A method and system for restoring satellite status data through historical enhancement

Abstract

The present invention relates to the technical field of satellite data processing and discloses a method and system for restoring satellite status data through historical enhancement. The method comprises the following steps: calculating the similarity between current satellite data and multiple segments of historical data, selecting the most similar segments of historical data as matched historical data; designing a tensor fusion filling scheme, determining the similarity between current satellite data and historical data, assigning different weights to the current satellite data and matched neighbor data, fusing the current satellite data with the weighted neighbor data, training a data completion model, and outputting complete satellite data through the data completion model. The method effectively improves the accuracy of the completion model, adopts a subspace clustering quantization method, and utilizes a distance table to efficiently calculate the distance between sparse data and historical data. This method improves computational efficiency while maintaining high retrieval accuracy, significantly improving retrieval performance.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite data processing, and in particular to a method and system for restoring satellite status data through historical enhancement. Background Art

[0002] With the rapid development of space technology, satellite networks and satellite computing have shown great potential. Real-time monitoring of satellite status is crucial to ensuring stable operation and mission success. Ensuring the accuracy of satellite data and the reliability of system performance is the basis for advancing these fields. However, due to factors such as space environment interference (such as electromagnetic interference and cosmic rays), hardware failures, and energy limitations, satellite status data is often incomplete. The loss of satellite status monitoring data often leads to various problems:

[0003] (1) Resource management failure: The lack of satellite status data may hinder the accurate assessment of the satellite's remaining power, storage capacity, or other resource utilization, thereby affecting effective resource allocation and optimization, resulting in some missions not being completed as expected;

[0004] (2) Delayed anomaly diagnosis: In the absence of complete satellite status data, the ground control center may not be able to detect equipment failures or other anomalies in a timely manner and miss the optimal intervention window, which increases the risk of fault propagation and may even cause satellite failure.

[0005] Therefore, it is crucial to ensure the accuracy and completeness of satellite status monitoring data. In recent years, missing data completion models based on neural networks have attracted widespread attention due to their advantages in capturing complex data patterns. Existing methods are mainly divided into two categories: neural network models based on multidimensional tensor decomposition and completion frameworks based on generative diffusion models.

[0006] (1) Neural network models based on multidimensional tensor decomposition: The core idea of ​​this method is to extend the linear assumption of traditional tensor decomposition through neural networks and explicitly model nonlinear interactions in multidimensional data. NTC constructs a three-dimensional interaction graph through outer products and extracts features through three-dimensional convolutions, but its cubic parameters grow with the feature dimension, and there is a risk of overfitting in sparse data, as seen in sensor networks with limited node observations, which weakens global learning. NTF integrates LSTM and MLP for temporal and nonlinear capture, but due to the gradient propagation problem of LSTM and the parameter redundancy of MLP, NTF performs poorly in high-dimensional sparse sequences, especially in heterogeneous data. NTM adds a generalized CP decomposition layer and a tensorized MLP to achieve better performance through multi-layer interactions. However, in sparse data, its local attention conflicts with the global pattern, such as missing key nodes in social networks. CoSTCo uses CNN to compress the tensor dimension and apply low-rank constraints, and performs well in regular grid data. In heterogeneous sparse data, fixed kernels cannot adapt well, and low-rank constraints may distort the results in ultra-low density cases.

[0007] (2) Completion framework based on generative diffusion model: To overcome the limitations of tensor decomposition methods, recent studies have turned to generative diffusion models, which learn the distribution of missing data through probabilistic modeling. TimesNet uses Fourier transform to convert time series data into 2D tensors to capture periodicity. In sparse non-periodic data (such as emergency monitoring), its forced periodicity assumption may distort the signal. CSDI combines diffusion and conditional attention to generate completion through weighted sampling. In high-dimensional sparse scenarios, its bidirectional attention context vector maintenance leads to exponential growth of memory, and weight calculation cannot reflect the true correlation of discontinuous data. PriSTI adds geospatial priors to improve diffusion robustness, but in non-spatial sparse data such as text, the geocoding module distorts features by mapping non-spatial data to space, and its strict spatial check of multi-step diffusion limits diversity. DAIN selects key entities with influence functions for data improvement. However, its local importance evaluation (e.g., relying on gradient magnitude) may miss global structural nodes, thereby reducing the diversity of generated data.

