A cross-section drift buoy isomer digital twin model construction method

Abstract

The present application relates to a kind of profile drift buoy isomerization digital twin model construction method, belong to model construction technical field, including the mechanical motion model of construction profile drift buoy, obtain actual operation data, according to mechanical motion model definition error function, according to error function the error value of mechanical motion model and actual operation data;The mean and standard deviation of error value are calculated by neural network, the error value is corrected according to mean, and isomerization digital twin model is obtained;According to isomerization digital twin model, the predicted running state is obtained, and the model error result is calculated according to the predicted running state and actual operation data, according to standard deviation definition error threshold, according to error threshold and model error result, loss function is calculated, according to loss function, the parameter of isomerization digital twin model is updated, the deviation quantification mechanism between physical model and real system can be established, high-precision modeling and prediction to the running state of profile drift buoy are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of model construction, in particular to a profile drift buoy heterogeneous digital twin model construction method. BACKGROUND

[0002] In the field of ocean observation, the profile drift buoy has been in a complex and changeable deep sea environment for a long time, and its motion state, structural health and environmental adaptability directly affect the accuracy of data collection. The traditional monitoring method relies on limited sensor data and is difficult to fully evaluate the real-time state of the buoy. The digital twin technology can dynamically simulate the mechanical behavior of the buoy, predict potential failures, and provide visual support for maintenance decisions, thereby significantly improving the reliability and efficiency of ocean observation.

[0003] At present, the construction of heterogeneous digital twin model mainly establishes an analytical model based on the mechanical equation of the buoy, which has strong interpretability and robustness, and is suitable for data-scarce scenarios. Or use deep learning to learn the system behavior from historical data, adapt to nonlinear and coupled relationships, and have high modeling efficiency.

[0004] However, the physical modeling method relies on accurate mechanism knowledge, but the complex working conditions of the buoy in the deep sea environment cause the model parameters to drift easily, and it is difficult to update in real time, and the adaptability is insufficient. The pure data-driven method lacks physical interpretability, requires a large amount of training data and high quality, has poor generalization ability when data is insufficient, and has high computational complexity, making it difficult to deploy in lightweight edge devices of the buoy. SUMMARY

[0005] In view of the problems in the related art, the present application provides a profile drift buoy heterogeneous digital twin model construction method to solve the technical problems that the physical modeling method in the prior art is difficult to adapt to complex working condition changes due to its dependence on accurate mechanism, and the data-driven method is difficult to deploy in lightweight devices due to its poor interpretability, strong data dependence and high computational complexity.

[0006] The present application provides a profile drift buoy heterogeneous digital twin model construction method, comprising the following steps:

[0007] Error calculation step: constructing a mechanical motion model of the profile drift buoy, obtaining actual running data of the profile drift buoy based on the mechanical motion model, defining an error function according to the mechanical motion model, and calculating an error value of the mechanical motion model and the actual running data according to the error function;

[0008] Model construction step: calculating the mean and standard deviation of the error value through a neural network, performing error correction on the error value according to the mean, and obtaining a heterogeneous digital twin model according to the error-corrected mechanical motion model.

[0009] The parameter updating step: obtaining the predicted running state of the profile drift buoy according to the heterogeneous digital twin model, calculating the model error result according to the predicted running state and the actual running data, defining the error threshold according to the standard deviation, calculating the loss function according to the error threshold and the model error result, updating the parameters of the heterogeneous digital twin model according to the loss function, and the heterogeneous digital twin model is used to generate decisions according to the profile drift buoy.

[0010] By constructing a mechanical motion model and defining an error function, the error value with the actual running data is calculated, a deviation quantification mechanism between the physical model and the real system is established, and accurate data basis is provided for subsequent error correction; then the mean and standard deviation of the error are calculated by the neural network, and the error is corrected, the complex nonlinear error which is difficult for the physical model to accurately describe is compensated, finally the heterogeneous digital twin model which fuses physical mechanism and data-driven is formed, and high-precision modeling and prediction of the running state of the profile drift buoy are realized.

