Data assimilation method and equipment based on digital twinning and virtual-real verification, and medium

Abstract

The invention relates to the technical field of data assimilation, and particularly provides a data assimilation method and equipment based on digital twinning and virtual-real verification, and a medium. The data assimilation method based on digital twinning and virtual-real verification comprises the following steps: constructing an agent model of an observation object, and obtaining a prediction result output by the agent model and an actual observation result of the observation object; consistency verification is carried out on the prediction result and the actual observation result, and the prediction result and the actual observation result which pass verification are obtained; performing sequence alignment processing on the predicted result and the actual observation result which pass the verification to obtain a processed predicted result and an actual observation result; and performing data assimilation on the processed prediction result based on the processed actual observation result to obtain a target prediction result of the observation object. According to the method, the technical problems that actual observation data has noise and has sparsity are solved through consistency verification and sequence alignment processing, so that the accuracy and reliability of an agent model prediction result are improved.

Description

Technical Field

[0001] The present disclosure relates to the technical field of data assimilation, and in particular to a data assimilation method, device and medium based on digital twins and virtual-reality verification. Background Art

[0002] Digital twin technology simulates or reflects the state and performance of the observed object by constructing a virtual model corresponding to the observed object. In the process of building and maintaining the virtual model, data assimilation technology is needed to effectively integrate the real-time observation data collected by the observed object into the virtual model, thereby ensuring that the model can accurately reflect the current and predicted state of the physical object.

[0003] In related technologies, observation data are usually input into filtering algorithms for data assimilation. However, since the observation data has local and unevenly distributed noise and is sparse, the virtual model cannot accurately capture the true state of the observed object, which in turn affects the accuracy and reliability of the virtual model's prediction results. Summary of the invention

[0004] In view of this, the exemplary embodiments of the present disclosure provide a data assimilation method, device and medium based on digital twins and virtual-reality verification to solve the problems existing in the related art.

[0005] One aspect of an exemplary embodiment of the present disclosure provides a data assimilation method based on digital twins and virtual-real verification, the method comprising:

[0006] Constructing a proxy model of the observed object, and obtaining the prediction result output by the proxy model and the actual observation result of the observed object;

[0007] Performing consistency check on the prediction result and the actual observation result to obtain the prediction result and the actual observation result that pass the check;

[0008] Perform sequence alignment processing on the predicted results and actual observation results that have passed the verification to obtain the processed predicted results and actual observation results;

[0009] The processed prediction results are data assimilated based on the processed actual observation results to obtain the target prediction results of the observed object.

[0010] Another aspect of the exemplary embodiments of the present disclosure provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described in the exemplary embodiments of the present disclosure.

[0011] According to another aspect of the exemplary embodiments of the present disclosure, a computer-readable storage medium is provided, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the method described in the exemplary embodiments of the present disclosure is implemented.

[0012] According to another aspect of the exemplary embodiments of the present disclosure, a computer program product is provided, including a computer program / instruction. When the computer program / instruction is executed by a processor, the method described in the exemplary embodiments of the present disclosure is implemented.

[0013] As will be described in detail below, according to a data assimilation method based on digital twins and virtual-reality verification in an embodiment of the present disclosure, a proxy model of the observed object is constructed, and the prediction results output by the proxy model and the actual observation results of the observed object are obtained; the prediction results and the actual observation results are checked for consistency to obtain the prediction results and the actual observation results that have passed the check; the prediction results and the actual observation results that have passed the check are sequence aligned to obtain the processed prediction results and the actual observation results; the processed prediction results are data assimilated based on the processed actual observation results to obtain the target prediction results of the observed object. Therefore, the data assimilation method provided by the present disclosure solves the technical problems of noise and sparsity in the actual observation data through consistency verification and sequence alignment processing, thereby improving the accuracy and reliability of the prediction results of the proxy model. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The above and other purposes, features and advantages of the present disclosure will become more apparent by describing the embodiments of the present disclosure in more detail in conjunction with the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of the present disclosure and constitute a part of the specification. Together with the embodiments of the present disclosure, they are used to explain the present disclosure and do not constitute a limitation of the present disclosure. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0015] Figure 1 A flowchart of a data assimilation method based on digital twins and virtual-real verification provided as an example in the present disclosure;

[0016] Figure 2 A schematic diagram of the working principle of a data assimilation method based on digital twins and virtual-real verification provided as an example in the present disclosure;

[0017] Figure 3 A schematic diagram of a data assimilation process provided as an example in the present disclosure;

[0018] Figure 4 This is a schematic block diagram of functional modules of a data assimilation device based on digital twin and virtual-real verification provided by an exemplary embodiment of the present disclosure;

[0019] Figure 5A structural block diagram of an electronic device provided as an example in the present disclosure;

[0020] Figure 6 A schematic diagram of a computer program product provided by an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0021] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein, which are instead provided for a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0022] It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0023] The term "including" and its variations used in this document are open inclusions, that is, "including but not limited to". The term "based on" means "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments". Relevant definitions of other terms will be given in the description below. It should be noted that the concepts of "first", "second", etc. mentioned in this disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0024] It should be noted that the modifications of "one" and "plurality" mentioned in the present disclosure are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0025] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0026] It is understandable that before using the technical solutions disclosed in the embodiments of the present disclosure, the types, scope of use, usage scenarios, etc. of the personal information involved in the present disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0027] For example, in response to receiving an active request from a user, a prompt message is sent to the user to clearly prompt the user that the operation requested to be performed will require obtaining and using the user's personal information. Thus, the user can autonomously choose whether to provide personal information to software or hardware such as an electronic device, application, server, or storage medium that performs the operation of the technical solution of the present disclosure according to the prompt message.

