Tower crane few sample adaptive digital twin modeling, calibration and operation monitoring method

Abstract

The application discloses a tower crane few-sample adaptive digital twin modeling, calibration and operation monitoring method, and belongs to the technical field of engineering structure digital twin and intelligent monitoring. The method comprises the following steps: establishing a tower crane working condition stress response data set; constructing a mechanics law guided residual neural network training agent model; deploying a digital twin and completing field residual calibration by using field stress; obtaining a few samples after the tower body configuration is changed, calculating the tower body configuration residual; learning the configuration residual to establish an adaptive digital twin; and correlating the tower body configuration and the running condition in real time, calling corresponding prediction relationship for continuous calibration and updating. The application distinguishes the real and virtual deviations in the field and the response deviations caused by the configuration change. After the tower body section is assembled and changed, the agent model does not need to be retrained as a whole, and only a few samples are used to complete residual migration and adaptation, so that the reuse capability of the digital twin is improved, and the digital twin can be used for tower crane tower body structure stress prediction, safety evaluation and intelligent operation.

Description

Technical Field

[0001] This invention relates to the field of digital twins and intelligent monitoring of engineering structures, belonging to the technical field of cross-integration of structural mechanics and artificial intelligence, and in particular to a method for modeling, calibrating and monitoring the operation and maintenance of a tower crane using a few-sample adaptive digital twin. Background Technology

[0002] Tower cranes are core equipment in building construction and large-scale engineering projects, responsible for material handling. Their safe and reliable operation directly affects construction safety and the normal progress of the project. During operation, the tower crane's working radius, lifting load, and slewing angle are constantly changing. Accurately and promptly grasping the stress state of key parts of the tower is the foundation for structural safety assessment and operation and maintenance. Digital twin technology combines operational monitoring data with structural response prediction models to represent the stress state of the tower crane in digital space, providing a basis for structural safety assessment and intelligent operation and maintenance. Therefore, establishing a reliable and effective digital twin of the tower crane is crucial for structural condition monitoring and safety assessment.

[0003] Existing technologies for digital twin modeling of tower cranes and related load-bearing structures have been researched in areas such as parametric model construction, structural response prediction, and calibration of measured information. Some existing solutions utilize finite element analysis data to establish a proxy model between the working conditions and structural response, and use sensor data to monitor structural performance and update the model. However, during construction, the number and combination of tower segments can change according to usage requirements. Even if the lifting load, working radius, and slewing angle are within the original operating range, changes in structural stiffness distribution and load transfer relationships will still render the working condition-stress relationship established under the original configuration invalid.

[0004] The search revealed the following published patent documents related to this application:

[0005] CN116654779A discloses a tower crane digital twin monitoring method and system, which collects tower crane operating status data through sensors and establishes a digital twin parameterized initialization model based on a tower crane template library to realize the association and visualization display of operating status data and digital model; however, it is only applicable to a fixed tower body configuration. When the tower body segments change, all need to be remodeled and trained, and it does not have the ability to adapt to configuration residual transfer.

[0006] CN117034687A discloses a tower crane digital twin monitoring system, which establishes a proxy model based on a simulation dataset and receives operational data to complete status monitoring; however, the neural network architecture is not guided by mechanical laws and does not distinguish between two independent residual correction mechanisms: the virtual and real residuals on site and the residuals due to configuration changes. As a result, the model fails after the configuration changes.

[0007] CN119783453A discloses a digital twin testing method and device for large-scale load-bearing structures, which integrates multi-task transfer learning with simulation and field testing to iterate the model of a given structure. The object is a fixed structure, and it does not address the scenario of variable configuration brought about by the modular assembly of the tower crane body, and there is no hierarchical update mechanism of "original model retention + configuration residual learning".

