Digital twinning performance deduction model of long short-term memory network based on physical information constraint and construction method of digital twinning performance deduction model

Abstract

The invention belongs to the technical field of fault prediction, and particularly discloses a physical information constraint-based digital twinning performance deduction model of a long short-term memory network and a construction method thereof, and the method comprises the steps: training a neural network model through a training set, and taking the trained neural network model as the digital twinning performance deduction model; the training set comprises a pavement physical parameter X and a pavement performance parameter Y; in the neural network model, a first LSTM network extracts time sequence characteristics of a parameter X to obtain a preliminary performance parameter predicted value # imgabs0 #; the automatic differential module adopts an automatic differential technology to calculate a partial derivative # imgabs2 # of # imgabs1 # for time t and a partial derivative # imgabs4 # of # imgabs3 # for a parameter X; and the second LSTM network obtains # imgabs6 # according to # imgabs5 # and the parameter X, and adjusts # imgabs8 # according to # imgabs7 #, so that the evolution of # imgabs8 # accords with physical priori knowledge, and physical information constraint is realized. According to the method, the digital twinning performance deduction precision in a complex environment can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of fault prediction technology, and more specifically, relates to a digital twin performance deduction model of a long short-term memory network based on physical information constraints and a construction method thereof. Background Art

[0002] Digital twin technology, by dynamically mapping physical entities and virtual models, has become a core technology in fields such as intelligent manufacturing, smart energy, and aerospace. Its core value lies in predicting and optimizing system states through real-time performance simulation, thereby reducing maintenance costs and improving operational efficiency. However, with the increasing complexity of industrial systems, traditional performance simulation methods face two major challenges: how to extract dynamic patterns from multi-source heterogeneous data, and how to ensure that the simulation results are consistent with the rationality of physical mechanisms.

[0003] Currently, mainstream digital twin performance simulation methods can be divided into two categories: model-based and data-driven. Traditional simulation methods based on explicit mathematical models construct deterministic mechanistic models based on physical laws, such as using differential equations or conservation equations to describe the system's dynamic behavior. While these methods offer high accuracy for simple systems or steady-state scenarios, they rely on explicit mathematical expressions for modeling complex dynamic systems, leading to significant limitations. For example, in the case of multi-body coupling, nonlinear, time-varying, or high-dimensional multi-physics systems, such as those in pavement engineering, the modeling process requires linearization assumptions, order reduction, or neglect of minor factors, leading to deviations from the actual physical process, resulting in model distortion. Furthermore, real-time solution of explicit models carries high computational costs, making them difficult to meet the low-latency requirements of digital twins. Furthermore, model parameter calibration relies on expert experience, making it less adaptable to unknown operating conditions. Purely data-driven methods utilize deep learning algorithms to directly learn the system's dynamic characteristics from historical data, avoiding the complexity of explicit modeling. Time series models, such as long-short-term memory networks, are widely used for equipment degradation prediction and dynamic parameter inversion due to their ability to capture long-term dependencies. However, such methods rely entirely on data distribution, ignore physical prior knowledge, and are easily affected by data noise, sparsity, and distribution offset. As a result, the deduction results may violate basic physical laws such as conservation of energy and conservation of mass, and the generalization ability in scenarios outside the training data is insufficient. Summary of the Invention

[0004] In response to the above-mentioned defects or improvement needs of the existing technology, the present invention provides a digital twin performance deduction model based on a long short-term memory network with physical information constraints and a construction method thereof, the purpose of which is to effectively improve the accuracy of digital twin road performance deduction in complex environments.

[0005] To achieve the above objectives, according to a first aspect of the present invention, a method for constructing a digital twin performance deduction model based on a long short-term memory network with physical information constraints is proposed, comprising the following steps: Train the neural network model using the training set, and use the trained neural network model as the digital twin performance deduction model; The training set includes road surface physical parameters X and road surface performance parameters Y, wherein the physical parameters X include time series dynamic parameters and static parameters; The neural network model includes a first LSTM network, an automatic differentiation module and a second LSTM network, wherein the first LSTM network is used to extract the time series characteristics of the parameter X and obtain the preliminary performance parameter prediction value ; The automatic differentiation module is used to calculate using automatic differentiation technology Partial derivative with respect to time t as well as Partial derivative with respect to parameter X ; The second LSTM network is used to And the parameter X, get the predicted value , according to the predicted value , the initial performance parameter prediction values Make adjustments so that The evolution of is consistent with physical prior knowledge and realizes physical information constraints.