[0028] As an optional but non-limiting implementation, in response to receiving an active request from the user, the method of sending a prompt message to the user may be, for example, a pop-up window, in which the prompt message may be presented in text. In addition, the pop-up window may also carry a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device. It is understandable that the above notification and the process of obtaining user authorization are only illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that meet relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0029] Digital twin technology simulates or reflects the state and performance of the observed object by constructing a virtual model corresponding to the observed object. In the process of building and maintaining the virtual model, data assimilation technology is needed to effectively integrate the real-time observation data collected by the observed object into the virtual model, thereby ensuring that the model can accurately reflect the current and predicted state of the physical object.

[0030] In related technologies, observation data are usually input into filtering algorithms for data assimilation. However, since the observation data has local and unevenly distributed noise and is sparse, the virtual model cannot accurately capture the true state of the observed object, which in turn affects the accuracy and reliability of the virtual model's prediction results.

[0031] Therefore, in order to solve the above problems, the exemplary embodiment of the present disclosure provides a data assimilation method based on digital twins and virtual-reality verification, which is trained by constructing a proxy model corresponding to the observed object and constructing a data set by collecting sample points. After the training is completed, the output results of the proxy model and the observation results of the observed object are checked for consistency and sequence alignment to ensure that the output of the proxy model matches the actual observation data. Finally, the improved Transformer model is used for data assimilation to optimize the performance of the proxy model and improve the real-time performance of data processing.

[0032] For example, Figure 1 This is a flow chart of a data assimilation method based on digital twins and virtual-real verification provided by an exemplary embodiment of the present disclosure, such as Figure 1 As shown, the following steps may be specifically included:

[0033] Step S110: construct a proxy model of the observed object, and obtain the prediction results output by the proxy model and the actual observation results of the observed object.

[0034] For example, Figure 2 The following is a schematic diagram of the working principle of the data assimilation method based on digital twin and virtual-real verification provided by the present disclosure as an example. Figure 2 As shown, the construction of the proxy model may specifically include:

[0035] Step S210: Collect sample points and construct a training data set.

[0036] The sample points can be collected by constructing the sample space through Latin Hypercube Sampling (LHS) in the modern Design of Experiments (DoE) method. The sample points in the sample space and their corresponding output responses constitute the data set {X h ,f h}, where X h represents the sample parameter set, that is, the set of input variables; f h Represents the state value or output response of a sample point.

[0037] When the boundary conditions and structural parameters are known, the constructed high-fidelity sample set can be specifically expressed as: h ={x1,x2,...,x N} represents the input of the sample set, f={f1,f2,...,f N} represents the output response at each sample point.

[0038] Step S220: construct a proxy model and perform training.

[0039] Exemplarily, a sample data set including real system data and high-fidelity sample set data generated by a simulation system is learned by using a generative adversarial network (GAN). The generative adversarial network iteratively learns the mapping relationship between input and output in the adversarial process between the generator and the discriminator until the error condition is met.

[0040] Since the surrogate model adopts a supervised learning strategy, in which the control task between input and output is handled by a deep learning network, the surrogate model is suitable for predicting low-fidelity outputs under new input conditions that have never appeared in the training sample set.

[0041] Exemplarily, the proxy model can be expressed as:

[0042] f l =Γ(Xh ,t,θ) (1)

[0043] Where Γ represents the nonlinear mapping relationship between input and output in the proxy model; θ represents the key parameter set in the proxy model; t represents time; f l represents the predicted low-fidelity response.

[0044] Specifically, GAN consists of two parts: Generator (G) and Discriminator (D).

[0045] The task of the generator (G) is to generate fake samples that are as close to real data as possible. The generator receives a random noise vector p(z) as input and generates new data samples through this noise. The noise vector p(z) can come from a predefined probability distribution (such as a normal distribution), and the generated data samples imitate the characteristics of the real data in the training set to deceive the discriminator.

[0046] The task of the discriminator (D) is to distinguish whether the input sample comes from the real dataset or fake data generated by the generator.

[0047] Step S230: Determine whether the error meets a preset condition.

[0048] During the training process, the generator tries to minimize the probability that its generated samples are identified as fake, while the discriminator distinguishes true and false samples as accurately as possible. Therefore, the loss functions corresponding to the generator and the discriminator and the corresponding preset conditions can be expressed as:

[0049] in, represents the discriminator loss function; ψ represents the discriminator network parameters. in, represents the generator loss function; θ represents the input key parameters, and φ represents the generator network parameters.