[0008] Existing technologies establish proxy models for given structures. Even with optimization, they struggle to directly adapt to changes in the load-stress mapping relationship caused by configuration variations. This often necessitates re-implementing the entire process under new structural conditions, limiting the reuse efficiency and continued application capability of deployed twins. Furthermore, discrepancies exist between the idealized finite element method (FEM) model and actual connection, manufacturing, installation, and measurement conditions. The model's prediction accuracy of the FEM calculation results cannot directly represent its consistency with the field response. While some existing patents fuse simulation and measured information through multi-task transfer learning, their focus remains on calibrating the virtual-to-real deviation of the given structure and iterating the prediction model.

[0009] Therefore, how to effectively utilize existing models under limited new data conditions, establish a digital twin modeling method that balances simulation prediction and on-site response consistency and can adapt to changes in tower configuration, and quickly restore the effective prediction capability of deployed twins for stress in key parts of new configurations are key technical problems that urgently need to be solved in the field of tower crane digital twins and intelligent operation and maintenance. Summary of the Invention

[0010] This invention addresses the issues of discrepancies between tower crane stress prediction models and on-site responses, the difficulty in directly reusing the original configuration proxy model after changes in tower segment combinations, and the limited availability of target configuration data. It proposes a method for modeling, calibrating, and monitoring tower cranes using a few-sample adaptive digital twin.

[0011] This invention uses the original configuration proxy model as a common foundation, establishes the relationship between on-site residual calibration and configuration residual update, and associates it with the current configuration and operating condition input of the tower crane to form a digital twin that can output the structural response at key locations, maintaining the continuous prediction effectiveness after the deployment of the digital twin.

[0012] To achieve the above objectives, the present invention adopts the following technical solution:

[0013] A method for few-sample adaptive digital twin modeling, calibration, and operation and maintenance monitoring of tower cranes, comprising the following steps:

[0014] S1. Establish a structural working condition-stress response dataset: Obtain the segment type, quantity, and connection order of the tower crane tower body, and establish a finite element model; use the lifting load Q, working radius R, and slewing angle θ as working condition input parameters, and obtain the stress response at key locations of the tower body through finite element calculation to form a working condition-stress response dataset.

[0015] S2. Construct a residual neural network guided by mechanical laws and train a surrogate model: Construct mechanical extended features for the input working condition parameters and build a stress baseline; use the difference between the structural response and the stress baseline as the learning target of the neural network to train a surrogate model;

[0016] S3. Deploy the digital twin and implement calibration using on-site stress: Deploy the surrogate model as a digital twin to receive the stress prediction results output by the tower crane in real time; collect the measured stress data on site, calculate the on-site residual between the measured response and the surrogate model prediction, train the on-site residual calibration model, and obtain the structural response output after on-site calibration.

[0017] S4. Obtain a small number of samples after the tower structure is changed and calculate the tower configuration residual; when the number of tower segments or the assembly method changes to form a new configuration, obtain a small number of new configuration working condition-stress response samples and calculate the tower configuration residual between the new configuration response and the original proxy model prediction output.

[0018] S5. Learn the configuration residuals and establish an adaptive digital twin; train an additional update model based on the tower configuration residuals, keep the parameters of the original proxy model unchanged, and superimpose the configuration residual mapping relationship to obtain an adaptive digital twin that adapts to the new configuration;

[0019] S6. Implement digital twin calls and continuous calibration updates for the associated tower structure and operation status; continuously receive tower crane operating conditions and current tower configuration information, call the corresponding prediction relationship according to the current tower assembly configuration, and output the stress at key locations of the tower; and support continuous updates to the model with subsequent additions of monitoring data and configuration data.

[0020] Furthermore, in step S2:

[0021] The mechanical extension feature is: ,in It is a vertical force. , Here is the unit conversion factor for the suspended weight, and g is the acceleration due to gravity; the stress baseline expression is... B is the coefficient matrix obtained by least squares calculation using the training dataset; the learning objective of the neural network. , The stress response is simulated for the original configuration. This represents the input vector for the i-th operating condition in the original configuration simulation dataset. Indicates that the input is The stress baseline prediction results obtained at that time.

