Space-time data filling method, system and device based on guidance information and medium

Abstract

This invention discloses a spatiotemporal data imputation method, system, device, and medium based on guidance information, relating to the interdisciplinary fields of intelligent transportation applications and spatiotemporal data mining. The method includes interpolating the spatiotemporal data to be imputed with missing values, training a spatiotemporal dependency learning neural network based on the interpolation information, constructing guidance information required for spatiotemporal data imputation, and training a diffusion-based spatiotemporal imputation model using the guidance information and the spatiotemporal data to be imputed. The trained spatiotemporal imputation model can generate estimates of the missing parts in the spatiotemporal data to be imputed. Because the imputation process fully considers spatiotemporal feature correlations and geospatial relationships, and effectively models and utilizes guidance information to guide the spatiotemporal dependency learning process with noisy information, the above-mentioned spatiotemporal imputation model can achieve high imputation accuracy even when imputing multiple missing patterns in spatiotemporal data from different domains.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation application and spatiotemporal data mining, and in particular to a spatiotemporal data filling method and system based on guide information, an electronic device and a computer readable storage medium. BACKGROUND

[0002] With the vigorous development of the world economy, more and more fields have entered the stage of informatization construction, such as air quality monitoring, intelligent transportation systems, urban power monitoring, etc. Infrastructure such as monitoring stations and sensors ensures the full collection of data in time and space. Such data can be modeled as spatiotemporal data in a broad sense, which is used for spatiotemporal pattern analysis, anomaly detection, behavior prediction, etc. However, due to external factors such as sensor failure or human error, missing data in spatiotemporal data is a common problem, which can cause deviations in the analysis of the spatiotemporal pattern of the related scene, and thus affect the downstream tasks. Therefore, it is urgent to design a reasonable and accurate method to accurately fill in the spatiotemporal data in space and time.

[0003] In recent years, with the rapid development of computer science, more and more research has begun to focus on spatiotemporal filling, which challenges the extraction of useful spatiotemporal pattern information from incomplete data, the modeling of spatiotemporal correlation, and the filling of missing data. Among the existing methods that can be used for spatiotemporal data filling, a typical autoregressive method mainly uses time correlation to fill in missing values, which uses recurrent neural networks as the core and recursively updates the hidden state of the missing part through available observation data. However, this autoregressive filling method is inevitably affected by the error accumulation problem when the data is sparse. In recent years, the use of a deep generative model called diffusion model for filling has achieved excellent performance in the field of multivariate time series filling, but it still has two problems: first, it mixes noisy information and observation values in the model learning, which affects the learning of spatiotemporal dependence during the model training process, and lacks sufficient complete and correct spatiotemporal information to guide the mapping of noise distribution to the true distribution of missing values; second, existing filling methods do not fully consider spatial correlation, which should not be limited to geographical proximity or feature correlation. SUMMARY

[0004] The purpose of the present application is to provide a spatiotemporal data filling method and system based on guide information, an electronic device and a computer readable storage medium, to solve the problem of low prediction accuracy of traditional autoregressive methods and existing generation methods when facing highly sparse data.

[0005] To achieve the above purpose, the present application provides the following solutions.

[0006] In a first aspect, the present embodiment provides a spatiotemporal data filling method based on guide information, comprising:

[0007] Obtain the spatiotemporal data to be filled;

[0008] The spatiotemporal data to be filled is input into the trained spatiotemporal data filling model to obtain the missing values ​​of the spatiotemporal data to be filled, and the spatiotemporal data to be filled is filled according to the missing values.

[0009] The training process of the spatiotemporal data imputation model is as follows:

[0010] Sample data is determined based on the original spatiotemporal data, and the sample data is divided to obtain first sample data and second sample data; the original spatiotemporal data is spatiotemporal sample data with missing values ​​to be filled; the sample data is data obtained by occluding the processed original spatiotemporal data.

[0011] The first sample data is processed to obtain interpolation information;

[0012] By utilizing the temporal global feature association, spatial global feature association, and geospatial relationship in the interpolation information, guiding information required for training the spatiotemporal data filling model is constructed.

[0013] The observations in the second sample data are deleted to obtain training data, and noise information is added to the unfilled part of the training data through the diffusion process in the conditional diffusion model to obtain sample data with noisy information.

[0014] The interpolation information and the noisy sample data are concatenated using tensors to obtain the input data for the inverse process.