[0050] When the preset conditions are met, the output results of the proxy model are subjected to consistency verification and sequence alignment processing in turn.

[0051] If the preset condition is not met, return to step S220 to continue training.

[0052] Finally, after the proxy model is trained, new data samples are generated and used for the next step of reliability analysis.

[0053] Specifically, according to the sampled latent variables, the trained generator can generate the conditional sample function y = f θ(y|x). The latent variables are sampled from the noise vector p(z). Latent variables refer to variables that cannot be observed and are used to explain the relationship between variables or to provide a deeper explanation of the observed data. Since latent variables are random, they can help the generator explore and learn the diversity of data. The conditional sample function refers to the function that the generator uses to map latent variables to the data space. It can generate new data samples based on the input noise conditions.

[0054] In the reliability analysis phase, a one-way transfer calculation method can be used. One-way transfer calculation refers to the forward propagation process from the input layer to the output layer in the generator network. In this process, the latent variables are converted into specific data samples through the network layer.

[0055] After obtaining the generated new data samples, various statistical analyses can be performed on these samples, such as calculating the mean and variance of the samples, to evaluate the accuracy of the proxy model's prediction results.

[0056] In an optional approach, in order to avoid mode collapse during training, the adversarial structure of the generator and discriminator networks, as well as the loss function, can be adjusted to improve the stability of the training process, so that the generator and discriminator maintain a good competitive relationship during the learning process, thereby generating high-quality and diverse data samples.

[0057] Step S120: Perform consistency check on the prediction results and the actual observation results to obtain the prediction results and actual observation results that pass the check.

[0058] Dynamic time warping (DTW) is used to calculate the DTW distance between the predicted results and the actual observed results;

[0059] When the DTW distance is within the preset confidence interval, it is confirmed that the predicted result and the actual observation result have passed the consistency check.

[0060] In this embodiment, the main purpose of the consistency check is to check the consistency between the time series data obtained by the proxy model and the observed object. In the process of analyzing time series data, it is necessary to consider a variety of practical factors that affect the results, such as noise, amplitude translation and expansion, linear drift, discontinuity and time axis expansion.

[0061] To overcome the above problems, Dynamic Time Warping (DTW) can be used to calculate the similarity of two time series. DTW can evaluate the similarity of two sequences by optimizing the alignment of the time axis to minimize the total distance when there is a time offset or expansion between the time series. If the DTW distance of two time series is within a preset confidence interval, the two sequences can be considered similar.

[0062] Specifically, first assume that the time series corresponding to the output of the proxy model is S, with a length of m, expressed as: S = {s1, s2, s3, ..., s m}; The time series corresponding to the observation results obtained by the sensor of the observed object is P, with a length of n, expressed as: P = {p1, p2, p3, ..., p n}.

[0063] Furthermore, the Euclidean distance between the two time series is calculated, that is, the sum of the distances between the corresponding points of the two time series is calculated, and at the same time, a point at a certain moment in the time series S corresponds to the points at multiple consecutive moments in the time series P.

[0064] The Euclidean distances between multiple moments of time series S and time series P form a distance matrix, and each element in the distance matrix represents the Euclidean distance between the i-th moment in time series S and the j-th moment in time series P.

[0065] For example, the Euclidean distance between two time series can be calculated using the following formula:

[0066] M(i,j)=|S(i)-P(j)| (2)

[0067] Among them, S(i) represents the predicted time series data of the proxy model at the i-th moment; P(j) represents the real time series data obtained by the sensor at the j-th moment.

[0068] Then, dynamic programming is used to populate the distance matrix (D), where each element (D(i,j)) represents the minimum cumulative distance from the top left corner (1,1) of the distance matrix to (i,j).

[0069] Then, the optimal path of the distance matrix is ​​found: the optimal path is the path from (1,1) to (m,n) through the distance matrix (D), and the optimal path reflects the minimum cumulative distance between the two time series.

[0070] Finally, calculate the DTW distance and determine the similarity of the two time series: the DTW distance is the value of the last element (D(m,n)) in the distance matrix (D), which represents the cumulative distance of the overall optimal alignment. Determine whether the DTW distance is within the preset credibility interval. If it is within the preset credibility interval, the two series can be considered similar.

[0071] Step S130: performing sequence alignment processing on the verified prediction results and actual observation results to obtain processed prediction results and actual observation results.

[0072] Determine the sampling frequencies corresponding to the predicted results and the actual observed results;

[0073] Based on the sampling frequency, interpolation or downsampling is used to align the predicted results and the actual observation results.

[0074] In the analysis of many dynamic systems, the time labels of the data are not synchronized due to different sampling frequencies, resulting in sequence jitter. When the data is sequence jittery, it means that there is a deviation in the alignment of the data on the time axis, which will make it impossible to perform subsequent data assimilation processing.

[0075] In order to solve the problem of inconsistent time tags caused by different sampling frequencies, a fast scanning algorithm can be used to align time series with different sampling frequencies. The purpose of the fast scanning algorithm is to align data series with different time tags into a unified time frame by intelligently resampling or interpolating.