[0008] Both types of methods face significant challenges when processing high-dimensional sparse data. Their core bottlenecks can be attributed to insufficient parameter efficiency, poor global pattern learning ability, and weak adaptability to heterogeneity. Summary of the Invention

[0009] The purpose of the present invention is to solve the above problems. A method and system for restoring satellite status data through historical enhancement are designed to solve the problem that the existing missing data completion methods have high missing rate data and long-term data missing, which makes it difficult for the completion model to learn accurate data features from limited information during training, thereby limiting the accuracy of the missing data completion results.

[0010] A first aspect of the present invention provides a method for recovering satellite status data by historical enhancement, the method comprising the following steps:

[0011] S1. Calculate the similarity between the current satellite data and multiple segments of historical data, and select the most similar segments of historical data as the matched historical data;

[0012] S2. Design a tensor fusion filling scheme to determine the similarity between the current satellite data and the historical data, assign different weights to the current satellite data and the matched neighbor data, fuse the current satellite data with the weighted neighbor data, train the data completion model, and output complete satellite data through the data completion model, where the neighbor data is the historical data obtained in step S1.

[0013] The specific implementation process of step S1 includes:

[0014] S1-1. Input historical data with missing values ​​into the missing data completion model to obtain complete historical data;

[0015] S1-2, set a sliding window with the same size as the current satellite data, start matching from the first time node of the complete historical data, and get Window data;

[0016] S1-3, using the subspace clustering quantization method, calculate the current satellite data and The similarity of the window data is measured, and the most similar satellite data is selected. Neighbor data.

[0017] Optionally, in a first implementation of the first aspect of the present invention, the specific implementation process of step S1-3 includes:

[0018] Divide each window data into subspaces, cluster the corresponding subspaces of all window data to construct a codebook, and then quantize the original space according to the codebook;

[0019] Build a distance table for the current satellite data, and quickly query the distance between the current satellite data and each window tensor according to the distance table, as a similarity measure, and select the closest distance. as neighbor data.

[0020] Optionally, in a second implementation of the first aspect of the present invention, dividing each window data into subspaces, clustering corresponding subspaces of all window data to construct a codebook, and then quantizing the original space according to the codebook includes:

[0021] Slice each high-dimensional window data and divide it into slices, and the time dimension of each slice is , the vacancies are filled with 0, and each slice is a subspace;

[0022] For each window data Perform K-means clustering on slices and get The first cluster center Cluster center slices, after After clustering, we finally get Cluster center slices Cluster centers are the codebooks;

[0023] For each window tensor slices, and compare them with The first cluster center The distance between the cluster center slices is calculated, and then the cluster center index corresponding to the nearest distance is recorded. Finally, each high-dimensional window data is converted into One-dimensional data consisting of index values.

[0024] Optionally, in a third implementation of the first aspect of the present invention, a distance table is constructed for the current satellite data, and the distance between the current satellite data and each window tensor is quickly queried according to the distance table as a similarity measurement standard, and the closest distance is selected from them. As neighbor data, including:

[0025] Perform subspace division on the current satellite data to obtain slices, and then calculate the distance between each slice and the slice of each cluster center respectively, and put the calculated distances into the corresponding positions of the distance table in turn;

[0026] For the distance calculation between the current satellite data and the data in each window, query the corresponding elements in the distance table and sum them;

[0027] Use the distance table to get the distance between the current satellite data and all window data, and select the one with the smallest distance. distance, this The window data corresponding to the distance is the neighbor data.

[0028] Optionally, in a fourth implementation of the first aspect of the present invention, the distance table includes OK Column, the distance table The row represents the current satellite data Slice and The cluster center The distance between the cluster center slices, the The column represents the current satellite data. Slices and The first cluster center The distance between cluster center slices.

[0029] Optionally, in a fifth implementation of the first aspect of the present invention, weights are assigned according to the distance between the neighbor data and the current satellite data, and the assigned weights are normalized so that they are distributed between.