[0022] Furthermore, in step S2, the residual neural network is a physics-guided attention residual network PG-ARNet-VH; the network input is a 10-dimensional mechanical extended feature, which is mapped to the latent space and passes through multiple residual blocks in sequence. Attention weights are used to fuse shallow and deep latent features, and vertical stress output heads and horizontal stress output heads are set respectively. The network training is completed using the Smooth-L1 loss function and the AdamW optimizer.

[0023] Furthermore, in step S3, when the field-monitored stress is the stress increment relative to the zero-adjustment reference condition... At that time, the difference between the prediction results of the surrogate model under the current working condition and the reference working condition is calculated and converted into stress increments under the same reference condition. Then calculate the on-site residuals. , This represents the field monitoring stress of the i-th sample. This represents the working condition input for the i-th sample. This indicates that the input is At that time, the prediction results of the proxy model.

[0024] Furthermore, in step S3, the field residual calibration model employs ridge regression, recursive least squares, or a small-scale neural network to establish residual mapping relationships using limited field-measured samples; the output response is then calibrated. ; Represents the original proxy model prediction, Output for on-site residual calibration model.

[0025] Furthermore, in step S5, all parameters of the original proxy model are kept unchanged, and only the configuration residual mapping relationship is learned. New configuration output response .

[0026] Furthermore, the method for obtaining a small number of new configuration working condition-stress response samples in step S4 includes: establishing a finite element simulation model for the new configuration to obtain samples through simulation calculation, or collecting on-site sensor measurements of the new configuration tower crane entity.

[0027] Further, in step S6, the original proxy model, the on-site residual calibration model, and the configuration residual update model corresponding to different tower assembly configurations are saved; during runtime, the tower assembly configuration record is read, and the corresponding model relationship is matched to output stress prediction.

[0028] This invention also discloses a tower crane few-sample adaptive digital twin monitoring system, including a memory, a processor, and a computer program stored in the memory and executable on the processor; the computer program performs all the steps of the above method.

[0029] The present invention also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method steps.

[0030] Beneficial effects;

[0031] Compared with existing digital twin methods that establish surrogate models for given tower configurations and perform experimental calibration, this invention has the following advantages:

[0032] 1) This invention proposes a residual neural network guided by mechanical laws to construct a mechanical baseline. The neural network only learns the residuals outside the baseline, integrating prior mechanical knowledge with data-driven approaches to improve the prediction accuracy of the surrogate model.

[0033] 2) Distinguish between two types of deviations: the field residual between finite element simulation and the actual site, and the configuration residual caused by changes in the tower assembly configuration; establish two sets of correction models respectively;

[0034] 3) When the assembly of tower segments changes, it is not necessary to retrain the original proxy model as a whole. Instead, the configuration residuals are learned using only a small number of samples, and adaptive adaptation is achieved through residual transfer, which greatly reduces the number of samples required for the new configuration.

[0035] 4) Save the corresponding model correction relationship for different tower configurations; automatically call the corresponding prediction relationship when reading tower assembly information during operation; realize continuous iterative updates of digital twin; improve the reusability of digital twin model, and provide reliable tower stress output for tower crane tower structure status monitoring, safety assessment and intelligent operation and maintenance. Attached Figure Description

[0036] Figure 1 is a flowchart of an adaptive digital twin modeling method for a tower crane according to an embodiment of the present invention, showing the overall process in the embodiment;

[0037] Figure 2 is a comparison chart of the test performance of the proxy model and other methods in the embodiments of the present invention, showing the bar error of the RMSE test of different models;

[0038] Figure 3 is a diagram of the error of the reserved working condition for the on-site incremental residual calibration in an embodiment of the present invention, showing the RMSE curve of the number of on-site calibration working conditions minus the reserved working condition.