[0076] Specifically, we first determine the longest and shortest time labels in all time series, as well as the minimum sampling interval, to set a benchmark for a unified time frame.

[0077] Then, for series with fewer sampling points, interpolation methods (such as linear interpolation, spline interpolation, etc.) are used to estimate the data values ​​of the missing time points. For series with denser sampling points, downsampling can be performed to reduce the number of data points so that they match a uniform time frame.

[0078] Finally, a fast scanning algorithm is applied to ensure that all sequences have data representation at the same time point, thus achieving temporal alignment.

[0079] Step S140: Perform data assimilation on the processed prediction results based on the processed actual observation results to obtain the target prediction results of the observation object.

[0080] Before starting data assimilation, we first need to clarify what are state quantities (variables predicted by the proxy model) and observation quantities (variables actually measured by the observed object). State quantities include process noise covariance, which helps to consider the randomness and uncertainty within the system during data assimilation.

[0081] For example, the observation equation can be expressed as:

[0082] v o =H(x o )+ε (3)

[0083] Among them, v o represents the observed quantity; x o Represents the observation state; H(x o ) represents the true value; ε represents the observation noise.

[0084] The above observation equation is a function that maps state variables to observation space, which is used to link the output of the proxy model with the actual observation data. Among them, the observation noise ε is a description of the uncertainty and error of the observation data in the real world, reflecting the random errors introduced by factors such as the observation equipment or environment. The observation noise is usually assumed to be Gaussian distributed with a mean of zero, indicating the random fluctuation of the observation data around the true value.

[0085] The data assimilation process also involves time series prediction. By using Γ and observations in the proxy model, the proxy model can obtain the optimal solution for the time series.

[0086] For example, for the time series t k ∈[0,T], the corresponding prediction result can be expressed as:

[0087] u k =Γu k-1 (4)

[0088] Among them, t k Indicates the time corresponding to the k index; [0, T] indicates the prediction interval; u k Indicates the system state value at this moment; u k-1 Indicates the status value at the previous moment.

[0089] Data assimilation is the integration of observed data and proxy model predictions through an optimization process. Now assume that the data assimilation process is u d ,u d The essence of this is the estimated system state u k and observation data v o Make a trade-off and adjust the parameters of the proxy model so that the output of the proxy model is closer to the actual observation value.

[0090] Exemplarily, the process of data assimilation can be expressed as:

[0091] u d =argmin u J(u) (5)

[0092] Among them, J(u) represents the cost function, which characterizes the state error and observation error. The solution to the appropriate cost function is also to minimize the difference between the numerical results and the measurements.

[0093] For example, the data assimilation algorithm can be a variational algorithm or a filtering algorithm. The essence of the variational algorithm is to minimize the functional to find the optimal solution. The filtering algorithm can be selected from Kalman filtering, extended Kalman filtering or ensemble Kalman filtering (Ensemble Kalman Filter, EnKF) and the like.

[0094] Since the ensemble Kalman filter has a good nonlinear effect and is faster than the particle filter, in this embodiment, the ensemble Kalman filter method is used for data assimilation. Within the defined time window, new observations are continuously assimilated into the model. Each time new observation data arrives, the state estimate is updated using the selected data assimilation algorithm to gradually improve the accuracy of the model prediction.

[0095] Specifically, the observations defined in the window [0,T] are assimilated into the time points. After assimilation, new observations are assimilated in a new time window [T,2T]. Each process starts from the end point of the previous process and assimilates the data calculated by EnKF. EnKF assumes that both the state distribution and the observation likelihood are Gaussian.

[0096] For example, Figure 3 A schematic diagram of a data assimilation process provided as an example in the present disclosure is shown in FIG. Figure 3 As shown in the figure, the proxy model curve shows the predicted state value before the model performs data assimilation, reflecting the prediction result based on the current proxy model parameters and initial conditions. The assimilation prediction curve shows the optimized prediction value after the data assimilation process. The optimized prediction value is the result of optimization by integrating the actual observation data with the model prediction.

[0097] Among them, the state value on the agent model curve shows the predicted state of the agent model at a specific point in time.

[0098] Observations are observational data collected from actual observed objects and are used to calibrate and optimize the surrogate model.

[0099] The optimized observation value on the assimilated prediction curve represents the adjusted result after data assimilation.

[0100] In yet another embodiment of the present disclosure, it further includes:

[0101] Obtain a pre-constructed data assimilation model; the input data of the data assimilation model includes original input, observation input and static covariates, the original input includes data samples generated by the proxy model and the input data of the proxy model, and the observation input includes actual observation results;

[0102] The data assimilation model is used to perform feature fusion processing on the original input, observation input and static covariate to obtain the feature fusion result;

[0103] The feature fusion results are processed in time series to obtain the target prediction results of the observed object.

[0104] Exemplarily, the data assimilation model may adopt an improved Transformer model.