[0011] In some embodiments of the application, the parameter updating step specifically includes:

[0012] The neural network is used to obtain the prediction error between the error value and the mean value, the maximum mean difference between the prediction error and the error value is calculated according to the loss function, and the parameters of the heterogeneous digital twin model are updated according to the maximum mean difference.

[0013] The model parameters are updated by calculating the prediction error by the neural network and combining the maximum mean difference, the distribution difference between the prediction error and the actual error is quantified, the performance degradation in the model updating process is avoided by forcibly maintaining the distribution consistency, so that the parameter optimization process not only considers the numerical value of the error but also pays attention to the statistical characteristics of the error, thereby improving the scientificity of the parameter updating and the stability of the model updating, and ensuring the reliability of the heterogeneous digital twin model in long-term operation.

[0014] In some embodiments of the application, the method further includes:

[0015] The neural network updating step: if there is a preset deviation between the updated parameters of the heterogeneous digital twin model and the parameters of the mechanical motion model, the parameters of the mechanical motion model are updated, the error value, the mean value and the standard deviation of the error value are recalculated, and the neural network is updated.

[0016] By triggering the retraining process when the model parameters have a preset deviation from the physical model parameters, the model drift problem caused by environmental changes or equipment aging is solved by periodically refreshing the entire modeling process to track system dynamics, thereby realizing the full life cycle adaptability of the heterogeneous digital twin model, prolonging the effective use period of the model, and ensuring the consistency of simulation accuracy and physical entities.

[0017] In some embodiments of the application, the calculation model of the loss function is:

[0018]

[0019] wherein, is a loss function; is a model error result; is an error threshold.

[0020] By defining the calculation model of the loss function, the loss calculation process is standardized, ensuring that the optimization behavior under large and small error conditions meets the design expectations, providing a clear and differentiable optimization target for parameter updating, thereby ensuring the convergence and stability of the entire adaptive system.

[0021] In some embodiments of the application, the dynamics of the mechanical motion model is represented as:

[0022]

[0023] wherein, is the mass of the sectioned drift buoy; is the density of water at a depth of is the volume of the glass sphere shell; is the compressibility of the glass sphere shell; is the volume of the outer oil bladder; is the compressibility of the hydraulic oil; is an unmodeled term; is the compressibility of the unmodeled term; is the drag coefficient of the sectioned drift buoy; is the movement speed of the sectioned drift buoy.

[0024] By defining the dynamics equation of the mechanical motion model, an accurate physical description of the buoy movement is established based on the principles of fluid mechanics and material mechanics, providing a reliable theoretical basis for the heterogeneous digital twin model.

[0025] In some embodiments of the application, the expression of the error function is:

[0026]

[0027] wherein,​ is an error function.

[0028] By defining the mathematical expression of the error function, the difference between the mechanical motion model and the actual observation value is formalized into a calculable function relationship, which provides a clear learning goal for the neural network, while retaining the physical interpretability of each parameter, so that error analysis can be traced back to specific physical factors, providing directional guidance for system optimization. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, specific embodiments of the present application will be described in detail below with reference to the drawings. For those skilled in the art, other drawings can also be obtained without creative labor on the premise that these drawings do not deviate from the scope of the present application.

[0030] Figure 1 A flowchart of a cross-section drift float buoy heterogeneous digital twin model construction method provided by the embodiment of the present application;

[0031] Figure 2 A flowchart of another cross-section drift float buoy heterogeneous digital twin model construction method provided by the embodiment of the present application;

[0032] Figure 3 A flowchart of another cross-section drift float buoy heterogeneous digital twin model construction method provided by the embodiment of the present application. DETAILED DESCRIPTION

[0033] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be described and explained below with reference to the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application, and are not intended to limit the present application. Based on the embodiments provided by the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0034] It should be noted that the terms used herein are only intended to describe specific embodiments, and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should be understood that the terms "comprise" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0035] Digital Twin technology is a modeling and simulation method that combines virtual and real worlds. By constructing a virtual model consistent with the physical entity, real-time perception, predictive analysis, health management and behavior decision of the equipment running state are realized, and good visualization, predictability and interpretability are provided, which provides a new solution for the state monitoring, fault diagnosis and intelligent decision of marine equipment.