[0030] A second aspect of the present invention provides a system for recovering satellite status data by historical enhancement, the system comprising:

[0031] The historical data matching module is used to calculate the similarity between the current satellite data and multiple historical data segments, and select the most similar historical data segments as the matched historical data: input the historical data with missing values ​​into the missing data completion model to obtain the complete historical data; set a sliding window of the same size as the current satellite data, start matching from the first time node of the complete historical data, and obtain Window data; use the subspace clustering quantization method to calculate the current satellite data and The similarity of the window data is measured, and the most similar satellite data is selected. Neighbor data;

[0032] The tensor fusion filling module is used to design a tensor fusion filling scheme, determine the similarity between the current satellite data and the historical data, assign different weights to the current satellite data and the matched neighbor data, fuse the current satellite data with the weighted neighbor data, train the data completion model, and output complete satellite data through the data completion model, where the neighbor data is the historical data obtained in step S1.

[0033] A third aspect of the present invention provides a device for recovering satellite status data through historical enhancement, the device comprising a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory to cause the device to perform each step of the method for recovering satellite status data through historical enhancement as described in any one of the above items.

[0034] A fourth aspect of the present invention provides a computer-readable storage medium having instructions stored thereon, which, when executed by a processor, implement the various steps of the method for recovering satellite status data through historical enhancement as described in any one of the above items.

[0035] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects:

[0036] (1) To address the challenges posed by the high missing rate and structural missing patterns in satellite data, which make existing data completion models difficult to apply, this paper proposes a universal model training method that augments the training dataset by finding the most similar data segments from historical data. This framework can be applied to any data completion model, effectively improving the accuracy of the completion model.

[0037] (2) In order to quickly match historical data similar to the current satellite data, the present invention adopts a subspace clustering quantization method and uses a distance table to efficiently calculate the distance between sparse data and historical data. While maintaining high retrieval accuracy, it improves the calculation efficiency and significantly improves the retrieval performance.

[0038] (3) In order to balance the importance of the current satellite data and the matched neighbor data in the historical data, the present invention designs a fusion completion strategy based on historical data enhancement, which defines weights according to the similarity with the current data, balances the importance of the matched neighbor data and the current data for model training, and thus further enhances the performance of the data completion model. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Various other advantages and benefits will become apparent to those skilled in the art by reading the following detailed description of the preferred embodiment.The accompanying drawings are only for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0040] Figure 1 The overall process of the historical data matching module and the tensor fusion filling module provided in the embodiment of the present invention;

[0041] Figure 2 A schematic diagram of the historical data matching module process provided by an embodiment of the present invention;

[0042] Figure 3 A schematic diagram of the historical data sliding window operation provided by an embodiment of the present invention;

[0043] Figure 4 A schematic diagram of a historical data matching process based on subspace clustering quantization provided by an embodiment of the present invention;

[0044] Figure 5A schematic diagram of subspace division in the early preparation stage provided by an embodiment of the present invention;

[0045] Figure 6 A schematic diagram of clustering codebook construction in the early preparation stage provided by an embodiment of the present invention;

[0046] Figure 7 A schematic diagram of quantization based on a codebook in the early preparation stage provided by an embodiment of the present invention;

[0047] Figure 8 The invention provides a method for constructing a distance representation scheme in the query phase;

[0048] Figure 9 A schematic diagram of querying distance using a distance table in the query phase provided by an embodiment of the present invention;

[0049] Figure 10 A schematic diagram of the structure of a device for restoring satellite status data through historical enhancement provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The terms "first," "second," "third," "fourth," and so forth (if any) in the description and claims of the present invention and in the accompanying drawings are used to distinguish similar items and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that shown or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, product, or device that includes a series of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, product, or device.

[0051] For ease of understanding, the specific process of an embodiment of the present invention is described below. This embodiment of the present invention provides a method for restoring satellite status data through historical enhancement. The method models satellite data as a three-dimensional tensor of sample, time node, and attribute. To address the high data missing rate and long-term data missing issues commonly encountered in real-world scenarios, a historical data matching module is designed to select multiple segments of historical data that best match the current satellite data to supplement the training data, thereby resolving the problem of insufficient training data during complete model training. The similarity between the current satellite data and multiple segments of historical data is calculated, and the most similar segments of historical data are selected as the matched historical data. A tensor fusion filling scheme is further designed, fully considering the similarity between the current satellite data and historical data. Different weights are assigned to the current satellite data and the matched historical data, respectively, to balance the importance of the current satellite data and historical data to model training. The method of this embodiment is applied to an existing complete data completion model to address the problem of low model training accuracy or even inability to train due to the high data missing rate of the existing complete model. Specifically, the following steps are included:

[0052] S1. Calculate the similarity between the current satellite data and multiple segments of historical data, and select the most similar segments of historical data as the matched historical data;

[0053] S2. Design a tensor fusion filling scheme to determine the similarity between the current satellite data and historical data, and assign different weights to the current satellite data and the matched neighbor data respectively. Fuse the current satellite data with the weighted neighbor data to train the data completion model.