[0105] In practical applications, due to the locality of the receiving domain and the misalignment of timestamps, the marks formed on a single time step of multivariate time series often fail to display useful information. Therefore, before applying data assimilation, the sequence alignment method can be used to clear obstacles for the attention mechanism. In addition, in order to avoid weakening the advantages of the Transformer model in processing the time dimension, it is necessary to redefine the expression of basic sequences and multivariate correlations.

[0106] In the field of deep learning, time series prediction usually requires encoding the time series and generating labels. This process is driven by the underlying complex mechanism of the deep learning network. Each event sequence is labeled by a descriptive variable and acts through a self-attention mechanism, while using a feedforward neural network as a sequence representation. All past information is merged into a finite window k, and a sequence forecast for a period of time in the future is output. The specific representation of data assimilation using the improved Transformer model is as follows:

[0107] u φ (q,t,τ)=f φ (x i,t-k:t+τ ,y i,t-k:t ,v i,t-k:t ,τ) (6)

[0108] Among them, u φ (q, t, τ) represents the output time series; q represents the quantile {0.1, 0.5, 0.9}, which is assigned different error weights; t represents the input time node; τ represents the prediction time length; f φ (x i,t-k:t+τ ,y i,t-k:t ,v i,t-k:t ,τ) represents the transformer assimilation process: k is a finite window, and x is known i,t-k:t+τ Represents the input data from tk to t+τ, y i,t-k:t represents the output of the GAN proxy model from time tk to time t, v i,t-k:t Represents the observation value from time tk to time t, and predicts the model sequence of length τ.

[0109] For example, Figure 2 The data assimilation part in the data assimilation can specifically include:

[0110] The original input includes the data samples y generated by the proxy model i,t-k:t and input data x i,t-k:t+τ . Input data x i,t-k:t+τ represents the input data between time tk and time t+τ, where x i,t-k:t is the input data in the past time period, x i,t:t+τ is the input data in the future time period. Data sample yi,t-k:t represents the data sample output by the proxy model between time tk and time t. In the original input, y i,t-k:t and x i,t-k:t The set of can be used as historical input data, x i,t:t+τ can be used as input data for future time periods. i,t-k:t+τ and i,t-k:t The set is the original data input, which is input into the first encoder to encode the original input and extract effective features.

[0111] Observation input v i,t-k:t Represents the observation value between time tk and time t, which is input into another first encoder to embed the observation input and map the observation data into a high-dimensional space for subsequent processing.

[0112] In this embodiment, the first encoder is an LSTM (Long Short-Term Memory) encoder, and the second encoder is a static covariate encoder.

[0113] The static covariates are input into the static covariate encoder, which converts the static covariates into high-dimensional feature representations to form static feature vectors. Static covariates refer to variables that remain unchanged during observation or data collection. Since static covariates have the property of not changing over time, the encoded results provide additional contextual information that helps enhance the predictive power of the model.

[0114] Furthermore, the outputs of the two LSTM encoders are fused with the encoding results of the static covariates to integrate the dynamic and static feature information, obtain the feature fusion result, and strengthen the encoding of the context vector to regulate the temporal dynamics.

[0115] Feature fusion strategies can use vector concatenation, weighted sum, or nonlinear combination.

[0116] Then, the feature fusion results are processed sequentially and input into the enhancement layer, residual connection layer and layer normalization, multi-head attention mechanism, residual connection layer and normalization layer, and feedforward neural network in sequence.

[0117] Among them, the enhancement layer is used to perform feature extraction processing on the feature fusion result. The enhancement layer may include an input layer and a hidden layer. The number of hidden layers is one or more. The hidden layer is responsible for automatically extracting features from the input data. Each layer will perform a certain transformation on the output of the previous layer, so as to learn different levels of abstract representation of the data. The hidden layer close to the input layer can learn relatively simple and intuitive features. The deeper hidden layer can learn more complex and abstract features, which are a combination of shallow features and can represent higher-level concepts.

[0118] The feature extraction results of the enhancement layer are normalized through residual connection and layer normalization (Addition and Normalization, Add&Norm).

[0119] In the network structure of the transformer model, the enhancement layer enhances the model's expressiveness by extracting complex features. In order to maintain the continuity of feature transfer and avoid the gradient vanishing problem, these features are often transferred through residual connections (Add). Subsequently, layer normalization is applied to improve the convergence speed of the network and the stability of the training process.

[0120] In practical applications, when variables with the same timestamp are normalized, it may be difficult to distinguish the characteristics of variables at different timestamps. To solve this problem, we can consider adjusting the normalization strategy to normalize each individual variable separately. This method is called "normalization at the individual variable level", which allows each variable to be normalized independently, thereby retaining the distinction between variables at different timestamps, while also taking advantage of the stability and convergence brought by normalization.

[0121] After residual connection and layer normalization (Add&Norm), the data is fed into the multi-head attention mechanism. The original input data is usually rich in historical information and features. The multi-head attention mechanism can effectively extract global and local features from these data and focus on different parts of the input data, thereby significantly improving the model's ability to understand complex data structures.

[0122] The core of the attention mechanism is to give greater weight to variables with higher correlation by calculating the correlation between input sequences.