[0036] Digital Twin can not only realize real-time perception, predictive analysis and health assessment of the system by constructing a virtual model highly consistent with the running state of the physical equipment, but also has good visualization, interpretability and scalability, which provides a new solution for the state monitoring, fault diagnosis and autonomous decision of complex systems such as marine equipment.

[0037] At present, the construction methods of digital twin mainly include two categories: physical modeling based method and artificial intelligence based method. The physical modeling based method relies on prior knowledge and mechanical equations of the equipment, which can provide high interpretability and strong robustness, and is suitable for data scarce and complex environment scenarios, but the modeling process is tedious and difficult to adapt to system working condition changes;

[0038] The artificial intelligence based method automatically learns the system behavior through deep learning and other data driven methods, which has high modeling efficiency and strong adaptability, especially in handling nonlinear and coupled systems, but also has poor interpretability, strong dependence on data, high computational complexity and difficulty in deploying lightweight devices.

[0039] In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0040] The technical solutions of the present application will be described in detail below in conjunction with specific embodiments and the accompanying drawings.

[0041] As shown in Figure 1 The present application provides a cross-section drift buoy heterogeneous digital twin model construction method, comprising:

[0042] Error calculation step S1: constructing a mechanical motion model of the cross-section drift buoy, obtaining actual running data of the cross-section drift buoy based on the mechanical motion model, defining an error function according to the mechanical motion model, and calculating the error value of the mechanical motion model and the actual running data according to the error function;

[0043] In some embodiments, the dynamics of the mechanical motion model is represented as:

[0044]

[0045] wherein, is the mass of the cross-section drift buoy; is the depth density of water under the glass sphere shell; volume of the glass sphere shell; compressibility of the glass sphere shell; volume of the outer oil bladder; compressibility of the hydraulic oil; unmodelable item; compressibility of the unmodelable item; drag coefficient of the cross-section drift buoy; moving speed of the cross-section drift buoy;

[0046] Specifically, other volume, the other volume being an unmeasurable volume in the process of balancing the mass and buoyancy of the cross-section drift buoy in the laboratory, i.e., the unmodelable item, other than the volume of the glass sphere shell and the volume of the outer oil bladder.

[0047] By defining the dynamic equation of the mechanical motion model, an accurate physical description of the buoy motion is established based on the principles of fluid mechanics and material mechanics, providing a reliable theoretical basis for the heterogeneous digital twin model.

[0048] In some embodiments, historical running data of the cross-section drift buoy is acquired, and parameters of the mechanical motion model are updated by least squares method OLS according to the historical running data and preset laboratory parameters, to obtain an updated mechanical motion model.

[0049] In some embodiments, the error function is used to represent the changes in volume error and different material compressibility at different depths.

[0050] wherein the expression of the error function is:

[0051]

[0052] wherein, error function; although the error function is composed of linear terms of errors, the existence of unmodelable items and ignored items in the mechanical motion model and different materials is considered, and therefore the error function is represented by .

[0053] By defining the mathematical expression of the error function, the difference between the mechanical motion model and the actual observation value is formalized as a calculable function relationship, providing a clear learning goal for the neural network, while retaining the physical interpretability of each parameter, so that error analysis can be traced back to specific physical factors, providing directional guidance for system optimization.

[0054] Error calculation step S2: Calculate the mean and standard deviation of the error values by the neural network, correct the error values according to the mean, and obtain the heterogeneous digital twin model according to the error-corrected mechanical motion model;

[0055] Optionally, the neural network is a Monte Carlo Dropout neural network. Compared with traditional methods that need a complex Bayesian framework or ensemble learning, the Monte Carlo Dropout technology does not need to increase the number of models or significantly change the network structure, can perform multiple sampling only by activating Dropout in the test phase, has small computational overhead, is very suitable for deployment of edge devices such as buoys with limited computing resources, and makes the final model not overly dependent on any specific neuron or feature by randomly discarding neurons during the training process, thereby enhancing the robustness to noise and outliers and improving the generalization ability on unseen data.