[0039] Figure 4 is a configuration update error curve under different target sample numbers in the embodiments of the present invention, showing the target configuration adaptation condition number - test RMSE curve;

[0040] Figure 5 shows the residual migration before and after the update under the new configuration in this embodiment of the invention. ) and directly using the original proxy model The prediction results comparison chart shows the scatter plot of the predicted and actual values ​​of the new configuration sample before and after residual migration. Detailed Implementation

[0041] This invention takes data processing, neural networks, residual calibration, and configuration residual transfer as its core technical features, and effectively implements them by using a tower crane as the application object in the embodiments.

[0042] This invention constructs an adaptive digital twin for predicting the structural response of tower cranes. It comprehensively considers the discrepancies between the finite element model and the actual on-site structural response, changes in the working condition-stress mapping caused by variations in the number and combination of tower segments, and the limited availability of data under new configurations. This achieves stress response prediction for the original configuration, on-site residual calibration, and few-sample prediction adaptation after configuration changes. This method maintains the predictive effectiveness of the digital twin under on-site deployment and after tower configuration changes while preserving existing surrogate models, improving model reusability and cross-configuration predictive adaptability. It can provide technical support for tower crane structural condition monitoring, safety assessment, and intelligent operation and maintenance.

[0043] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are merely for explaining the present invention and do not limit the scope of protection of the present invention.

[0044] The tower body of a tower crane is assembled from modular segments, including foundation segments, standard segments, and reinforcing segments; as construction requirements change, the assembly configuration of the tower body changes accordingly.

[0045] The configuration of the tower crane refers to the assembly method of the tower structure. In the formula, the superscript or subscript S represents the original configuration simulation, F represents the original configuration measured on-site, and T represents the new configuration after modification; bold symbols represent vectors or matrices. The surrogate model is used to quickly obtain the stress response output through the input of working condition parameters. After being deployed on the digital twin platform, it serves as a digital twin describing the stress response of the structure.

[0046] like Figure 1 As shown, an embodiment of the present invention provides a method comprising the following steps:

[0047] (1) Establish a structural load-stress response dataset;

[0048] The segment types, number, and connection sequence of the tower crane's tower body are obtained. A finite element model is established by combining operating parameters, geometric dimensions, material parameters, and constraint relationships. The operating parameters refer to the crane's working load Q, luffing radius R, and slewing angle θ. Based on these parameters, boundary conditions are applied to the finite element model to calculate the structural response. The structural response is the vertical, lateral, or Von Mises stress at one or more key locations. Thus, using the operating parameters as input and the structural response as output, a structural operating-stress response dataset is constructed.

[0049] In one implementation scenario, based on the rated lifting performance of the tower crane during its design, the allowable combination range of lifting weight and working radius is determined. Within this range, uniform sampling, Latin hypercube sampling, or other space-filling sampling methods are used to obtain working condition samples. For the obtained working condition samples, the stress response at a predetermined position is calculated using a finite element model corresponding to the current tower configuration, forming a dataset with a one-to-one correspondence between input working conditions and output stress.

[0050] Based on the tower crane lifting performance curve, a feasible operating condition range is defined. Operating condition samples covering this range are selected, and corresponding boundary loads are applied one by one to extract the stress response, forming the original configuration dataset. Wherein, S refers to the original configuration simulation data. This represents the i-th operating condition input vector of the original configuration simulation data. This represents the simulated stress response of the original configuration corresponding to the input vector of the stated operating condition. It is the total number of samples. Indicates the sample number, from 1 to .

[0051] (2) Construct a residual neural network guided by mechanical laws and train a surrogate model;

[0052] Based on the dataset established in step (1), a residual neural network guided by mechanical laws is constructed and a surrogate model is trained.

[0053] In practice, feature transformations are constructed on the input vector to obtain mechanical extended features.