[0015] The spatiotemporal dependency learning neural network is trained based on the guidance information, the inverse process input data, and the noise information to obtain the trained spatiotemporal data filling model.

[0016] Secondly, the present invention provides a spatiotemporal data completion system based on guidance information, comprising:

[0017] The data acquisition module is used to acquire the spatiotemporal data to be filled.

[0018] The missing value imputation module is used to input the spatiotemporal data to be imputed into the trained spatiotemporal data imputation model, obtain the missing values ​​of the spatiotemporal data to be imputed, and impute the spatiotemporal data to be imputed according to the missing values;

[0019] The training process of the spatiotemporal data imputation model is as follows:

[0020] Sample data is determined based on the original spatiotemporal data, and the sample data is divided to obtain first sample data and second sample data; the original spatiotemporal data is spatiotemporal sample data with missing values ​​to be filled; the sample data is data obtained by occluding the processed original spatiotemporal data.

[0021] The first sample data is processed to obtain interpolation information;

[0022] By utilizing the temporal global feature association, spatial global feature association, and geospatial relationship in the interpolation information, guiding information required for training the spatiotemporal data filling model is constructed.

[0023] The observations in the second sample data are deleted to obtain training data, and noise information is added to the unfilled part of the training data through the diffusion process in the conditional diffusion model to obtain sample data with noisy information.

[0024] The interpolation information and the noisy sample data are concatenated using tensors to obtain the input data for the inverse process.

[0025] The spatiotemporal dependency learning neural network is trained based on the guidance information, the inverse process input data, and the noise information to obtain the trained spatiotemporal data filling model.

[0026] Thirdly, the present invention provides an electronic device including a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the spatiotemporal data filling method based on boot information according to the first aspect.

[0027] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the spatiotemporal data filling method based on boot information as described in the first aspect.

[0028] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0029] This invention can fill in multiple missing patterns in spatiotemporal data from different domains. It interpolates the spatiotemporal data containing missing values, trains a spatiotemporal dependency learning neural network based on the interpolation information, constructs guiding information required for spatiotemporal data filling, and trains a diffusion-based spatiotemporal filling model using the guiding information and the spatiotemporal data to be filled. The trained spatiotemporal filling model can generate estimates of the missing parts in the spatiotemporal data to be filled. Because the filling process fully considers spatiotemporal feature correlations and geospatial relationships, and effectively models and utilizes guiding information to guide the spatiotemporal dependency learning process with noisy information, the above-mentioned spatiotemporal filling model can achieve high filling accuracy even when facing multiple missing patterns in spatiotemporal data from different domains. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 The flowchart of the spatiotemporal data filling method based on guidance information provided by the present invention is shown below.

[0032] Figure 2 This is a schematic diagram illustrating the principle of constructing a spatiotemporal data filling model based on a conditional diffusion model, as described in this invention.

[0033] Figure 3 This is a schematic diagram of the spatiotemporal dependency learning module of the present invention;

[0034] Figure 4 This invention provides a structural diagram of a spatiotemporal data filling system based on guidance information. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Example 1

[0038] Figure 1 The flowchart of the spatiotemporal data filling method based on guidance information provided by the present invention is as follows: Figure 1 As shown in the figure, this embodiment provides a spatiotemporal data imputation method based on guidance information, including:

[0039] Step 100: Obtain the spatiotemporal data to be filled;

[0040] Step 200: Input the spatiotemporal data to be filled into the trained spatiotemporal data filling model to obtain the missing values ​​of the spatiotemporal data to be filled, and fill in the missing values ​​of the spatiotemporal data to be filled.

[0041] The training process of the spatiotemporal data imputation model is as follows:

[0042] Sample data is determined based on the original spatiotemporal data, and the sample data is divided to obtain first sample data and second sample data; the original spatiotemporal data is spatiotemporal sample data with missing values ​​to be filled; the sample data is data obtained by occluding the processed original spatiotemporal data.

[0043] The first sample data is processed to obtain interpolation information.

[0044] By utilizing the temporal global feature association, spatial global feature association, and geospatial relationship in the interpolation information, the guiding information required for training the spatiotemporal data filling model is constructed.

[0045] The observations in the second sample data are deleted to obtain training data, and noise information is added to the unfilled part of the training data through the diffusion process in the conditional diffusion model to obtain sample data with noisy information.

[0046] The interpolation information and the noisy sample data are concatenated using tensors to obtain the input data for the reverse process.