[0123] The multi-head attention mechanism introduces multiple attention heads, allowing the model to capture the dependencies between different parts of the input sequence at the same time. In this way, variables with high correlation will receive greater weights. The attention score is performed by query (Q), key value (K), and feature value (V):

[0124]

[0125] Among them, Q represents query; K represents key value; V represents feature value; d k Represents the dimension of the key value K.

[0126] In practical applications, the dynamic features output by the LSTM encoder can be selected as the query (Q). These dynamic features represent the key information of the current time step and are used to calculate the attention weight. The key value (K) is composed of the dynamic features output by the LSTM encoder and the static features output by the static covariate encoder. It is used to match the query (Q) and determine the relevance of the query to all time step features. The selection of the feature value (V) is consistent with the key value (K), which contains dynamic and static features and is used for weighted summation to form the final output.

[0127] After the above features (Q, K, V) are input into the multi-head attention mechanism, the attention weights are calculated and the features are fused. In this way, the multi-head attention mechanism can process the input features in parallel, with each head focusing on different aspects of the data to extract richer feature representations.

[0128] After the output of the multi-head attention mechanism, it is processed again through residual connection and layer normalization (Add&Norm). This step not only maintains the continuity of information, but also helps to suppress the gradient disappearance or explosion problems that may occur during training, thereby enhancing the stability and performance of the model.

[0129] Then, the output of residual connection and layer normalization is fed into the feed-forward neural network (FFNN) for in-depth processing and extraction of advanced features. The feed-forward neural network consists of multiple fully connected layers and is combined with nonlinear activation functions to introduce necessary nonlinear characteristics, so that the network can capture and express more complex data patterns and features. The activation function can use ReLU (Rectified Linear Unit).

[0130] The feedforward neural network connects the multi-head attention mechanism and the output layer. During the data assimilation process, it processes and refines features to ensure that the model can extract useful information from the original input and the data processed by the attention mechanism. In addition, the ReLU activation function helps the model avoid the gradient vanishing problem while maintaining computational efficiency. By applying ReLU after each fully connected layer, the network can pass positive gradients while maintaining sparse activation, thereby accelerating the learning process and improving the convergence speed of the model.

[0131] The output layer is responsible for receiving the output from the feedforward neural network and converting it into a final prediction result, which represents a fully processed estimate of the system state. The design of the output layer depends on the application scenario, such as classification, regression, or other specific tasks.

[0132] Based on this, the Transformer model with optimized network structure greatly improves memory utilization and significantly enhances computational efficiency. Because its design allows highly parallel operations, the Transformer model can effectively process large amounts of data while optimizing response time. This parallelization capability not only improves data processing speed, but also ensures efficient performance of the model in real-time tasks, thereby better solving real-time problems.

[0133] One or more technical solutions provided in the exemplary embodiments of the present disclosure are trained by constructing a proxy model corresponding to the observed object and constructing a data set by collecting sample points. After the training is completed, the output results of the proxy model and the observation results of the observed object are checked for consistency and sequence alignment to ensure that the output of the proxy model matches the actual observation data. Finally, the improved Transformer model is used for data assimilation to optimize the performance of the proxy model and improve the real-time performance of data processing.

[0134] Therefore, the data assimilation method based on digital twins and virtual-reality verification provided in the exemplary embodiment of the present disclosure effectively solves the real-time problem of the data assimilation process through the optimized Transformer model, and overcomes the limitation of the prior art that the observation data cannot be effectively integrated.

[0135] The above mainly introduces the solutions provided by the exemplary embodiments of the present disclosure. It is understandable that in order to implement the above functions, the electronic device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should easily realize that, in combination with the units and algorithm steps of each example described in the embodiments disclosed herein, the present disclosure can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present disclosure.

[0136] The exemplary embodiments of the present disclosure can divide the functional units of the electronic device according to the above method examples. For example, each functional module can be divided according to each function, or two or more functions can be integrated into one processing module. The above integrated modules can be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the exemplary embodiments of the present disclosure is schematic and is only a logical function division. There may be other division methods in actual implementation.

[0137] In the case of dividing each functional module according to each function, an exemplary embodiment of the present disclosure provides a data assimilation device based on digital twins and virtual-reality verification. The data assimilation device based on digital twins and virtual-reality verification can be a server or a chip applied to a server. Figure 4 This is a schematic block diagram of the functional modules of a data assimilation device based on digital twin and virtual-real verification provided by an exemplary embodiment of the present disclosure. Figure 4 As shown, the data assimilation device 400 based on digital twin and virtual-real verification includes:

[0138] The data acquisition module 410 is used to construct a proxy model of the observed object and obtain the prediction results output by the proxy model and the actual observation results of the observed object;

[0139] The data processing module 420 is used to perform consistency check on the prediction result and the actual observation result to obtain the prediction result and the actual observation result that pass the check;

[0140] The data processing module 420 is further used to perform sequence alignment processing on the verified prediction results and actual observation results to obtain processed prediction results and actual observation results;

[0141] The data processing module 420 is further used to perform data assimilation on the processed prediction result based on the processed actual observation result to obtain the target prediction result of the observation object.