[0054] In this embodiment, the existing trained missing data completion model is used to feed the historical data with missing values ​​into the model to obtain complete historical data; a sliding window of the same size as the current satellite data is set, and matching is started from the first time node of the complete historical data to obtain Window data; use the subspace clustering quantization method to calculate the current satellite data and The similarity of the window data is measured, and the most similar satellite data is selected. Neighbor data.

[0055] In this embodiment, the calculation formula for the completeness of historical data is: ;in For historical data, It can be any deep learning based completion model.

[0056] In this embodiment, the window data quantity The calculation formula is: ;in is the number of time nodes of historical data, is the sliding window size, is the sliding window step size.

[0057] In this embodiment, (1) the preliminary preparation stage: each window data is divided into subspaces, the corresponding subspaces of all window data are clustered to construct a codebook, and then the original space is quantized according to the codebook. (2) the query stage: a distance table is constructed for the current satellite data, and the distance between the current satellite data and each window tensor is quickly queried according to the distance table as a similarity measurement standard, and the closest distance is selected from them. as neighbor data.

[0058] In this embodiment, the specific implementation process of the preliminary preparation stage includes: (1) subspace division: slicing each high-dimensional window data, dividing it into slices, and the time dimension of each slice is , empty space Filling, each slice is a subspace; (2) Clustering to build codebook: For each window data, the first Perform K-means clustering on slices and get The first cluster center Cluster center slices. After clustering, we finally get Cluster center slices cluster centers, which is the codebook; (3) Codebook-based quantization: For each window tensor slices, we compare them with The first cluster center The distance between the cluster center slices is calculated, and then the cluster center index corresponding to the nearest distance is recorded. Finally, each high-dimensional window data is converted into One-dimensional data consisting of index values.

[0059] In this embodiment, the formula for constructing one-dimensional data based on codebook quantization is: ;in Representing one-dimensional data No. element, that is, The first window data The slice index of the closest cluster center for each slice cluster. Indicates the The first window data slices, Indicates the The first cluster center Cluster center slices.

[0060] In this embodiment, the specific implementation process of the query phase includes: (1) Constructing a distance table: We divide the current satellite data into subspaces and obtain Slices, and then calculate the distance between each slice and the slice of each cluster center respectively, and put the calculated distances into the corresponding positions of the distance table in turn; (2) Use the distance table to query the distance: For the distance calculation between the current satellite data and each window data, we only need to query the corresponding elements in the distance table and then sum them; (3) Determine the neighbor data: Use the distance table to get the distance between the current satellite data and all window data, and select the one with the smallest distance. distance, this The window data corresponding to the distance is the neighbor data.

[0061] In this embodiment, a distance table with C rows and P columns is constructed, where each element of the distance table is represented by express, The calculation formula is: ;in Indicates the current satellite data slices, Indicates the The first cluster center After calculating all the elements, we finally get OK The distance table of the column The row represents the current satellite data Slice and The cluster center The distance between the cluster center slices, the The column represents the current satellite data. Slices and The first cluster center The distance between cluster center slices.

[0062] In this embodiment, the distance query formula is: ;in Indicates the current satellite data and the The distance between window data, The distance table Rank The element corresponding to the column.

[0063] In this embodiment, (1) weights are assigned to the current satellite data in a predefined manner. ; (2) to Neighbor data assignment weight ;(3)The above The weights are normalized so that they are distributed in between.

[0064] In this embodiment, The formula for weight distribution of neighbor data is: ,in Indicates the distance between the nth neighbor data and the current satellite data.

[0065] In this embodiment, The formula for normalizing the weights is: ; ;in express The sum of weights.