[0047] The spatiotemporal dependency learning neural network is trained based on the guidance information, the inverse process input data, and the noise information to obtain the trained spatiotemporal data filling model.

[0048] Furthermore, the step of determining sample data based on the original spatiotemporal data specifically includes:

[0049] The raw spatiotemporal data was organized into a shape as follows: N x d The continuous time series values ​​of multiple nodes are used to obtain the processed spatiotemporal data to be filled. Then, the processed spatiotemporal data to be filled is manually masked to obtain sample data.

[0050] Further, the processing of the first sample data to obtain interpolation information specifically includes:

[0051] Based on the processed spatiotemporal data to be filled, a structure is constructed to describe the shape of the missing patterns. N x d A 0-1 matrix is ​​generated, where 0 represents missing data and 1 represents no missing data. Then, based on the 0-1 matrix, a linear interpolation algorithm in the time dimension is used to interpolate the first sample data to generate interpolation information for spatiotemporal data filling model training.

[0052] Furthermore, the step of utilizing the temporal global feature association, spatial global feature association, and geospatial relationship in the interpolation information to construct the guiding information required for training the spatiotemporal data filling model specifically includes:

[0053] The interpolated information is converted into the original high-dimensional hidden state; the original high-dimensional hidden state is a... N x L x d Tensor data, in which, N This refers to the number of sensor or monitoring station nodes in the spatiotemporal data. L The length of the time window, d The dimension of the high-dimensional hidden state data.

[0054] Organize the original high-dimensional hidden state along the time dimension. L The distributed tensor data is fed into the temporal attention module to obtain a high-dimensional hidden state that aggregates global temporal features.

[0055] Organize the original high-dimensional hidden state along spatial dimensions N The distributed tensor data is fed into the spatial attention module to obtain a high-dimensional hidden state that aggregates global spatial features.

[0056] Organize the original high-dimensional hidden state along the time dimension. N Distributed tensor data is fed into a graph convolutional neural network to obtain a high-dimensional hidden state that aggregates geospatial relationships. Specifically, the interpolated information utilizes geographic information and a graph convolutional network model to learn the geospatial relationships within the data.

[0057] The high-dimensional hidden states of aggregated temporal global features, aggregated spatial global features, and aggregated geospatial correlation are aggregated through residual connection and standardization, and then fed into a multilayer perceptron to obtain the high-dimensional hidden states of aggregated spatiotemporal global features and geospatial correlation.

[0058] The high-dimensional hidden state, which aggregates spatiotemporal global features and correlates them with geospatial data, serves as the guiding information required for training the spatiotemporal data-filling model.

[0059] Furthermore, for the spatiotemporal data to be filled, in order to cope with various possible missing situations, during model training, some values ​​are deleted from the observations according to a specific masking strategy as the filling target in the training phase, and the remaining observations are used to model spatiotemporal pattern information. For the spatiotemporal filling model, noise information is added to the clean data distribution of the filling target through a diffusion process, and then denoising is performed based on the guidance information through an inverse process. For the inverse process, the filling target, Gaussian noise, and diffusion step are sampled in each training step, and the guidance information is constructed based on the remaining observations except for the filling target, realizing the conversion of noisy information to the true value of the missing part. For the conversion of noisy information to the true value of the missing part, the process of calculating the spatiotemporal attention score is separated from the process of learning the spatiotemporal distribution of the noise signal, so that the noise signal fully participates in the spatiotemporal dependency learning process and protects the effective information from the interference of the noise signal.

[0060] Furthermore, the step of training a spatiotemporal dependent learning neural network based on guidance information, inverse process input data, and noise information to obtain a trained spatiotemporal data imputation model specifically includes:

[0061] The temporal attention weights are calculated using the guidance information, and the temporal attention module is updated according to the temporal attention weights. Then, the input data of the reverse process is fed into the updated temporal attention module to obtain noisy information that aggregates global temporal features.

[0062] Spatial attention weights are calculated using guidance information, and the spatial attention module is updated based on these weights. Then, the noisy information of aggregated temporal global features is fed into the updated spatial attention module to obtain the noisy information of aggregated spatial global features.

[0063] Organize the input data of the reverse process along the time dimension N The distributed tensor data is fed into a graph convolutional neural network to obtain a high-dimensional hidden state that aggregates geospatial relationships.

