Space environment data and navigation time-frequency field data correlation analysis method

Abstract

The application provides a space environment data and navigation time-frequency field data correlation analysis method, adopts a graph machine learning method, extracts space environment index abnormal moment data and navigation time-frequency field abnormal moment data respectively, and corresponds according to corresponding time. The application aims to predict navigation field data changes through space environment data, but because different data structures are not the same, simple mathematical relationships and regression learning cannot be used. Different structures of data are regarded as nodes, the nodes continue to contain sub-nodes, the relationship between nodes is learned, the above difficulties can be overcome and effective correlation rules and relationships can still be obtained. The model quantity of the application is small, the model applicability is wide, and the correlation between numbers is processed by a natural language processing method.

Description

Technical Field

[0001] This invention belongs to the field of data processing, and in particular relates to a method for correlation analysis of space environment data and navigation time and frequency domain data. Background Technology

[0002] As the breadth and depth of BeiDou applications continue to increase, satellite navigation systems are no longer limited to simply acquiring position, velocity, and time information. Satellite navigation signals are affected directly or indirectly by various factors, such as the ionosphere, troposphere, climate, and space environment. Therefore, monitoring and early warning of the satellite navigation system is of great significance for better serving the performance and accuracy of BeiDou navigation and positioning. This aims to improve the reliability of satellite navigation services by spatiotemporally correlating and fusing meteorological, hydrological, and space environment data with a large amount of navigation time-frequency monitoring data, learning from this data, and predicting the relevant performance of satellite navigation services.

[0003] Despite continuous technological advancements, their application in navigation remains relatively uncommon. Satellite navigation systems, however, have existed for decades, accumulating massive amounts of data. This includes not only China's BeiDou and the US GPS systems, but also the EU's Galileo, Russia's GLONASS, and other satellite-based augmentation systems, providing a foundation for big data learning. At the algorithmic level, numerous machine learning techniques, including clustering, regression, and neural network models, have been extensively applied in fields such as computer science.

[0004] Traditional correlation methods often rely on manual searching and labeling, which becomes increasingly difficult and inefficient when processing massive amounts of data, lacking a holistic perspective. Furthermore, different types of data often have different descriptive dimensions, increasing the number of judgment criteria and making the methods cumbersome and complex. In addition, there is currently no unified method for correlation analysis of both space environment data and navigation time-frequency domain data. With the continuous development of big data technology, accurately and efficiently solving these pain points and problems will become possible. Summary of the Invention

[0005] To address the aforementioned technical challenges and better serve satellite navigation positioning and services, this invention proposes a correlation analysis method between space environment data and navigation time-frequency domain data. This method involves learning the correlation patterns between historical space environment indices and navigation time-frequency domain data, and then predicting future navigation time-frequency data. The invention employs graph machine learning, first extracting data from anomaly times in both the space environment index and navigation time-frequency domains, and then mapping them according to their corresponding times. While the invention aims to predict changes in navigation domain data using space environment data, the different data structures prevent the use of simple mathematical relationships and regression analysis. By treating data with different structures as nodes, each containing child nodes, and learning the relationships between nodes, the aforementioned difficulties can be overcome while still obtaining effective correlation patterns and relationships.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for correlation analysis between space environment data and navigation time-frequency domain data includes the following steps:

[0008] Step 1: Establish the relationship between anomalous moments in space environment data and navigation time-frequency domain data;

[0009] Step 2: Learn the association patterns;

[0010] Step 3: Make a prediction.

[0011] Furthermore, in step 1, each pair of space environment data and navigation time-frequency domain data at the same time is established as a tuple, with time as the boundary;

[0012] Create N tuples in the manner described above, and set the distance between nodes to 1.

[0013] Further, step 2 includes:

[0014] The space environment data and navigation time-frequency domain data are jointly established in one model; after the nodes are formed, the theoretical probability of node a generating node b is:

[0015]

[0016] In the formula u b v a Let be the spatial embedding vectors of nodes b and a, respectively; k represents all nodes adjacent to node a; X represents the set of nodes adjacent to a; and the superscript T denotes matrix transpose. The actual calculated probability is:

[0017]

[0018] In the formula ε abis the edge weight of node a and node b, and X represents the set of nodes adjacent to a, which is obtained from the actual statistical tuple library. The learning process is to make the theoretical probability as close as possible to the actual probability.

[0019] Define the loss function G as:

[0020]

[0021] In the formula, KuL represents the KL divergence. The complement of the probability is represented by ||, which represents the logical operation "OR", X represents the set of nodes adjacent to a, and Y represents the set of nodes adjacent to b.