[0142] In another embodiment provided by the present disclosure, the data processing module 420 is also used to collect sample points and construct a training data set; the training data set includes sample points in the sample space and output responses corresponding to the sample points; the proxy model is constructed and trained to obtain an error value of the proxy model; the proxy model includes a generator and a discriminator, the generator is used to simulate real data to generate false samples, and the discriminator is used to distinguish the source of the input samples; when the error value meets a preset condition, the trained proxy model is obtained.

[0143] In another embodiment provided by the present disclosure, the data processing module 420 is also used to calculate the DTW distance between the prediction result and the actual observation result using dynamic time warping DTW; when the DTW distance is within a preset credibility interval, it is confirmed that the prediction result and the actual observation result have passed the consistency check.

[0144] In another embodiment provided by the present disclosure, the data processing module 420 is further used to determine the sampling frequencies corresponding to the prediction results and the actual observation results respectively; based on the sampling frequencies, interpolation or downsampling is used to perform sequence alignment processing on the prediction results and the actual observation results.

[0145] In another embodiment provided by the present disclosure, the data processing module 420 is also used to obtain a pre-constructed data assimilation model; the input data of the data assimilation model includes original input, observation input and static covariates, the original input includes data samples generated by the proxy model and the input data of the proxy model, and the observation input includes the actual observation result; using the data assimilation model, feature fusion processing is performed on the original input, observation input and static covariates to obtain a feature fusion result; and time series processing is performed on the feature fusion result to obtain a target prediction result of the observed object.

[0146] In another embodiment provided by the present disclosure, the data assimilation model includes a first encoder, a second encoder, a feature fusion module, a multi-head attention mechanism and a feedforward neural network; the first encoder is used to encode the original input and the observed input respectively; the second encoder is used to encode the static covariate; the feature fusion module performs feature fusion processing on the encoding results corresponding to the original input, the observed input and the static covariate respectively to obtain the feature fusion result; the multi-head attention mechanism is used to perform feature extraction on the feature fusion result in parallel to obtain multi-head feature information; the feedforward neural network obtains the target prediction result of the observed object based on the multi-head feature information.

[0147] In another embodiment provided by the present disclosure, the data assimilation process of the data assimilation model is:

[0148] u φ (q,t,τ)=f φ (x i,t-k:t+τ ,y i,t-k:t ,v i,t-k:t ,τ)

[0149] Among them, u φ (q, t, τ) represents the output time series; q represents the quantile {0.1, 0.5, 0.9}, which is assigned different error weights; t represents the input time node; τ represents the prediction time length; f φ (x i,t-k:t+τ ,y i,t-k:t ,v i,t-k:t ,τ) represents the transformer assimilation process: k is a finite window, and x is known i,t-k:t+τ Represents the input data from tk to t+τ, y i,t-k:t represents the output of the proxy model from time tk to time t, v i,t-k:t Represents the observation value from time tk to time t, and predicts the model sequence of length τ.

[0150] The exemplary embodiment of the present disclosure also provides an electronic device, comprising: at least one processor; and a memory connected to the at least one processor in communication. The memory stores a computer program that can be executed by the at least one processor, and the computer program is used to cause the electronic device to perform the method according to the embodiment of the present disclosure when executed by the at least one processor.

[0151] The exemplary embodiments of the present disclosure also provide a non-transitory computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor of a computer, is used to cause the computer to perform the method according to the embodiments of the present disclosure.

[0152] Figure 5 The block diagram of the electronic device provided as an example of the present disclosure, the block diagram of the electronic device 500 that can be used as the server or client of the present disclosure will now be described, which is an example of a hardware device that can be applied to various aspects of the present disclosure. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples, and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0153] like Figure 5 As shown, the electronic device 500 includes a computing unit 501, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 502 or a computer program loaded from a storage unit 508 into a random access memory (RAM) 503. In the RAM 503, various programs and data required for the operation of the electronic device 500 can also be stored. The computing unit 501, the ROM 502, and the RAM 503 are connected to each other via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0154] A plurality of components in the electronic device 500 are connected to the I / O interface 505, including: an input unit 506, an output unit 507, a storage unit 508, and a communication unit 509. The input unit 506 may be any type of device capable of inputting information to the electronic device 500, and the input unit 506 may receive input digital or character information, and generate key signal inputs related to user settings and / or function control of the electronic device. The output unit 507 may be any type of device capable of presenting information, and may include, but is not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 508 may include, but is not limited to, a disk, an optical disk. The communication unit 509 allows the electronic device 500 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and may include, but is not limited to, a modem, a network card, an infrared communication device, a wireless communication transceiver, and / or a chipset, such as a Bluetooth™ device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0155] The computing unit 501 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 501 performs the various methods and processes described above. Each of the methods described above may be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 508. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 500 via the ROM 502 and / or the communication unit 509.

[0156] Figure 6 This is a schematic diagram of an exemplary computer program product provided by the present disclosure. The exemplary embodiment of the present disclosure also provides a computer program product 600, including a computer program 601, wherein the computer program 601 is used to enable the computer to perform a method according to an embodiment of the present disclosure when executed by a processor of the computer.