[0064] The noisy information of aggregated temporal global features, the noisy information of aggregated spatial global features, and the high-dimensional hidden state of aggregated geospatial association are aggregated through residual connection and standardization, and then fed into a multilayer perceptron to obtain the noisy information of aggregated spatiotemporal global features and geospatial association.

[0065] The noisy information associated with spatiotemporal global features and geospatial features is used as the module input. The residual connections and skip connections are obtained through gating activation units. The residual connections of each layer are used as the input of the next layer. The skip connections of each layer are added together and passed through multiple multilayer perceptrons to output the predicted noise information.

[0066] The loss value is calculated based on the predicted noise information and the inserted noise information, the model parameters are updated, and the trained spatiotemporal filling model is obtained.

[0067] This invention is applicable to various spatiotemporal data imputation tasks, such as urban air quality PM2.5 values ​​and highway network vehicle speeds. The data to be imputed consists of spatiotemporal data with various missing patterns across different fields. By filtering, interpolating, and extracting features from the original spatiotemporal data, a deep generative model based on a conditional diffusion model can learn the global spatiotemporal dependencies of the spatiotemporal data and the geospatial associations between different sensors or monitoring stations, thereby accurately imputing spatiotemporal data with missing values.

[0068] Example 2

[0069] Figure 2 A schematic diagram illustrating the principle of constructing a spatiotemporal data imputation model based on a conditional diffusion model, as shown below. Figure 2As shown, the conditional diffusion model is a deep generative model that adds noise to sample data and learns a denoising process. It offers highly flexible modeling and can generate probability distributions for sample data, making it well-suited for spatiotemporal probability interpolation tasks. The workflow for building a spatiotemporal data imputation model based on the conditional diffusion model can be represented by the following steps:

[0070] Step 1.1: Process the original spatiotemporal data; the original spatiotemporal data is spatiotemporal data with missing values ​​to be filled; the processing procedure is as follows: organize the original spatiotemporal data into a shape of... N x d The continuous time series values ​​of multiple nodes are used to obtain the processed spatiotemporal data to be imputed. Based on the processed spatiotemporal data to be imputed, a model is constructed to describe the shape of the missing pattern. N x d A 0-1 matrix, where 0 represents missing data and 1 represents no missing data.

[0071] Step 1.2: Manually mask the processed spatiotemporal data to be filled to obtain the spatiotemporal data for model training, i.e., sample data, for the model to learn fully; the method of manual masking depends on the type and missing pattern of the data, including random masking, continuous masking, mixed masking and other modes.

[0072] Step 1.3: Divide the sample data, use one part of the sample data to determine interpolation information, and delete the observations of the other part of the sample data to obtain the processed sample data, i.e., the training data. The deleted observations are used as the filling target in the training stage. Noise information is added to the part to be filled in the training data through the diffusion process in the conditional diffusion model to obtain sample data with noisy information. Specifically, the diffusion process is a Markov process, and the added noise information conforms to a Gaussian random distribution. In each Markov step, a quantitative amount of noise information is sampled according to the pre-set hyperparameters, and finally the sample is approximately a random normal distribution.

[0073] Step 1.4: Use a linear interpolation algorithm in the time dimension to interpolate a portion of the sample data to generate interpolation information for spatiotemporal data filling model training; specifically, for a portion of the sample data, if there are observations before and after the missing part in the current time window, then fill the missing value using linear interpolation based on the observation and time step; if there is no observation at one end of the missing part, then fill the missing part with the existing observation.

[0074] Step 1.5: Utilize the temporal global feature association, spatial global feature association, and geospatial relationship in the interpolation information to construct the guiding information required for training the spatiotemporal data filling model.

[0075] use Figure 3The spatiotemporal dependency learning module described herein extracts spatiotemporal features to obtain the guiding information required to fill the model with training spatiotemporal data, including: learning the temporal global feature association of interpolation information through a temporal attention model, and learning the spatial feature association of interpolation information through a spatial attention model; and for the interpolation information, using geographic information, learning the geographic spatial relationship of the data through a graph convolutional network model.

[0076] The construction steps of the spatiotemporal dependency learning module are as follows:

[0077] Step A: The spatiotemporal data to be filled is continuously updated in the model as a high-dimensional hidden state, and organized into a... N x L x d tensor H ,in N This refers to the number of sensor or monitoring station nodes in the spatiotemporal data. L The length of the time window, d Let be the dimension of the high-dimensional hidden state, and let be the model degree after training.