[0054] The mechanical extension feature is: Among them, vertical force and bending moments along the X and Z axes Together, they constitute the quantities describing the axial force and overturning moment corresponding to the suspended load. . κ Q This represents the unit conversion factor for the suspended weight, and g represents the acceleration due to gravity. The mechanical extension features are used to organize the input to the surrogate model, guiding the influence of working condition parameters on the structural response from a mechanical perspective. A stress baseline is then established based on this. B is the coefficient matrix obtained by least squares calculation of the training data. For the i-th sample, the original configuration structure response is... The difference from the stress baseline is used as the learning target of the neural network. , ,in, The stress response is simulated for the original configuration. This represents the input vector for the i-th operating condition in the original configuration simulation dataset. Indicates that the input is The stress baseline prediction results are obtained at that time. This completes the training of the neural network, resulting in the final surrogate model. The surrogate model is then compared with other methods, for example... Figure 2 The results show that its prediction accuracy is higher.

[0055] (3) Deploy the digital twin and perform calibration using on-site stress;

[0056] The original configuration proxy model trained in step (2) is deployed to the tower crane digital twin system to receive the tower crane's operating conditions in real time and output the stress prediction results at the predetermined positions through the proxy model. To improve the consistency between the surrogate model's predictions and the actual structural response, further field stress data corresponding to the model's output locations were collected. And the residual between the measured response and the surrogate model predicted response. Establish an on-site residual calibration model. When the on-site monitored stress is the stress increment relative to the zero-adjustment reference condition, subtract the prediction results of the surrogate model under the current condition and the reference condition, convert it to the stress increment under the same reference condition, and then calculate the on-site residual. , residual For the i-th sample, The residual calibration model is based on the collected data and obtains the mapping relationship through ridge regression, recursive least squares, small neural networks, or other regression models suitable for limited data. , making This demonstrates that limited field data can effectively improve the consistency between digital models and actual tower cranes.

[0057] (4) Obtain a small number of samples after the structural change and calculate the tower body configuration residuals;

[0058] When the number or combination of tower segments changes, a new configuration is determined based on tower configuration records, segment information in the digital model, or the construction management system. The new configuration refers to a tower segment composition state different from when the original configuration proxy model was established. A corresponding structural numerical model is established or updated for the target configuration, and the load-stress response under a small number of new configurations is obtained. Samples. Samples can be obtained by establishing a finite element model based on information related to the new configuration, applying boundary conditions to the finite element model based on the load parameters, and calculating the structural response; alternatively, field stress data corresponding to the model output location under the new configuration can be collected. After obtaining a small number of target configuration load-stress response samples, for the i-th sample, the residual between the new configuration response and the twin prediction is calculated. The residual structure of the tower body was obtained. .

[0059] (5) Learn configurational residuals and build adaptive digital twins;

[0060] Based on the original configuration surrogate model, and using the small number of samples obtained in step (4) and the calculated configuration residuals... An update mechanism for the new configuration is established. This update mechanism is used to obtain the mapping relationship. , making In one implementation, the parameters of the original surrogate model are updated based on a small number of samples to fine-tune the model. In another implementation, the parameters of the original surrogate model are kept unchanged, and the difference between the target configuration response and the prediction result of the original surrogate model is used as the learning objective to establish an additional update model for training to obtain the mapping relationship. The digital twin outputs the updated structural response. .

[0061] During implementation, model fine-tuning and residual transfer were carried out separately. The configuration update error curves under different target sample numbers are shown in the figure below. Figure 4 As shown in the figure. Compared to directly using the original model and model fine-tuning, residual transfer achieved a lower average prediction error. The comparison of model prediction performance before and after the update under the new configuration is shown in the figure below. Figure 5 As shown.

[0062] (6) Implement digital twin calls and continuous calibration and updates by associating the tower structure and operation status at each moment;

[0063] During the operation of the digital twin, the corresponding prediction relationship is invoked based on the current configuration to output the stress response at key locations of the tower body in real time. The prediction relationship refers to the calibrated structural response when the configuration remains unchanged. The updated structural response is obtained by re-executing steps (4) and (5) each time the configuration changes. This avoids the need to redeploy the entire prediction system after configuration changes.

[0064] In practice, the original surrogate model and its field residual calibration relationship are preserved for the original test configuration. When segments are added to the tower body based on the original configuration, the original surrogate model is used as the basis for shared prediction, and the updated structural response is obtained through residual transfer. This allows the digital twin of the tower crane to be continuously updated as on-site data accumulates and the tower structure changes.