[0066] In this embodiment, an existing, trained missing data completion model is used to input historical data with missing values ​​into the model to obtain complete historical data. This step repairs the original missing historical data, providing a complete data foundation for subsequent matching and ensuring the availability of historical data. The missing data completion model used here can be any deep learning-based completion model. Its purpose is to properly fill in missing values ​​in historical data so that the historical data more accurately reflects the actual satellite status. For example, if the satellite power value at a certain time point in the historical data is missing, processing it with the model can generate a reasonable estimate, making the data at that time point complete. A sliding window of the same size as the current satellite data is set, and matching begins at the first time point in the complete historical data. The number of window data is calculated based on the number of time points in the historical data, the sliding window size, and the sliding window step size. The specific calculation formula is: Number of window data = (Number of time points in the historical data - Sliding window size) ÷ Sliding window step size + 1, with the result rounded down. In this way, the complete historical data is divided into multiple windows of the same size as the current satellite data, facilitating subsequent searches for historical data segments similar to the current satellite data. Taking the current satellite data with 10 time nodes, a sliding window size of 5, and a sliding window step of 2 as an example, if the historical data has 20 time nodes, then the number of window data is (20-5) ÷ 2 + 1 = 8.5, and then rounded down to 8. Each window data contains 5 consecutive time node data, corresponding to the time scale of the current satellite data.

[0067] The following combination Figures 1 to 9 The present invention describes a fast and general framework for recovering satellite status data through historical augmentation.

[0068] like Figure 1The figure shows the overall process of the historical data matching module and the tensor fusion filling module, clearly presenting the entire process from historical data processing to data fusion filling, which facilitates understanding of the overall architecture and data flow of the invention, including:

[0069] (1) Historical data matching module: To address the satellite data matching problem in different missing data scenarios, a historical data matching module is designed to effectively extract similar data from historical satellite data for data enhancement;

[0070] (2) Tensor fusion filling module: Through the above method, we found Neighborhood data that matches the current satellite data. However, the similarity between each neighbor data and the current satellite data varies. If all data are treated equally, prediction accuracy will be limited. Intuitively, the training model needs to be set closer to the data characteristics at the current moment. Therefore, the present invention designs a tensor fusion filling module to assign different weights to the data, balancing the importance of the data, thereby further improving filling accuracy.

[0071] like Figure 2 As shown, the historical data matching module includes:

[0072] (1) Historical data pre-filling: This paper uses the existing missing data completion model to estimate the missing values ​​in the historical data to ensure its integrity;

[0073] (2) Historical data partitioning: A historical data partitioning method based on a sliding window is designed to divide the historical data into multiple window data with the same scale as the current data, ensuring that the historical data segments belonging to the same time period as the current satellite data can be identified;

[0074] (3) Fast historical data matching: This paper proposes a fast historical data matching method based on subspace clustering quantization. This method can effectively match the neighboring data that is most similar to the current satellite data from the window data. This method can not only minimize storage overhead but also significantly reduce computational costs.

[0075] The historical data pre-filling is specifically implemented by using a simple missing data filling model to estimate the missing values ​​in the historical data. The formula is: ;in For historical data, It can be any deep learning based completion model.

[0076] like Figure 3 As shown, the historical data division is specifically implemented as follows: given a complete historical data, which contains Time nodes, for the current satellite data, contain time nodes, among which Set a sliding window with the same size as the current satellite data, and slide the entire historical data along the dimension Divided into window data, the sliding window step size is , the calculation formula for the number of window data is: .

[0077] like Figure 4 As shown in the figure, historical data fast matching includes the following two stages:

[0078] (1) Preliminary preparation stage: This stage includes three parts: subspace partitioning, clustering to construct codebook, and quantization based on codebook;

[0079] (2) Query phase: This phase includes three parts: building a distance table, using the distance table to query the distance, and determining neighbor data.

[0080] The subspace partitioning is as follows: Figure 5 As shown, the specific implementation is as follows: Directly quantizing high-dimensional data will encounter the dimensionality curse, resulting in a sharp increase in computational and storage costs. Therefore, we propose to decompose the high-dimensional space into multiple low-dimensional subspaces, and quantize each subspace independently, which significantly reduces the amount of computation. The present invention slices the high-dimensional window data along the dimension Divided into slices, and the time dimension of each slice is , empty space Filled, each slice is a subspace.

[0081] For example, consider the 3D window data , whose size is 4×5×2. When we divide the subspace along the time dimension, we get three slices, namely , each slice has a size of 4×2×2. It is worth noting that the third slice Empty positions in the second dimension of are filled with elements 0.