[0078] Step B: Organize the high-dimensional hidden state along the time dimension L The distributed tensor is fed into the temporal attention module to obtain a high-dimensional hidden state that aggregates global temporal features.

[0079] Step C: Organize the high-dimensional hidden state along the spatial dimensions N The distributed tensor is fed into the spatial attention module to obtain a high-dimensional hidden state that aggregates global spatial features.

[0080] Step D: Organize the high-dimensional hidden state along the time dimension N The distributed tensor is fed into a graph convolutional neural network to obtain a high-dimensional hidden state that aggregates geospatial relationships.

[0081] Step E: The high-dimensional hidden state, which aggregates temporal global features, spatial global features, and geospatial correlations, is aggregated through residual connections and normalization, and then fed into a multilayer perceptron to obtain a hidden state that aggregates spatiotemporal global features and geospatial correlations.

[0082] Step 1.6: Concatenate the interpolation information and the noisy sample data into a tensor to serve as the module input data.

[0083] Step 1.7: Train the spatiotemporal dependency learning neural network based on the guidance information and module input data to obtain the trained spatiotemporal data imputation model, specifically including:

[0084] Denoising is performed based on the guiding information through the inverse process. Specifically, for the inverse process, the target filling, Gaussian noise and diffusion steps are sampled in each training step, and the guiding information is constructed based on the remaining observations except for the target filling. This realizes the transformation of noisy information into the true value of the missing part, and realizes the prediction task of input data into normally distributed noise.

[0085] The task of predicting normally distributed noise using guided information is to separate the process of calculating the spatiotemporal attention score from the process of learning the spatiotemporal distribution of the noise signal, so that the noise signal can fully participate in the spatiotemporal dependent learning process while protecting the effective information from interference by the noise signal. The specific steps are as follows:

[0086] Step (1): Calculate the temporal attention weights using the guiding information, and obtain the noisy information that aggregates the global temporal features through step B.

[0087] Step (2): The noisy information that aggregates the global temporal features obtained in step (2) is used as the module input. The spatial attention weight is calculated using the guidance information. The noisy information that aggregates the global spatial features is obtained through step C. The noisy information that aggregates the global temporal features and geospatial relationships is obtained through steps D and E.

[0088] Step (3): The noisy information obtained in step (2) that aggregates spatiotemporal global features and geospatial relationships is used as module input, and residual connections and skip connections are obtained through gated activation units;

[0089] Step (4): Stack steps (1)-(3) into multiple layers. The residual connections of each layer are used as the input to the next layer of step (1). The skip connections of each layer are added together and passed through multiple multilayer perceptrons. The output is the predicted noise information. Calculate the loss value based on the predicted noise information and the inserted noise information, update the model parameters, and obtain the trained spatiotemporal filling model.

[0090] The trained spatiotemporal imputation model is used to generate estimates of the missing parts in the spatiotemporal data to be imputed, as follows:

[0091] Step 1): For the trained spatiotemporal imputation model, the missing data pattern is known, and its imputation target is set to all missing values ​​in the spatiotemporal data.

[0092] Step 2): Construct interpolation information and guiding information based on all observations of the data to be filled. Through the inverse process described above, use a multi-step Markov process to generate the true values ​​of the spatiotemporal data of the missing part based on random Gaussian noise. Specifically, for the data to be filled, first construct the corresponding interpolation information, and then construct the corresponding guiding information based on the interpolation information; the initial part to be filled is random Gaussian noise, and then through the inverse process described above, the random Gaussian noise is gradually transformed into the true values ​​of the missing values ​​through a multi-step Markov process.

[0093] This invention is based on computer science and various machine learning algorithms. Its implementation requires a certain level of programming, machine learning, and deep learning knowledge. This invention is implemented using the Python programming language and the open-source machine learning library PyTorch. For verification... Figure 2 and Figure 3 The spatiotemporal data imputation model based on guidance information, as shown, was tested using air quality data and highway vehicle speed data collected in a certain city. To evaluate the model's imputation capability, three evaluation metrics were used: Mean Absolute Error (MAE), Root Mean Square Absolute Error (RMSE), and Continuous Probability Distribution Score (CRPS). MAE and RMSE measure the error between the estimated and true values ​​of the data to be imputed, while CRPS calculates the compatibility of the probability distribution of the imputed result with the true values ​​of the data to be imputed. To verify that this invention performs better than other models, comparative experiments were conducted using other commonly used prediction models such as Vector Autoregression, Linear Interpolation, and K-Nearest Neighbors, as well as several advanced deep learning imputation models from recent years. The experiments demonstrate that the imputation error of this invention is the lowest.

