structures based on field data inversion as described in claim 8, characterized in that: The load-bearing capacity includes ultimate load, buckling load, stress distribution, and risk of failure.

Images