[0022] By using negative sampling to reduce computation, the loss function becomes F:

[0023]

[0024] In the formula, σ represents the sigmoid function, L represents the number of negative samples used, and v a u b u c All are embedding vectors, where c represents all nodes adjacent to a;

[0025] After learning is complete, the spatial vector representation of the node in the same mapping space is obtained. All spatial vector representations constitute the model jointly established by the space environment data and the navigation time and frequency domain data. During the first learning, the historical navigation time and frequency domain data and historical space environment data are correlated and learned. After the first model is generated, the correlation and learning is repeated on the real navigation time and frequency domain data and real space environment data input in real time to form a new model. The new model is then fused with the original model to complete the incremental model.

[0026] Further, step 3 includes:

[0027] The model represents the learning of nodes in the same space, so that the spatial vector representation of nodes can be stored and form correlation patterns. When future prediction data with the same format as space environment data is input into the model, its spatial distance with various navigation time and frequency domain data is calculated. The closer the distance, the more likely it is to occur or the greater the correlation.

[0028] Beneficial effects:

[0029] 1. The number of models in this invention is small: The data types involved in this invention theoretically require 5 × 9 = 45 models, but the method of this invention can use a single model for learning and prediction.

[0030] 2. The model of this invention has wide applicability: This invention involves a large number of quantities such as numbers and words, which are difficult to analyze and learn using algorithms such as regression. It regards each variable as a node and maps quantities with different units and measures to the same space. It has strong applicability and preserves the original relationship, making it easy to apply further.

[0031] 3. This invention uses natural language processing to process the relationships between numbers: traditional methods tend to use equations and other methods to describe the relationships between numbers, while natural language processing tends to process characters. Since time is an intermediate quantity, this invention uses natural language processing to process the relationships between numbers, which is a new perspective compared to traditional methods. Attached Figure Description

[0032] Figure 1 This is a flowchart of a method for correlation analysis between space environment data and navigation time-frequency domain data according to the present invention;

[0033] Figure 2 This is a diagram of the feature extraction process;

[0034] Figure 3 This is a two-dimensional diagram showing the learning results for the nodes. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0036] The space environment data involved in this invention includes:

[0037] Kp index: A 3-hour index derived from the Ks index (data from 13 geomagnetic observatories). It is the average of the disturbance levels in the horizontal two components of the three geomagnetic observations. The Kp index is expressed as: 0-2 calm, 2-3 disturbance, 3-4 active, 4-5 minor geomagnetic storm, 5-6.5 major geomagnetic storm, and above 6.5 severe geomagnetic storm;

[0038] AE Index: Polar substorm intensity index, measured in nT, measured once per minute;

[0039] Dst index: Magnetic storm loop current index, measured in nT, is obtained hourly from the average of the H-component disturbances of the geomagnetic field at four low-latitude magnetic observatories (Honolulu, San Juan, Hermanus, and Kashiwagi). The Dst index is essentially a measure of the longitudinally averaged ground disturbances at low-latitude magnetometer stations, thus indiscriminately measuring the effects of many ground and magnetospheric current systems.

[0040] SSN: Sunspot Count, no unit, measured once a day;

[0041] F10.7: The radio flux index of the Sun at a wavelength of 10.7 cm (2800 MHz), measured in sfu (solar flux units), and determined daily (by the intensity of solar radio emissions in a 100 MHz bandwidth).

[0042] The navigation time and frequency domain data involved in this invention includes:

[0043] Satellite segment data: broadcast orbit and satellite clock errors;

[0044] Space segment data: ionospheric grid data, tropospheric delay data;

[0045] Ground segment data: standard single-point positioning accuracy, signal integrity rate, effectiveness, carrier-to-noise ratio, pseudorange noise, and carrier noise.

[0046] like Figure 1 As shown, the correlation analysis method between space environment data and navigation time-frequency domain data of the present invention includes the following steps:

[0047] Step 1: Establish relationships between abnormal moments:

[0048] The above data are all composed of time and numerical values ​​(some data include location information). This invention uses time as a medium to establish a tuple for each pair of space environment data and navigation time-frequency domain data at the same time. Taking 2020 / 1 / 1, Kp index and broadcast orbit error as an example, it is expressed as: 2020 / 1 / 1, Kp value 3, broadcast orbit error 1.2m.

[0049] Create N tuples in the manner described above, and set the distance between nodes to 1.