[0065] Through the above implementation methods, the present invention uses the original configuration mechanical guided proxy model as a reusable prediction basis, and expresses the finite element-site deviation and the working condition-stress mapping deviation caused by configuration change as additional residuals, and performs corresponding model calls and incremental updates according to the current tower configuration, thereby realizing the continuous adaptation of the tower crane digital twin after on-site deployment and tower reassembly.

[0066] In practice, the original tower structure will consist of one foundation section, two reinforcing sections, and three standard sections. A finite element model of the original tower structure will be established using ANSYS. The tower material has an elastic modulus of 206 GPa, a Poisson's ratio of 0.3, and a density of 7850 kg / m³. The tower arm, counterweight arm, slewing superstructure, counterweight, and suspended weight are applied to the top of the tower through equivalent quasi-static forces and overturning moments.

[0067] S1. Establish a load condition-stress response dataset;

[0068] Within the tower crane's permissible lifting range, 400 sets of original configuration working condition samples were extracted using Latin hypercube sampling; the working radius was 21.41~80m, and the lifting load was sampled based on a load utilization rate of 0.10-0.95 times the rated lifting capacity; the slewing angle was 0°-360°; the output consisted of vertical and lateral stresses at preset positions of the main chords on sides A and C of the tower body reinforcement section and foundation section, forming a dataset. Input vector Output vector .

[0069] S2 constructs a residual neural network guided by mechanical laws to train a surrogate model;

[0070] A 10-dimensional mechanical extended feature was constructed to establish a stress baseline; the neural network adopted was PG-ARNet-VH, a physics-guided attention residual network; the 10-dimensional mechanical features were mapped to a 48-dimensional latent space; three residual blocks were used, each including a linear layer, GELU activation function, Dropout, and LayerNorm; an attention mechanism was used to fuse deep and shallow latent features; vertical stress output heads and horizontal stress output heads were set respectively; Smooth-L1 loss and AdamW optimizer were used for training; the RMSE on the test set reached 0.246±0.048MPa.

[0071] S3 deploys digital twins to conduct on-site residual calibration;

[0072] Strain sensors were installed on the main chords of the tower crane's foundation section, reinforced section (side A and side C), with a sampling frequency of 50Hz. Field stress data was collected; the residual between the measured stress and the surrogate model prediction was calculated; and a field residual calibration model was established using ridge regression. After training with 13 sets of operating condition samples, the measured RMSE decreased from 5.085±0.949MPa to 2.616±0.940MPa. The calibration response was output. .

[0073] The S4 tower segment assembly was modified to obtain a new configuration; 60 sets of working condition samples of the new configuration were extracted by Latin hypercube, and the stress was obtained by finite element simulation to calculate the tower configuration residuals.

[0074] S5 learns configuration residuals to build an adaptive digital twin;

[0075] Train the configuration residual mapping model while keeping all parameters of the original proxy model unchanged. New configuration output Compared to directly using the original model and fine-tuning the model, the residual transfer scheme significantly reduces test errors.

[0076] S6 continuously receives tower crane operating conditions and tower assembly configuration information; calls the corresponding prediction output based on the current tower configuration; and continuously updates the residual mapping with subsequent new field data and configuration data to complete continuous calibration and update.

[0077] This embodiment achieves adaptive adaptation of the tower crane digital twin with fewer samples under the condition of changes in the tower assembly configuration through hierarchical residual correction, outputs the stress at key positions of the tower body, and serves the tower crane safety assessment and intelligent operation and maintenance.