[0094] Example 3

[0095] In order to implement the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, a spatiotemporal data filling system based on guidance information is provided below.

[0096] like Figure 4 As shown, the system includes:

[0097] Data acquisition module 1 is used to acquire the spatiotemporal data to be filled;

[0098] Missing value completion module 2 is used to input the spatiotemporal data to be filled into the trained spatiotemporal data completion model, obtain the missing values ​​of the spatiotemporal data to be filled, and fill in the missing values ​​of the spatiotemporal data to be filled.

[0099] The training process of the spatiotemporal data imputation model is as follows:

[0100] Sample data is determined based on the original spatiotemporal data, and the sample data is divided to obtain first sample data and second sample data; the original spatiotemporal data is spatiotemporal sample data with missing values ​​to be filled; the sample data is data obtained by occluding the processed original spatiotemporal data.

[0101] The first sample data is processed to obtain interpolation information;

[0102] By utilizing the temporal global feature association, spatial global feature association, and geospatial relationship in the interpolation information, guiding information required for training the spatiotemporal data filling model is constructed.

[0103] The observations in the second sample data are deleted to obtain training data, and noise information is added to the unfilled part of the training data through the diffusion process in the conditional diffusion model to obtain sample data with noisy information.

[0104] The interpolation information and the noisy sample data are concatenated using tensors to obtain the input data for the inverse process.

[0105] The spatiotemporal dependency learning neural network is trained based on the guidance information, the inverse process input data, and the noise information to obtain the trained spatiotemporal data filling model.

[0106] Example 4

[0107] This invention provides an electronic device including a memory and a processor. The memory stores a computer program, and the processor runs the computer program to enable the electronic device to perform a spatiotemporal data filling method based on boot information according to Embodiment 1.

[0108] Alternatively, the aforementioned electronic device may be a server.

[0109] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements a spatiotemporal data filling method based on boot information according to Embodiment 1.

[0110] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0111] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A spatiotemporal data imputation method based on guidance information, characterized in that, include: Obtain the spatiotemporal data to be filled; The spatiotemporal data to be filled is input into the trained spatiotemporal data filling model to obtain the missing values ​​of the spatiotemporal data to be filled, and the spatiotemporal data to be filled is filled according to the missing values. The training process of the spatiotemporal data imputation model is as follows: Sample data is determined based on the original spatiotemporal data, and the sample data is divided to obtain first sample data and second sample data; the original spatiotemporal data is spatiotemporal sample data with missing values ​​to be filled; the sample data is data obtained by occluding the processed original spatiotemporal data. The first sample data is processed to obtain interpolation information, specifically including: constructing a 0-1 matrix of shape N×d to describe the missing pattern based on the processed spatiotemporal data to be filled, where 0 represents missing data and 1 represents no missing data; and using a linear interpolation algorithm in the time dimension to interpolate the first sample data based on the 0-1 matrix to generate interpolation information for training the spatiotemporal data filling model. By utilizing the temporal global feature correlation, spatial global feature correlation, and geospatial relationship in the interpolation information, guiding information required for training the spatiotemporal data filling model is constructed. The observations in the second sample data are deleted to obtain training data, and noise information is added to the unfilled part of the training data through the diffusion process in the conditional diffusion model to obtain sample data with noisy information. The interpolation information and the noisy sample data are concatenated using tensors to obtain the input data for the inverse process. The spatiotemporal dependency learning neural network is trained based on the guidance information, the inverse process input data, and the noise information to obtain the trained spatiotemporal data filling model.

2. The spatiotemporal data imputation method based on guidance information according to claim 1, characterized in that, The process of determining sample data based on original spatiotemporal data specifically includes: The original spatiotemporal data is organized into continuous time series values ​​of multiple nodes with shape N×d to obtain the processed spatiotemporal data to be filled. The processed spatiotemporal data to be filled was manually masked to obtain sample data.

