Star Chart Recognition Method Based on the Fusion of Neural Network and Hausdorff Distance

Abstract

The present invention discloses a star map recognition method based on the fusion of neural network and Hausdorff distance, belonging to the fields of astronomical navigation, guidance and control. The method includes the following steps: constructing a navigation feature database and a neighboring star index library; building a self-organizing competitive neural network and integrating the calculation method of Hausdorff distance into neurons; training the neural network model; star map recognition, adopting the same feature extraction method for observed stars as that for navigation stars, extracting the feature vectors of observed stars and inputting them into the neural network to obtain the star map recognition result. The present invention combines the respective advantages of neural network and Hausdorff distance in the field of star map recognition. The feature information of navigation stars is complete and sufficient, and the matching process is simple. Under the influence of factors such as parameter deviation, star point position deviation, magnitude deviation, and false star magnitude interference in star sensors, it has strong robustness and a high recognition success rate.

Description

Technical Field

[0001] The present invention belongs to the field of celestial navigation, and particularly relates to a star map recognition method based on the fusion of neural network and Hausdorff distance. Background Art

[0002] During the assembly, processing, transportation and operation of a star sensor, it is prone to be interfered by various factors. During the assembly and processing of the star sensor, due to the process defects and installation deviations of components, no matter how precise the alignment is, there will always be deviations such as focal length, principal point, image plane tilt and distortion; during the launch with the carrier, even if the internal parameters to be on-orbit serviced have been calibrated on the ground, it will be interfered by factors such as vibration and deviate from the calibrated value; during the on-orbit operation of the star sensor, it will be interfered by stray light such as sunlight, moonlight and earth-reflected light, and its internal electronic devices will also age due to factors such as temperature and space radiation. The interference of these factors will lead to the reduction of the performance of the star sensor. Star map recognition is the premise and key for the accurate output of the attitude of a spacecraft. The performance of the star map recognition algorithm (including performance such as recognition accuracy, robustness and rapidity) largely determines the attitude output accuracy and stability of the star sensor. Therefore, further improving the autonomy and robustness of star map recognition is of great significance for enhancing the reliability of the star sensor.

[0003] With the continuous progress of space technology, various algorithms have been applied to star map recognition to enhance the autonomy and reliability of star sensors. Star map recognition algorithms can be roughly divided into three categories according to different feature types and matching methods: algorithms based on constellation features represented by the triangle algorithm, algorithms based on pattern recognition represented by the grid algorithm, and algorithms based on intelligent behavior represented by neural networks and genetic algorithms. Generally speaking, the robustness, reliability, autonomy and intelligence of star map recognition algorithms are showing an increasing trend.

[0004] The star chart recognition method based on the Hausdorff distance does not require a point-to-point correspondence relationship during feature matching and has good robustness for the recognition of star charts contaminated by noise and with large distortions. However, the traditional star chart recognition method based on the Hausdorff distance needs to traverse the star library, set relevant thresholds, and continuously access memory during operation, which to a certain extent affects the speed of star chart recognition. For the existing star chart recognition methods based on neural networks, most neural networks will first divide the navigation stars into many large categories, but still need to perform secondary matching, and the matching process is relatively complex. In addition, for the existing star chart recognition methods based on neural networks, during feature matching, only the point-to-point correspondence relationship of feature elements can be used for calculation. Especially when the adjacent stars are interfered and have large position deviations, the feature elements of the feature vectors of the relevant navigation stars often change greatly and cannot be correctly recognized. Considering the respective advantages and disadvantages of neural networks and the Hausdorff distance in the field of star chart recognition, the present invention proposes a star chart recognition method based on the fusion of neural networks and partial Hausdorff distance. This method has strong robustness and a high recognition success rate under the influence of factors such as parameter deviation in the star sensor, star point position deviation, star magnitude deviation, and false star magnitude interference. Summary of the Invention

[0005] The purpose of the present invention is to overcome the above-mentioned deficiencies of the prior art and provide a star chart recognition method based on the fusion of neural networks and partial Hausdorff distance. This method fully complements the advantages and disadvantages of the Hausdorff distance and neural networks in star chart recognition, can still achieve correct recognition under the condition that there are certain errors in the observed star feature patterns, the recognition process is fast, simple, has a high recognition rate, and has strong robustness to factors such as parameter deviation in the star sensor, star point position deviation, star magnitude deviation, and pseudo-star magnitude interference.

[0006] A star chart recognition method based on the fusion of neural networks and the Hausdorff distance includes the following steps:

[0007] 1. Construct a navigation star feature database

[0008] 1.1 Select the navigation stars that meet the requirements from the SAO star catalog, regard the star pairs with an angular distance less than a certain angular distance threshold as double stars, and delete the companion star with a higher magnitude in the double stars. After processing, there are still 5046 navigation stars left in the star catalog.

[0009] 1.2 Let each navigation star become the main star in turn. Starting from the main star along its radial direction, select the n nearest neighbor stars to the main star according to the angular distance, and transform the main star and the n neighbor stars from the celestial coordinate system to a certain virtual image coordinate system.

[0010] 1.3 Starting from the radial direction of the navigation main star S0, n neighboring stars with increasing distances from it are arranged as {S'1, S'2, …, S' n}, and the direction vector matrix of each neighboring star relative to the main star S0 is Select the connection line between any neighboring star, such as S'1, and the main star S0 as the starting side, with the counterclockwise direction as the positive direction. Taking the main star S0 as the center, calculate the angle θ' between other neighboring stars and S'1 (m,1) , according to formula (1):

[0011]

[0012] where θ' (m,1) represents the angle between the neighboring star ranked as and the neighboring star ranked as 1, × represents the cross product between two vectors, and ∑ represents the summation of array elements.