[0056] Parameter updating step S3: Obtain the predicted running state of the profile drift buoy according to the heterogeneous digital twin model, calculate the model error result according to the predicted running state and the actual running data, define the error threshold according to the standard deviation, calculate the loss function according to the error threshold and the model error result, and update the parameters of the heterogeneous digital twin model according to the loss function.

[0057] Specifically, the profile drift buoy may face the problem of parameter drift during actual use, that is, there is a preset deviation between the parameters of the updated heterogeneous digital twin model and the parameters of the mechanical motion model. Therefore, in order to always ensure the prediction accuracy of the heterogeneous digital twin model, the robust least squares method RLS is used for parameter identification when the parameters of the heterogeneous digital twin model drift.

[0058] However, the defect of the robust least squares method RLS is that the threshold of outliers is fixed. Therefore, the error threshold is defined according to the standard deviation by the 3-sigma criterion, the error threshold is used as the threshold of the Huber loss function in the robust least squares method, and the threshold adaptation in the parameter identification process is realized.

[0059] By dynamically defining the error threshold by the 3-sigma criterion and calculating the Huber loss function, adaptive processing of abnormal data is realized. When the error is within a reasonable range, the square loss is used to ensure the accuracy of parameter updating, and when an abnormal large error occurs, the linear loss is switched to suppress the interference of outliers, so that the parameter updating process not only maintains the optimization characteristics of the least squares method, but also has the robustness against data noise.

[0060] In some embodiments, the calculation model of the Huber loss function is:

[0061]

[0062] wherein, is a Huber loss function; is a model error result; is an error threshold.

[0063] Specifically, the Huber loss function is a piecewise function, when the absolute value of the model error result is greater than the error threshold, a linear loss calculation is adopted, which can ensure robustness, so that when the error is maximum, the loss value increases linearly with the error, rather than quadratically, reducing the weight of large errors and preventing them from excessively affecting parameter updates;

[0064] When the absolute value of the model error result is less than or equal to the error threshold, a square loss calculation is adopted, which can ensure accuracy, so that when the error is small, its behavior is consistent with ordinary least squares, which can provide accurate and fast convergence.

[0065] By defining the calculation model of the Huber loss, the loss calculation process is standardized, ensuring that the optimization behavior under large and small error conditions meets the design expectations, providing a clear and differentiable optimization target for parameter updates, thereby ensuring the convergence and stability of the entire adaptive system.

[0066] In some embodiments, the parameter updating step S3 specifically includes:

[0067] The prediction error between the error value and the mean value is obtained through the neural network, the maximum mean difference between the prediction error and the error value is calculated according to the loss function, and the parameters of the heterogeneous digital twin model are updated according to the maximum mean difference, and the heterogeneous digital twin model is used to generate decisions according to the profile drift buoy.

[0068] The model parameters are updated by calculating the prediction error through the neural network and combining the maximum mean difference, which quantifies the distribution difference between the prediction error and the actual error, and avoids performance degradation in the model updating process by forcing to maintain distribution consistency, so that the parameter optimization process not only considers the numerical size of the error but also focuses on the statistical characteristics of the error, thereby improving the scientificity of parameter updating and the stability of model updating, and ensuring the reliability of the heterogeneous digital twin model in long-term operation.

[0069] Specifically, although the influence of the error value is considered when updating the parameters, the adaptability of the error value output by the updated mechanical motion model to the neural network still needs to be considered, that is, although the updated mechanical motion model has smaller errors, due to inconsistent error distribution, the output of the mechanical motion model and the neural network will have larger errors.

[0070] Therefore, in order to avoid this situation, the maximum mean difference is introduced to measure the distribution deviation between the error value of the mechanical motion model and the actual running data and the prediction error output by the neural network, and the error value and the prediction error are forced to be consistent in distribution, and the maximum mean difference between the errors is taken as the loss term of parameter update to constrain the parameter update process.