3. The spatiotemporal data imputation method based on guidance information according to claim 1, characterized in that, The process of constructing guiding information for training the spatiotemporal data filling model by utilizing the temporal global feature correlation, spatial global feature correlation, and geospatial relationships in the interpolation information specifically includes: The interpolated information is converted into the original high-dimensional hidden state; the original high-dimensional hidden state is an N×L×d tensor data, where N is the number of sensor or monitoring station nodes in the spatiotemporal data, L is the length of the time window, and d is the dimension of the high-dimensional hidden state data. The original high-dimensional hidden state is organized into tensor data distributed along the time dimension L, and fed into the time attention module to obtain a high-dimensional hidden state that aggregates global time features. The original high-dimensional hidden state is organized into tensor data distributed along the spatial dimension N, and fed into the spatial attention module to obtain a high-dimensional hidden state that aggregates global spatial features. The original high-dimensional hidden state is organized into tensor data distributed along the time dimension N, and fed into a graph convolutional neural network to obtain a high-dimensional hidden state that aggregates geospatial associations. The high-dimensional hidden states of aggregated temporal global features, aggregated spatial global features, and aggregated geospatial associations are aggregated through residual connection and standardization, and then fed into a multilayer perceptron to obtain the high-dimensional hidden states of aggregated spatiotemporal global features and geospatial associations. The high-dimensional hidden state, which aggregates spatiotemporal global features and correlates them with geospatial data, serves as the guiding information required for training the spatiotemporal data-filling model.

4. The spatiotemporal data imputation method based on guidance information according to claim 1, characterized in that, The step of training a spatiotemporal dependent learning neural network based on guidance information, reverse process input data, and noise information to obtain a trained spatiotemporal data imputation model specifically includes: The temporal attention weights are calculated using the guidance information, and the temporal attention module is updated according to the temporal attention weights. Then, the inverse process input data is fed into the updated temporal attention module to obtain noisy information that aggregates global temporal features. Spatial attention weights are calculated using guidance information, and the spatial attention module is updated based on the spatial attention weights. Then, the noisy information of aggregated temporal global features is fed into the updated spatial attention module to obtain the noisy information of aggregated spatial global features. The inverse process input data is organized into tensor data distributed along the time dimension N, and fed into a graph convolutional neural network to obtain a high-dimensional hidden state that aggregates geospatial associations. The noisy information of aggregated temporal global features, the noisy information of aggregated spatial global features, and the high-dimensional hidden state of aggregated geospatial association are aggregated through residual connection and standardization, and then fed into a multilayer perceptron to obtain the noisy information of aggregated spatiotemporal global features and geospatial association. The noisy information associated with spatiotemporal global features and geospatial features is used as module input, and residual connections and skip connections are obtained through gated activation units. The residual connections of each layer are used as input to the next layer, and the skip connections of each layer are added and passed through multiple multilayer perceptrons to output the predicted noise information. The loss value is calculated based on the predicted noise information and the inserted noise information, the model parameters are updated, and the trained spatiotemporal filling model is obtained.

5. A spatiotemporal data completion system based on guidance information, characterized in that, include: The data acquisition module is used to acquire the spatiotemporal data to be filled. The missing value imputation module is used to input the spatiotemporal data to be imputed into the trained spatiotemporal data imputation model, obtain the missing values ​​of the spatiotemporal data to be imputed, and impute the spatiotemporal data to be imputed according to the missing values; The training process of the spatiotemporal data imputation model is as follows: Sample data is determined based on the original spatiotemporal data, and the sample data is divided to obtain first sample data and second sample data; the original spatiotemporal data is spatiotemporal sample data with missing values ​​to be filled; the sample data is data obtained by occluding the processed original spatiotemporal data. The first sample data is processed to obtain interpolation information, specifically including: constructing a 0-1 matrix of shape N×d to describe the missing pattern based on the processed spatiotemporal data to be filled, where 0 represents missing data and 1 represents no missing data; and using a linear interpolation algorithm in the time dimension to interpolate the first sample data based on the 0-1 matrix to generate interpolation information for training the spatiotemporal data filling model. By utilizing the temporal global feature correlation, spatial global feature correlation, and geospatial relationship in the interpolation information, guiding information required for training the spatiotemporal data filling model is constructed. The observations in the second sample data are deleted to obtain training data, and noise information is added to the unfilled part of the training data through the diffusion process in the conditional diffusion model to obtain sample data with noisy information. The interpolation information and the noisy sample data are concatenated using tensors to obtain the input data for the inverse process. The spatiotemporal dependency learning neural network is trained based on the guidance information, the inverse process input data, and the noise information to obtain the trained spatiotemporal data filling model.

6. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the spatiotemporal data filling method based on boot information according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the spatiotemporal data filling method based on boot information as described in any one of claims 1 to 4.

Images