[0050] Step 2: Learn the association patterns:

[0051] Theoretically, different associations in this invention require different model learning methods. Using the method of this invention, information from different dimensions can be mapped to the same space and represented by vectors of the same dimension. This allows the five types of space environment data and nine types of navigation time-frequency domain data from this invention to be jointly established in one model. After nodes are formed, the theoretical probability of node a generating node b is:

[0052]

[0053] In the formula u b v aLet be the spatial embedding vectors of nodes b and a, respectively; k represents all nodes adjacent to a; X represents the set of nodes adjacent to a; and the superscript T denotes matrix transpose. These are the theoretical generation probabilities, and the actual calculated probabilities are:

[0054]

[0055] In the formula ε ab Let X represent the edge weights of nodes a and b, and let X represent the set of nodes adjacent to a, derived from a database of actual statistical tuples. The learning process aims to make the theoretical probability as close as possible to the actual probability. This invention defines the loss function G as follows:

[0056]

[0057] In the formula, KuL represents the KL divergence. The complement of the probability is represented by ||, which represents the logical operation "OR", X represents the set of nodes adjacent to a, and Y represents the set of nodes adjacent to b.

[0058] Negative sampling can effectively reduce the amount of computation, and the loss function becomes F:

[0059]

[0060] In the formula, σ represents the sigmoid function, L represents the number of negative samples used, and v a u b u c All are embedding vectors, where c represents all nodes adjacent to a;

[0061] After learning, the vector representations of nodes in the same mapping space are obtained. All spatial vector representations constitute the model of this invention. Since the method used in this invention can perform incremental training without changing the original learning results, association learning can be performed on historical navigation time-frequency domain data and historical space environment data during the initial learning. Considering the huge amount of data, to reduce storage pressure, the model can delete the original data after calculating and saving the corresponding number of associations without having to save it. After generating the initial model, the association learning method of this invention can be repeated at any time on real-time input real navigation time-frequency domain data and real space environment data to form a new model, which is then fused with the original model to complete the incremental model. For both the fused model and the new model generated using real-time data, the similarity between nodes can be determined by calculating the Euclidean distance between vectors. However, the continuously incremental model will be more accurate and more meaningful for big data.

[0062] Step 3: Make a prediction:

[0063] The model learns to represent the aforementioned nodes in the same spatial space, allowing the vector representations of the nodes to be stored. It's worth noting that because the model learns the spatial representation of the nodes' mappings, it can perform incremental learning without relearning the mapping representations of existing data. For prediction, new space environment prediction data is input, and the model calculates the five closest spatially distributed data points for each type of data in the navigation time-frequency domain, representing these as the most likely occurrences. It's important to note that the more data learned, the more accurate the prediction results will be.

[0064] like Figure 2 The diagram shows the association pattern learning process of this invention, i.e. the feature extraction process. For related nodes (node ​​1 to node n-1), the core is to find the set theoretical probability that is closest to the actual probability. Figure 2 In this context, N1-N6 represent the relationships between nodes.

[0065] like Figure 3 As shown, compression technology is used to compress a 256-dimensional node vector into a 2-dimensional spatial display, where each point in the figure represents a node.

[0066] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for correlation analysis of space environment data and navigation time-frequency domain data, characterized in that, Comprise the following steps: Step 1, establish the abnormal time relationship between space environment data and navigation time-frequency field data, comprising: With time as the boundary, each pair of space environment data and navigation time-frequency field data at the same time is established into a group of tuples; According to the above method, N groups of tuples are established, and the distance between nodes is set to 1; Step 2, correlation rule learning, comprising: The space environment data and navigation time-frequency field data are jointly established in a model; After the formation of the node, the theoretical probability of node a generating node b is: where u b and v a are the spatial embedding vectors of nodes b and a, respectively, k denotes all nodes adjacent to node a, X denotes the set of adjacent nodes to a, and the superscript T denotes matrix transposition; the actual calculated probability is: where ε ab is the edge weight of node a and node b, X represents the set of adjacent nodes of a, which is obtained from the actual statistical tuple library, and the learning process is to make the theoretical probability close to the actual probability to the greatest extent. Define the loss function G as: where KuL represents the KL divergence, where P(X) represents the complement of the probability, || represents the logical operation "or", X represents the set of neighboring nodes of a, and Y represents the set of neighboring nodes of b. Use negative sampling to reduce the amount of calculation, then the loss function becomes F: where σ represents a sigmoid function, L represents the number of negative sampling adopted, v a , u b , u c are embedding vectors, and c represents all nodes adjacent to a. After learning, the spatial vector representation of the node in the same mapping space is obtained, and all the spatial vector representations constitute the model jointly established by the space environment data and the navigation time-frequency field data; At the first time of learning, the historical navigation time-frequency field data and the historical space environment data are associated and learned, after the first model is generated, the real-time input of real navigation time-frequency field data and real space environment data is repeatedly associated and learned to form a new model, and the new model is fused with the original model to complete the incremental model; Step 3, prediction, comprising: The model represents the learning of the node in the same space, so that the spatial vector representation of the node is stored and the correlation rule is formed; When the future prediction data in the same format as the space environment data is input into the model, the spatial distance between the future prediction data and each type of navigation time-frequency field data is calculated respectively; The closer the distance, the more likely it is to occur or the greater the correlation.

Images