[0071] Specifically, as shown in the figure, Figure 2 The flow of the heterogeneous digital twin model construction method is as follows:

[0072] A mechanical motion model of the profile drift buoy is constructed, and actual running data of the profile drift buoy is obtained based on the mechanical motion model;

[0073] An error function is defined according to the mechanical motion model, and an error value between the mechanical motion model and the actual running data is calculated according to the error function;

[0074] The mean and standard deviation of the error value are calculated by a neural network, the error value is corrected according to the mean, and a heterogeneous digital twin model is obtained according to the error-corrected mechanical motion model;

[0075] The predicted running state of the profile drift buoy is obtained according to the heterogeneous digital twin model, and a model error result is calculated according to the predicted running state and the actual running data;

[0076] An error threshold is defined by the 3-sigma rule and the standard deviation, it is judged whether the model error result exceeds the threshold, and a Huber loss function is calculated;

[0077] The prediction error between the error value and the mean is obtained by a neural network, and the maximum mean difference MMD of the prediction error and the error value is calculated in combination with the Huber loss function;

[0078] The Huber loss function and the maximum mean difference MMD are combined as an optimization target, and data noise constraint and error distribution constraint are realized by robust recursive least squares RRLS, so as to update the parameters of the heterogeneous digital twin model.

[0079] Specifically, the error value is dynamically weighted by the Huber loss function to suppress the influence of abnormal values, and the distribution consistency of the error value and the prediction error is forced by the maximum mean difference MMD.

[0080] Based on the above heterogeneous digital twin model construction method, the error value between the mechanical motion model and the actual running data is calculated by constructing the mechanical motion model and defining the error function, which can establish the deviation quantification mechanism between the physical model and the real system, and provide accurate data basis for subsequent error correction; then the mean and standard deviation of the error are calculated by the neural network, and the error is corrected, which compensates for the complex nonlinear error that the physical model is difficult to accurately describe, and finally forms a heterogeneous digital twin model that integrates physical mechanism and data-driven, realizing high-precision modeling and prediction of the running state of the profile drift buoy.

[0081] In some embodiments, as shown in Figure 3 The heterogeneous digital twin model construction method further comprises:

[0082] Neural network updating step S4: if the parameters of the updated heterogeneous digital twin model and the parameters of the mechanical motion model have a preset deviation, the parameters of the mechanical motion model are updated, the error value and the mean and standard deviation of the error value are recalculated, and the neural network is updated.

[0083] Specifically, further considering that the parameters of the updated heterogeneous digital twin model after the error value correction by the Monte Carlo Dropout neural network and the parameters of the mechanical motion model still have a preset deviation, the Monte Carlo Dropout neural network needs to be updated online, the parameters of the mechanical motion model are repeatedly updated, the error value between the mechanical motion model and the actual running data and the mean and standard deviation of the error value are recalculated, thereby updating the Monte Carlo Dropout neural network, and the parameters are fine-tuned to adapt to the new state;

[0084] According to the feature extraction layer of the fixed parameter frozen neural network, the output layer is fine-tuned to prevent overfitting and reduce the amount of calculation;

[0085] Define the loss function of the neural network;

[0086] Enable Dropout for training, and keep the Dropout layer in the on state during fine-tuning. Each time the forward propagation is still random neuron dropout, so that the training process can continue to maintain the uncertainty estimation ability of the model.

[0087] Use an optimization algorithm to perform a few rounds of iteration on a new data set to adjust the parameters of the unfrozen layers.

[0088] Evaluate the performance of the fine-tuned network on a validation set that did not participate in training;

[0089] If the performance meets the standard, replace the original parameters with the fine-tuned neural network parameters to complete the online update; if the performance does not meet the standard, adjust the hyperparameters for fine-tuning again, or trigger a more complex update process.