[0006] As a further preferred method, when training the neural network model, the loss function used is as follows: ; ; ; in, represents the data-driven loss term, 、 Respectively represent i The actual value of the pavement performance parameters and the preliminary performance parameter prediction values ​​corresponding to the group data, N represents the amount of data in the training set, represents a hyperparameter; represents the physical constraint loss term, The definition is as follows: ; in, 、 Represent the first LSTM network and the second LSTM network respectively, 、 Represent the parameters of the first LSTM network and the second LSTM network respectively.

[0007] As a further preference, the training set includes a simulation training set and a measured training set, the parameters X and Y in the simulation training set are obtained by simulation after constructing a pavement finite element model, and the parameters X and Y in the measured training set are obtained by actual monitoring of the pavement.

[0008] As a further preferred method, when training the neural network model using the training set, a phased training strategy is adopted: Pre-training stage: The neural network model is pre-trained through the simulation training set. ; Joint training phase: The pre-trained neural network model is trained using the actual training set, and the dynamic adjustment is made during the process. , and obtain the trained neural network model.

[0009] As a further preferred method, the joint training phase dynamically adjusts The method is: ,in and is a hyperparameter with a range of , is the error of the neural network model on the validation set.

[0010] As a further preferred method, the normal degradation and extreme working conditions of the road surface are covered during the simulation based on the finite element model, and the amount of simulation training set data is expanded; the data in the training set are all processed by maximum and minimum normalization.

[0011] As a further preferred embodiment, the static parameters are broadcasted to the time step dimension and concatenated with the time series dynamic parameters to form the road surface physical parameters.

[0012] As a further preference, the time-series dynamic parameters include traffic load and climate environment, the static parameters include pavement structure attributes; and the pavement performance parameters are pavement roughness index IRI, pavement ride quality index RQI or pavement quality index PQI.

[0013] According to the second aspect of the present invention, a digital twin performance deduction model is provided, which is constructed using the above-mentioned method for constructing a digital twin performance deduction model based on a long short-term memory network with physical information constraints.

[0014] According to a third aspect of the present invention, a digital twin road surface performance deduction method is provided, and the digital twin system of the road surface uses the above-mentioned digital twin performance deduction model to perform road surface performance deduction.

[0015] In general, the above technical solutions conceived by the present invention have the following technical advantages compared with the existing technology: 1. Aiming at the complex dynamic characteristics of pavement digital twin performance deduction, the present invention adopts multi-source physical data that combines time series dynamics and statics, calculates the sensitivity of performance indicators to input physical quantities through automatic differentiation technology, and uses a dual LSTM network to learn the multi-source parameter coupling effect, effectively improving the accuracy of digital twin pavement performance deduction in complex environments.

[0016] 2. This invention uses automatic differentiation to calculate the sensitivity of performance indicators to key physical quantities, dynamically analyzing the inherent relationship between physical laws and performance degradation. This implicit constraint is embedded into the neural network training process, avoiding the limitations of explicit equation modeling. Compared to traditional methods, this invention utilizes implicit physical constraints and a multi-source data collaboration mechanism to improve robustness to data noise and sparsity while ensuring the consistency of physical constraints, providing reliable technical support for digital twin-driven performance deduction.

[0017] 3. The present invention adopts a phased training strategy, learning the prior knowledge of physical laws through pre-training with simulation data, and then dynamically optimizing with measured data, breaking through the bottleneck of physical law representation of traditional models and effectively solving the problem of data sparsity in complex scenarios.

[0018] 4. In view of the heterogeneous characteristics of multi-source data, an input feature vector integrating physical information is constructed based on a dynamic-static parameter fusion strategy, covering multi-dimensional information such as pavement structure properties, climate environment, and traffic load. Combined with the long short-term memory network's ability to model the temporal dependence of the degradation process, high-precision performance deduction of the digital twin system is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a method for constructing a digital twin performance deduction model of a long short-term memory network based on physical information constraints in an embodiment of the present invention.

[0020] Figure 2 Schematic diagram of error changes during model training in an embodiment of the present invention.

[0021] Figure 3 This is a comparison chart of the actual value and the predicted value of the embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0023] The embodiment of the present invention provides a method for constructing a digital twin performance deduction model based on a long short-term memory network with physical information constraints, such as Figure 1 As shown, the following steps are included: (1) Construction of training set.