[0157] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0158] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0159] As used in this disclosure, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and / or device (e.g., disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0160] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0161] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0162] A computer system may include clients and servers. Clients and servers are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship to each other.

[0163] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instruction is loaded and executed on a computer, the process or function described in the embodiment of the present disclosure is executed in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, a terminal, a user device or other programmable device. The computer program or instruction may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer program or instruction may be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server, data center, etc. that integrates one or more available media. The available medium may be a magnetic medium, for example, a floppy disk, a hard disk, a tape; it may also be an optical medium, for example, a digital video disc (DVD); it may also be a semiconductor medium, for example, a solid state drive (SSD).

[0164] Although the present disclosure has been described in conjunction with specific features and embodiments thereof, it is apparent that various modifications and combinations may be made thereto without departing from the spirit and scope of the present disclosure. Accordingly, this specification and the drawings are merely exemplary illustrations of the present disclosure as defined by the appended claims, and are deemed to have covered any and all modifications, variations, combinations or equivalents within the scope of the present disclosure. Obviously, those skilled in the art may make various modifications and variations to the present disclosure without departing from the spirit and scope of the present disclosure. Thus, if these modifications and variations of the present disclosure fall within the scope of the claims of the present disclosure and their equivalents, the present disclosure is also intended to include these modifications and variations.

Claims

1. A data assimilation method based on digital twins and virtual-real verification, characterized in that: The method comprises: Constructing a proxy model of the observed object, and obtaining the prediction result output by the proxy model and the actual observation result of the observed object; Performing consistency check on the prediction result and the actual observation result to obtain the prediction result and the actual observation result that pass the check; Perform sequence alignment processing on the predicted results and actual observation results that have passed the verification to obtain the processed predicted results and actual observation results; The processed prediction results are data assimilated based on the processed actual observation results to obtain the target prediction results of the observed object.

2. The method according to claim 1, characterized in that The step of constructing a proxy model of the observed object includes: Collecting sample points and constructing a training data set; the training data set includes sample points in a sample space and output responses corresponding to the sample points; Constructing the proxy model and training it to obtain the error value of the proxy model; the proxy model includes a generator and a discriminator, the generator is used to simulate real data to generate fake samples, and the discriminator is used to distinguish the source of the input samples; When the error value meets a preset condition, a trained proxy model is obtained.

3. The method according to claim 1, characterized in that The performing consistency check on the prediction result and the actual observation result includes: Calculating the DTW distance between the predicted result and the actual observed result using dynamic time warping (DTW); When the DTW distance is within a preset credibility interval, it is confirmed that the prediction result and the actual observation result pass the consistency check.

4. The method according to claim 1, characterized in that: The performing sequence alignment processing on the prediction result and the actual observation result comprises: Determine the sampling frequencies corresponding to the prediction result and the actual observation result respectively; Based on the sampling frequency, interpolation or downsampling is used to perform sequence alignment processing on the prediction result and the actual observation result.

5. The method according to claim 1, characterized in that The performing data assimilation on the processed prediction result based on the processed actual observation result comprises: Acquire a pre-constructed data assimilation model; the input data of the data assimilation model includes original input, observation input and static covariates, the original input includes data samples generated by the proxy model and input data of the proxy model, and the observation input includes the actual observation result; Using the data assimilation model, performing feature fusion processing on the original input, the observed input and the static covariate to obtain a feature fusion result; The feature fusion result is subjected to time series processing to obtain a target prediction result of the observed object.

6. The method according to claim 5, characterized in that The data assimilation model includes a first encoder, a second encoder, a feature fusion module, a multi-head attention mechanism and a feedforward neural network; the first encoder is used to encode the original input and the observed input respectively; the second encoder is used to encode the static covariate; the feature fusion module performs feature fusion processing on the encoding results corresponding to the original input, the observed input and the static covariate respectively to obtain the feature fusion result; The multi-head attention mechanism is used to perform feature extraction on the feature fusion results in parallel to obtain multi-head feature information; the feedforward neural network obtains the target prediction result of the observed object based on the multi-head feature information.

7. The method according to claim 5, characterized in that The data assimilation process of the data assimilation model is: u φ (q,t,τ)=f φ (x i,t-k:t+τ ,y i,t-k:t ,v i,t-k:t ,τ) Among them, u φ (q, t, τ) represents the output time series; q represents the quantile {0.1, 0.5, 0.9}, which is assigned different error weights; t represents the input time node; τ represents the prediction time length; f φ (x i,t-k:t+τ ,y i,t-k:t ,v i,t-k:t ,τ) represents the transformer assimilation process: k is a finite window, and x is known i,t-k:t+τ Represents the input data from tk to t+τ, y i,t-k:t represents the output of the proxy model from time tk to time t, v i,t-k:t Represents the observation value from tk to time t, and predicts the model sequence of length τ.

8. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the method of claim 1.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the method of claim 1 is implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the method of claim 1 is implemented.

Images