[0078] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for modeling, calibrating, and monitoring the operation and maintenance of a tower crane using a few-sample adaptive digital twin, characterized in that, Includes the following steps: S1. Establish a structural working condition-stress response dataset: Obtain the segment type, quantity, and connection order of the tower crane tower body, and establish a finite element model; use the lifting load Q, working radius R, and slewing angle θ as working condition input parameters, and obtain the stress response at key locations of the tower body through finite element calculation to form a working condition-stress response dataset. S2. Construct a residual neural network guided by mechanical laws and train a surrogate model: Construct mechanical extended features for the input working condition parameters and build a stress baseline; use the difference between the structural response and the stress baseline as the learning target of the neural network to train a surrogate model; S3. Deploy the digital twin and perform calibration using on-site stress: Deploy the proxy model as a digital twin to receive the stress prediction results output by the tower crane in real time. Collect on-site measured stress data, calculate the on-site residual between the measured response and the surrogate model prediction, train the on-site residual calibration model, and obtain the on-site calibrated structural response output. S4. Obtain a small number of samples after the tower structure is changed and calculate the tower configuration residual; when the number of tower segments or the assembly method changes to form a new configuration, obtain a small number of new configuration working condition-stress response samples and calculate the tower configuration residual between the new configuration response and the original proxy model prediction output. S5. Learn the configuration residuals and establish an adaptive digital twin; train an additional update model based on the tower configuration residuals, keep the parameters of the original proxy model unchanged, and superimpose the configuration residual mapping relationship to obtain an adaptive digital twin that adapts to the new configuration; S6. Implement digital twin calls and continuous calibration updates for the associated tower structure and operation status; continuously receive tower crane operating conditions and current tower configuration information, call the corresponding prediction relationship according to the current tower assembly configuration, and output the stress at key locations of the tower; and support continuous updates to the model with subsequent additions of monitoring data and configuration data.

2. The method according to claim 1, characterized in that, In step S2: The mechanical extension feature is: ,in It is a vertical force. , Here is the unit conversion factor for the suspended weight, and g is the acceleration due to gravity; the stress baseline expression is... B is the coefficient matrix obtained by least squares calculation using the training dataset; the learning objective of the neural network. , The stress response is simulated for the original configuration. This represents the input vector for the i-th operating condition in the original configuration simulation dataset. Indicates that the input is The stress baseline prediction results obtained at that time.

3. The method according to claim 2, characterized in that, In step S2, the residual neural network is a physics-guided attention residual network PG-ARNet-VH; the network input is a 10-dimensional mechanical extended feature, which is mapped to the latent space and passes through multiple residual blocks in sequence. Attention weights are used to fuse shallow and deep latent features, and vertical stress output heads and horizontal stress output heads are set respectively. The Smooth-L1 loss function and AdamW optimizer are used to complete the network training.

4. The method according to claim 1, characterized in that... In step S3, when the field-monitored stress is the stress increment relative to the zero-adjustment reference condition... At that time, the difference between the prediction results of the surrogate model under the current working condition and the reference working condition is calculated and converted into stress increments under the same reference condition. Then calculate the on-site residuals. , This represents the field monitoring stress of the i-th sample. This represents the working condition input for the i-th sample. This indicates that the input is At that time, the prediction results of the proxy model.

5. The method according to claim 1, characterized in that... In step S3, the field residual calibration model employs ridge regression, recursive least squares, or a small-scale neural network to establish residual mapping relationships using limited field measurement samples; the output response is then calibrated. ; Represents the original proxy model prediction, Output for on-site residual calibration model.

6. The method according to claim 1, characterized in that... In step S5, all parameters of the original proxy model remain unchanged, and only the configuration residual mapping relationship is learned. New configuration output response .

7. The method according to claim 1, characterized in that... The small number of new configuration working condition-stress response samples mentioned in step S4 include: samples obtained by simulation calculation of the new configuration using a finite element simulation model, or samples measured by sensors on the actual site of the new configuration tower crane.

8. The method according to any one of claims 1-7, characterized in that... Step S6: Save the original proxy model, on-site residual calibration model, and configuration residual update model corresponding to different tower assembly configurations; during runtime, read the tower assembly configuration record, match the corresponding model relationship, and output stress prediction.

9. A tower crane few-sample adaptive digital twin monitoring system, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor; the computer program performs the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1-8.

Images