[0082] Clustering constructs codebook, such as Figure 6 As shown, the specific implementation is: the present invention sequentially processes the first Perform K-means clustering on slices and get The first cluster center Cluster center slices. After clustering, we finally get Cluster center slices The cluster centers are the codebooks.

[0083] For example, consider 4 windows of data belonging to the space 4×5×2 , the 4 window data are divided into 3 slices, namely The goal is to cluster all slices of the 4-window tensor into a codebook consisting of 2 cluster centers. When , the first slice of the 4-window tensor Perform K-means clustering to obtain the first cluster center slice of the two cluster centers ,when Time and The same operation is performed. Finally, these cluster center slices are restored to the corresponding 2 cluster centers and ,in , .

[0084] According to codebook quantization, Figure 7 As shown, the specific implementation is: for each window tensor slices, we compare them with The first cluster center The distance between the cluster center slices is calculated, and then the cluster center index corresponding to the nearest distance is recorded. Finally, each high-dimensional window data is converted into One-dimensional data consisting of index values.

[0085] For example, in the clustering stage, two cluster centers are obtained: and Window data Corresponding to three slices .when When calculating the slice Slices with cluster centers and The distance between and .if ,but ;when When calculating the slice Slice with cluster center and The distance between and .if ,but ;when When we calculate the slice Slices with cluster centers and The distance between and .if ,but Finally, the multidimensional window tensor Transformed into one-dimensional data .

[0086] According to codebook quantization, Figure 7 As shown in , the specific implementation is: for each window tensor’s p-th subspace, we calculate its distance to the C cluster centers in the codebook. Using the index of the nearest neighbor cluster center, its mathematical expression is: ,in represents the pth slice of the dth window, Represents the pth slice of the cth cluster center. By calculating the distance between the pth slice of the current window tensor and the pth slice of different cluster centers in sequence, the nearest neighbor cluster center is found and its corresponding index value is assigned to .

[0087] Construct a distance table such as Figure 8 As shown, the specific implementation is: divide the current satellite tensor into P sub-tensors, and then calculate the distance between these sub-tensors and the centroid of each cluster center to obtain the distance table elements, where the mathematical expression of each element of the distance table is: .

[0088] Use the distance table to query the distance, such as Figure 9 As shown, the specific implementation is: given a distance table, to calculate the distance between the current satellite tensor and the nth window tensor, we only need to query the corresponding elements in the distance table and sum them. The mathematical expression for the distance calculation between the current satellite tensor and the dth window tensor is: ,in Indicates the distance table The element value corresponding to row p column, where Calculated using the previous formula.

[0089] Determine the neighbor data. In order to find the N most similar historical window tensors, we use the distance table to sequentially query the distances between the D window tensors and the current satellite tensor, and then select the smallest N values ​​from them. The window tensors corresponding to these minimum distances are the final similarity tensors.

[0090] Figure 10FIG. 1 is a schematic diagram of the structure of a device for recovering satellite status data through historical augmentation, provided by an embodiment of the present invention. The device 600 for recovering satellite status data through historical augmentation may vary significantly depending on configuration or performance. The device may include one or more central processing units (CPUs) 610 (e.g., one or more processors), a memory 620, and one or more storage media 630 (e.g., one or more mass storage devices) storing application programs 633 or data 632. The memory 620 and storage medium 630 may be either transient or persistent storage. The program stored in the storage medium 630 may include one or more modules (not shown), each of which may include a series of instructions for operating on the device 600 through historical augmentation. Furthermore, the processor 610 may be configured to communicate with the storage medium 630, so that the device 600 executes the series of instructions stored in the storage medium 630 to implement the method provided in the above embodiment.

[0091] The device 600 for recovering satellite status data through historical enhancement may further include one or more power supplies 640, one or more wired or wireless network interfaces 650, one or more input and output interfaces 660, and / or one or more operating devices 631, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 10 The structure of the device for recovering satellite status data through historical enhancement does not limit the computer device provided by the present invention, and may include more or fewer components than shown in the figure, or combine certain components, or arrange components differently.

[0092] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions. When the instructions are executed on a computer, the computer executes the steps of the method for recovering satellite status data through historical enhancement provided in the above embodiments.