[0024] (1.1) Data acquisition: Simulation data and measured data are acquired through pavement finite element model simulation and field measurement, forming a dataset. Both simulation data and measured data include pavement physical parameters X and corresponding pavement performance parameters Y. Physical parameters X include time-series dynamic parameters and static parameters.

[0025] Furthermore, a finite element model of the pavement is constructed based on the pavement physical characteristics and failure mechanism, and multi-physics field simulation is carried out to generate a data set of physical parameters and performance parameters. , subscript Refers to the time step. Finite element model simulation covers normal degradation and extreme operating conditions, generates sample data, and implements neural network data augmentation.

[0026] Physical parameters Contains timing dynamic parameters With static parameters , performance parameters These are parameters that can reflect the health status of the digital twin system corresponding to the road surface; for the static parameters in the physical parameters, they are broadcast to the time step dimension and spliced ​​with the time series dynamic parameters.

[0027] Specifically, time-series dynamic parameters include traffic load and climate. Static parameters include pavement structural attributes, including the pavement's layered structure, total number of pavement layers, the functional type of each layer, and material type, i.e., the material category of each structural layer. Pavement performance parameters include the roughness index (IRI), the ride quality index (RQI), or the pavement quality index (PQI).

[0028] (1.2) Data preprocessing: The time series segmentation method is used to divide the acquired data set into continuous intervals in chronological order. The first 70% of the period is the training set, the middle 15% is the validation set, and the last 15% is the test set to avoid future information leakage.

[0029] Use maximum and minimum value normalization for preprocessing, and the calculation formula is as follows: ; Where, is the global minimum value of the corresponding feature in the data set, which is calculated by traversing the time series data of all samples; is the global maximum value of the corresponding feature in the data set, calculated in the same way as ; is the normalized eigenvalue, and its range is .

[0030] (2) Model training.

[0031] The neural network model is trained through the training set, and the trained neural network model is used as the digital twin performance deduction model.

[0032] (2.1) Model structure: The neural network model includes the first LSTM network , automatic differentiation module and the second LSTM network , where the first LSTM network is used to extract the time series features of parameter X and obtain the preliminary performance prediction value ; The automatic differentiation module is used to calculate using automatic differentiation technology Partial derivative with respect to time t as well as Sensitivity to parameter X ; The second LSTM network is used to And the parameter X, get the predicted value , and then based on the predicted value , the predicted value of performance parameters Adjust to get the final output, so as to embed physical information and realize physical information constraints. The evolution of is consistent with physical a priori knowledge. Specifically as follows: 1) First LSTM network :Through the time series feature extraction network , extract the long-term dependency of parameters and output the initial performance characterization prediction value: ; Where, is the predicted value of the performance characterization quantity; is the input characteristic parameter; t is the time variable; are network parameters.

[0033] 2) Automatic differentiation module: Calculate using automatic differentiation technology Partial derivative with respect to time t and its sensitivity to the physical parameter X .

[0034] 3) Second LSTM network :Will Together with the original feature X, it is input into the second LSTM network , implicitly fitting the dynamic relationship between degradation rate and physical parameters: ; Where, is the network parameter.

[0035] In order to balance the computational accuracy and computational complexity, only the first-order partial derivative is taken here. Direct online learning The nonlinear mapping replaces the traditional explicit equation and avoids the simplification error of manually defined equations.

[0036] (2.2) Loss function construction: Define the joint loss function , integrating data-driven errors with physical degradation law constraints: ; ; ; in, is the data-driven loss term, calculating the predicted value and the true value The mean square error of is the physical constraint loss term, which constrains the matching of the degradation rate with the physical model; is a hyperparameter optimized by grid search.

[0037] (2.3) Phased training strategy: During the pre-training phase, only simulated data is used for training and Network, fixed , learn the prior knowledge of physical laws. In the joint training phase, the measured data is introduced according to Dynamic modulation weights, and is a hyperparameter with a range of , is the error of the network on the validation set. When the validation error increases, Reduce, thereby balancing physical constraints with data-driven adaptive learning. An early stopping mechanism is used during training to prevent overfitting. When the model's loss function value on the validation set does not show a downward trend for five consecutive training cycles, the training process is automatically terminated.