[0090] By triggering the retraining process when the preset deviation between the model parameters and the physical model parameters is detected, the model drift problem caused by environmental changes or equipment aging is solved by periodically refreshing the entire modeling process to track system dynamic changes, thereby realizing the full life cycle adaptability of the heterogeneous digital twin model, prolonging the effective use period of the model, and ensuring the consistency of simulation accuracy and physical entities.

[0091] It should be noted that the above is a reference mode of the cross-section drift buoy heterogeneous digital twin model construction method, and the present application is not limited thereto.

[0092] The embodiment of the present application realizes the complementary advantages of physical mechanism and data-driven, so that the heterogeneous digital twin model not only retains the explainability and robustness of the physical model, but also compensates for the complex nonlinear error that the physical model is difficult to accurately describe through the neural network, finally achieves the purpose of high-precision modeling, and at the same time, through the parameter updating mechanism, ensures that the heterogeneous digital twin model can adapt to the complex and changeable marine environment, provides a more reliable virtual simulation platform for buoy state monitoring, solves the technical problems in the prior art that the physical modeling method relies on accurate mechanism knowledge, but the complex working conditions of the buoy in the deep sea environment cause the model parameters to drift easily, and it is difficult to update in real time, and the adaptability is insufficient; the pure data-driven method lacks physical explainability, requires extremely high training data volume and quality, has poor generalization ability when the data is insufficient, and has high computational complexity, which is difficult to deploy on the lightweight edge device of the buoy.

[0093] Finally, it should be noted that: each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to.

[0094] The above embodiments are only used to illustrate the technical solutions of the present application and not to limit them; although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the specific embodiments of the present application can be modified or some technical features can be replaced by equivalent; without departing from the spirit of the technical scheme of the present application, they should be covered in the technical scheme range of the present application.

Claims

1. A cross-section profile drift buoy isomer digital twin model construction method, characterized in that, The method comprises the following steps: An error calculation step: constructing a mechanical motion model of a profile drift buoy, obtaining actual running data of the profile drift buoy based on the mechanical motion model, defining an error function according to the mechanical motion model, and calculating an error value of the mechanical motion model and the actual running data according to the error function; A model construction step: calculating a mean value and a standard deviation of the error value through a neural network, performing error correction on the error value according to the mean value, and obtaining a heterogeneous digital twin model according to the error-corrected mechanical motion model; A parameter updating step: obtaining a predicted running state of the profile drift buoy according to the heterogeneous digital twin model, calculating a model error result according to the predicted running state and the actual running data, defining an error threshold value according to the standard deviation, calculating a loss function according to the error threshold value and the model error result, updating parameters of the heterogeneous digital twin model according to the loss function, and using the heterogeneous digital twin model to generate a decision according to the profile drift buoy; The dynamics of the mechanical motion model is represented as: wherein, is the mass of the profiled drift float; is the depth of the water; is the volume of the glass sphere shell; is the compressibility of the glass sphere shell; is the volume of the outer oil bladder; is the compressibility of the hydraulic oil; is the non-modeled item; is the compressibility of the non-modeled item; is the drag coefficient of the profiled drift float; is the velocity of the profiled drift float in motion; The expression of the error function is: wherein is the error function.

2. The cross-sectioned drift float isomeric digital twin model construction method according to claim 1, characterized in that, The parameter updating step specifically comprises: Obtaining a prediction error between the error value and the mean value through the neural network, calculating a maximum mean difference between the prediction error and the error value according to the loss function, and updating the parameters of the heterogeneous digital twin model according to the maximum mean difference.

3. The cross-sectioned drift float isomeric digital twin model construction method according to claim 1, characterized in that, The method further comprises: A neural network updating step: if there is a preset deviation between the updated parameters of the heterogeneous digital twin model and the parameters of the mechanical motion model, updating the parameters of the mechanical motion model, recalculating the error value, the mean value and the standard deviation of the error value, and updating the neural network.

4. The cross-sectioned drift float isomeric digital twin model construction method according to claim 1, characterized in that, The calculation model of the loss function is: wherein, is a loss function; is a model error result; is an error threshold.

Images