[0093] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment, devices, and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0094] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0095] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for recovering satellite status data by historical enhancement, characterized in that: The method comprises the following steps: S1. Calculate the similarity between the current satellite data and multiple segments of historical data, and select the most similar segments of historical data as the matched historical data; S2. Design a tensor fusion filling scheme to determine the similarity between the current satellite data and the historical data, assign different weights to the current satellite data and the matched neighbor data, fuse the current satellite data with the weighted neighbor data, train a data completion model, and output complete satellite data through the data completion model, where the neighbor data is the historical data obtained in step S1; The specific implementation process of step S1 includes: S1-1. Input historical data with missing values ​​into the missing data completion model to obtain complete historical data; S1-2, set a sliding window with the same size as the current satellite data, start matching from the first time node of the complete historical data, and get Window data; S1-3, using the subspace clustering quantization method, calculate the current satellite data and The similarity of the window data is measured, and the most similar satellite data is selected. Neighbor data.

2. The method for recovering satellite status data by historical enhancement according to claim 1, characterized in that: The specific implementation process of step S1-3 includes: Divide each window data into subspaces, cluster the corresponding subspaces of all window data to construct a codebook, and then quantize the original space according to the codebook; Build a distance table for the current satellite data, and quickly query the distance between the current satellite data and each window tensor according to the distance table, as a similarity measure, and select the closest distance. as neighbor data.

3. The method for recovering satellite status data by historical enhancement according to claim 2, wherein: The subspace division of each window data is performed, the corresponding subspaces of all window data are clustered to construct a codebook, and then the original space is quantized according to the codebook, including: Slice each high-dimensional window data and divide it into slices, and the time dimension of each slice is , the vacancies are filled with 0, and each slice is a subspace; For each window data Perform K-means clustering on slices and get The first cluster center Cluster center slices, after After clustering, we finally get Cluster center slices Cluster centers are the codebooks; For each window tensor slices, and compare them with The first cluster center The distance between the cluster center slices is calculated, and then the cluster center index corresponding to the nearest distance is recorded. Finally, each high-dimensional window data is converted into One-dimensional data consisting of index values.

4. The method for recovering satellite status data by historical enhancement according to claim 3, characterized in that: The distance table is constructed for the current satellite data, and the distance between the current satellite data and each window tensor is quickly queried according to the distance table as a similarity measure, and the closest distance is selected from them. As neighbor data, including: Perform subspace division on the current satellite data to obtain slices, and then calculate the distance between each slice and the slice of each cluster center respectively, and put the calculated distances into the corresponding positions of the distance table in turn; For the distance calculation between the current satellite data and the data in each window, query the corresponding elements in the distance table and sum them; Use the distance table to get the distance between the current satellite data and all window data, and select the one with the smallest distance. distance, this The window data corresponding to the distance is the neighbor data.

5. The method for recovering satellite status data by historical enhancement according to claim 4, characterized in that: The distance table includes OK Column, the distance table The row represents the current satellite data Slice and The cluster center The distance between the cluster center slices, the The column represents the current satellite data. Slices and The first cluster center The distance between cluster center slices.

6. The method for recovering satellite status data by historical enhancement according to claim 1, wherein: The specific implementation process of step S2 includes: Assign weights based on the distance between neighbor data and current satellite data, and normalize the assigned weights so that they are distributed between.

7. A system for recovering satellite status data by historical augmentation, characterized in that: The system includes: The historical data matching module is used to calculate the similarity between the current satellite data and multiple historical data segments, and select the most similar historical data segments as the matched historical data: input the historical data with missing values ​​into the missing data completion model to obtain the complete historical data; set a sliding window of the same size as the current satellite data, start matching from the first time node of the complete historical data, and obtain Window data; use the subspace clustering quantization method to calculate the current satellite data and The similarity of the window data is measured, and the most similar satellite data is selected. Neighbor data; The tensor fusion filling module is used to design a tensor fusion filling scheme, determine the similarity between the current satellite data and historical data, assign different weights to the current satellite data and the matched neighbor data, fuse the current satellite data with the weighted neighbor data, train the data completion model, and output complete satellite data through the data completion model, where the neighbor data is the matched historical data.

8. A device for recovering satellite status data through historical enhancement, characterized in that: The device for recovering satellite status data through historical augmentation includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory to cause the device for recovering satellite status data through historical augmentation to perform each step of the method for recovering satellite status data through historical augmentation according to any one of claims 1 to 6.

9. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by a processor, the steps of the method for recovering satellite status data through historical enhancement as described in any one of claims 1 to 6 are implemented.

Images