[0038] The present invention calculates the sensitivity of performance indicators to key physical quantities through automatic differentiation technology, dynamically analyzes the intrinsic relationship between physical laws and performance degradation, and embeds implicit constraints into the neural network training process to avoid the limitations of explicit equation modeling. At the same time, in view of the heterogeneous characteristics of multi-source data on the road surface in the time and space dimensions, an input feature vector that integrates physical information is constructed based on the time and space alignment strategy, covering multi-dimensional information such as material properties, environmental loads, and monitoring signals. Combined with the long and short-term memory network's ability to model the temporal dependence of the degradation process, high-precision performance deduction of the digital twin system is achieved. Compared with traditional methods, the present invention uses implicit physical constraints and a multi-source data collaboration mechanism to improve the robustness to data noise and sparsity while ensuring the consistency of physical constraints, providing reliable technical support for performance deduction driven by road digital twins.

[0039] The following are specific embodiments: The Long-Term Pavement Performance (LTPP) database from the United States was selected as the data source for method validation due to its authority and data completeness in the field of pavement engineering. As the world's largest long-term pavement monitoring project, the LTPP database systematically collects performance data throughout the entire life cycle, covering different climate zones, pavement structure types, and traffic load levels. Its multi-dimensional, long-term characteristics provide an ideal data foundation for validating prediction models constrained by physical information. The database not only contains regular inspection records of performance indicators such as the International Roughness Index (IRI) and rutting depth, but also integrates the material properties of each structural layer of the pavement, the annual traffic axle load spectrum, and environmental parameters continuously monitored by meteorological stations. It can fully depict the dynamic relationship between pavement performance degradation and multi-source physical factors.

[0040] In terms of the selection of performance characterization quantities, IRI, as one of the core monitoring indicators of LTPP, has a clear physical meaning and a standardized quantification method. IRI inverts the elevation changes of the longitudinal section of the pavement through the dynamic response of the vehicle, and its numerical growth directly represents the cumulative effect of material damage in the pavement structure layer and the decline in driving comfort. Compared with other performance indicators, IRI has the characteristics of continuous monitoring throughout the entire life cycle, and can capture the full-stage degradation trajectory from the initiation of initial microcracks to the later structural fragmentation. In addition, the evolution process of IRI is driven by the coupling of multiple physical fields such as repeated traffic loads, temperature and stress alternation, and moisture penetration. The complexity of its inherent degradation law can effectively verify the modeling ability of the physical information constraint mechanism for nonlinear dynamic systems.

[0041] The input data relies on the multi-source heterogeneous data structure of the LTPP database, extracting key physical parameters from three dimensions: structural properties, traffic loads, and climate. By integrating these multi-source features, the model is able to analyze the nonlinear evolution of IRI under the interaction of traffic, climate, and materials, thereby maintaining prediction accuracy under complex operating conditions. The structured storage and quality control system of LTPP data further ensure the physical rationality and temporal integrity of the characteristic parameters, providing reliable data support for verifying the generalization capabilities of the proposed method.

[0042] During data preprocessing, categorical features, such as the material types of the various pavement layers, were converted into binary vectors using one-hot encoding to eliminate the interference of unordered categorical variables on model training. Numerical features, such as axle load times and temperature, were normalized using maximum and minimum values ​​to linearly map the data to the range of 0 to 1. Normalization parameters were calculated based on global statistics of the training set to prevent information leakage in the test set. The dataset was partitioned chronologically into training, validation, and test sets, with a split ratio of 70%, 15%, and 15%, respectively, to ensure the continuity of the time series.

[0043] The model uses a temporal feature extraction network and physical constraint networks The hyperparameter configurations of the two networks are shown in Table 1: the number of hidden units is set to 64 to balance model capacity and computational efficiency; the dropout rate is set to 0.1 to suppress overfitting; the training batch size is set to 2, and the time step size is set to 2 to capture the evolution of road performance. .

[0044] Model training is divided into two stages: pre-training and joint optimization. In the pre-training stage, only simulation data is used for training. and network, initialize network parameters, fix physical constraint weights, and learn prior knowledge of physical laws; in the joint training phase, introduce measured data, according to Dynamic modulation weight, where the initial value , attenuation coefficient , The network's error on the validation set is expressed as ∑(x) / (y) / (z ...

[0045] Figure 2The figure shows the trend of mean squared error during training. The model's loss quickly converged to a stable range within 30 training epochs, and the validation set error fluctuated below the expected threshold for multiple training epochs, demonstrating the model's excellent generalization performance and training stability.

[0046] Figure 3 Comparison of the temporal evolution of the actual and predicted IRI values ​​on the test set revealed high consistency between the predicted and measured values ​​in both normal and accelerated degradation phases. The model was particularly able to accurately capture sudden changes in pavement performance in regions of abrupt nonlinear changes. The overall mean squared error (MSE) on the test set was 0.0051, significantly lower than that of the traditional LSTM model, demonstrating the effectiveness of the physical information constraint mechanism in improving prediction accuracy.

[0047] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing a digital twin performance deduction model based on a long short-term memory network with physical information constraints, characterized by: The steps include: Train the neural network model using the training set, and use the trained neural network model as the digital twin performance deduction model; The training set includes road surface physical parameters X and road surface performance parameters Y, wherein the physical parameters X include time series dynamic parameters and static parameters; The neural network model includes a first LSTM network, an automatic differentiation module and a second LSTM network, wherein the first LSTM network is used to extract the time series characteristics of the parameter X and obtain the preliminary performance parameter prediction value ; The automatic differentiation module is used to calculate using automatic differentiation technology Partial derivative with respect to time t as well as Partial derivative with respect to parameter X ; The second LSTM network is used to And the parameter X, get the predicted value , according to the predicted value , the initial performance parameter prediction values Make adjustments so that The evolution of is consistent with physical prior knowledge and realizes physical information constraints.

2. The method for constructing a digital twin performance deduction model based on a long short-term memory network with physical information constraints according to claim 1 is characterized in that: The loss function used when training the neural network model as follows: in, represents the data-driven loss term, 、 Respectively represent i The actual value of the pavement performance parameters and the preliminary performance parameter prediction values ​​corresponding to the group data, N represents the amount of data in the training set, represents a hyperparameter; represents the physical constraint loss term, The definition is as follows: in, 、 Represent the first LSTM network and the second LSTM network respectively, 、 Represent the parameters of the first LSTM network and the second LSTM network respectively.

3. The method for constructing a digital twin performance deduction model based on a long short-term memory network with physical information constraints according to claim 2 is characterized in that: The training set includes a simulation training set and a measured training set. The parameters X and Y in the simulation training set are obtained by simulation after constructing a road surface finite element model, and the parameters X and Y in the measured training set are obtained by actual monitoring of the road surface.

4. The method for constructing a digital twin performance deduction model based on a long short-term memory network with physical information constraints according to claim 3 is characterized in that: When training the neural network model through the training set, a phased training strategy is adopted: Pre-training stage: The neural network model is pre-trained through the simulation training set. ; Joint training phase: The pre-trained neural network model is trained using the actual training set, and the dynamic adjustment is made during the process. , and obtain the trained neural network model.

5. The method for constructing a digital twin performance deduction model based on a long short-term memory network with physical information constraints according to claim 4 is characterized in that: Dynamic adjustment during joint training phase The method is: ,in and is a hyperparameter with a range of , is the error of the neural network model on the validation set.

6. The method for constructing a digital twin performance deduction model based on a long short-term memory network with physical information constraints according to claim 3 is characterized in that: Finite element model-based simulation covers normal road degradation and extreme working conditions, expanding the amount of simulation training set data; all data in the training set are processed by maximum and minimum normalization.

7. The method for constructing a digital twin performance deduction model based on a long short-term memory network with physical information constraints according to claim 1 is characterized in that: The static parameters are broadcast to the time step dimension and concatenated with the time series dynamic parameters to form the road surface physical parameters.

8. The method for constructing a digital twin performance deduction model based on a long short-term memory network with physical information constraints according to any one of claims 1 to 7, characterized in that: The time series dynamic parameters include traffic load and climate environment, the static parameters include pavement structure attributes; the pavement performance parameters are pavement roughness index IRI, pavement ride quality index RQI or pavement quality index PQI.

9. A digital twin performance deduction model, characterized in that: It is constructed using the method for constructing a digital twin performance deduction model of a long short-term memory network based on physical information constraints as described in any one of claims 1 to 8.

10. A digital twin performance deduction method, characterized in that: The digital twin system of the road surface uses the digital twin performance deduction model as described in claim 9 to deduce road surface performance.

Images