[0013] 1.4 Rearrange the n neighboring stars in ascending order of θ' (m,1) , representing the order of neighboring stars starting from S'1 in the counterclockwise direction, i.e., {S1, S2, …, S n}, and its corresponding direction vector matrix relative to S0 is SS n The column vectors in are circularly shifted left by one position, i.e.,

[0014] Calculate the proportional distance elements in the navigation star feature vector, that is, divide the modulus of each vector in SS n _1 by the modulus of the corresponding vector in SS n ;

[0015] Calculate the cosine value of the angle between neighboring stars in the navigation star feature vector, that is, perform the dot product after normalizing the corresponding column vectors in SS n _1 and SS n . The obtained feature elements are in the corresponding relationship. Thus, the feature vectors of each navigation star are obtained, i.e.,

[0016] where,

[0017] 2. Construct the neighboring star index library

[0018] Record the star numbers of the n neighboring stars in ascending order of distance from the main star to construct a preliminary neighboring star index library.

[0019] Arrange the angles θ i between adjacent neighboring stars in ascending or descending order, and change the order of the star numbers of neighboring stars according to the sorting method of θ i to form the final neighboring star index library.

[0020] 3. Build a self-organizing competitive neural network and incorporate the calculation method of Hausdorff distance into the neurons

[0021] The neural network model adopted by this method is a self-organizing competitive neural network. This network has two layers. The first layer is the input layer, and the number of neurons is the same as the length of the navigation star feature vector, that is, 2n. The number of output nodes in the competitive layer is the same as the total number of navigation stars to be classified, that is, N. In this method, the number of neurons in the output layer is 5046.

[0022] Incorporate part of the Hausdorff distance into the calculation method of each neuron. Modify the calculation method of the weight vector and input vector of the neuron;

[0023] In the competitive layer of the self-organizing neural network, when the input vector is closest to the weight vector of the neuron, the neuron wins the competition. Change the measurement method of the closeness between the input vector and the neuron to part of the Hausdorff distance. The weight vector of a single neuron is w = [w1, w'1, …, w n , w' n . First, calculate the matrix A according to Equation (2);

[0024]

[0025] Take the minimum value of each row of the matrix A to form A' = [a'1, a'2, …, a' n-1 , a' n ;

[0026] Arrange the elements in A' in ascending order to form B = [a1, a2, …, a n-1 , a n ;

[0027] Set the value k, 0 < k ≤ 1, and select the first elements in B, denoted as floor division, that is,

[0028] Sum the elements in C, that is, ∑C;

[0029] Neuron activation and output; O = g(∑C + b), where b is the neuron threshold and g(·) is the activation function, and O is the output of the neuron.

[0030] 4. Train the neural network

[0031] Since each neuron in the competitive layer corresponds to a category of navigation stars, and the weight vector of each neuron is in one-to-one correspondence with the feature vector of the corresponding navigation star, the training task can be completed by simply assigning the feature database of the navigation stars to the weight matrix of the neural network.

[0032] 5. Star Chart Recognition

[0033] 5.1 Screen the observed stars; after extracting the centroids of the star points, consider the star points with the centroid distance less than or equal to 4 pixels as double stars, and lock and remove the dimmer one of the double stars.

[0034] 5.2 Extract the feature vectors of the observed stars; sort the observed stars in ascending order according to the distance from the main point, and take each observed star as the observed main star in turn. From the radial direction of the observed main star, take the n nearest neighboring stars to the observed main star, and calculate the distance b from each of the n observed neighboring stars to the observed main star i 。

[0035] Along the circumferential direction with the counterclockwise direction as positive and with the observed main star as the center, calculate the ratio of the distances from adjacent neighboring stars to the main star At the same time, calculate the cosine value of the angle between adjacent observed neighboring stars With the feature vector of this observed star, that is

[0036] 5.3 Establish an index library for the neighboring stars of the observed stars; record the serial numbers of the observed main star and its neighboring stars, and this serial number represents the order of the observed stars from the nearest to the farthest from the main point. Sort the serial numbers of the neighboring stars in ascending order according to the distance from the observed neighboring stars to the observed main star for the first time; with the counterclockwise direction as positive, sort the serial numbers of the observed neighboring stars according to the size in ascending or descending order to form an index library for the neighboring stars of the observed stars.

[0037] 5.4 Match the observed main star; input the feature vector of the observed star into the trained neural network. Each neuron in the competitive layer calculates the partial Hausdorff distance between the input vector and its own weight vector, and the neuron with the smallest partial Hausdorff distance between the weight vector and the input vector wins the competition; since the serial number of the neuron is in one-to-one correspondence with the navigation star number, query the neuron with the output of 1, and it can be considered that the corresponding navigation star has been found for this observed star.

[0038] 5.5 Match the observed neighboring stars; when the observed main star is recognized, query the n corresponding navigation neighboring stars in the navigation star neighboring star index library for this star, which correspond one by one to the n neighboring stars of this star in the observed star neighboring star index library, that is, the observed neighboring stars are matched to the corresponding navigation neighboring stars. Description of the Drawings

[0039] Figure 1 Shows the overall flow block diagram of star map recognition;

[0040] Figure 2 Shows the influence of focal length change on the recognition rate under different parameter modes;

[0041] Figure 3 Shows the influence of the standard deviation of star point positions on the recognition rate under different parameter modes;

[0042] Figure 4 Shows the influence of the standard deviation of star magnitudes on the recognition rate under different parameter modes;

[0043] Figure 5 Shows the influence of false stars on the recognition rate under different parameter modes;

[0044] Figure 6 Shows the influence of missing stars on the recognition rate under different parameter modes. Detailed implementation manners

[0045] The present invention will be further described below in conjunction with the detailed implementation manners with reference to the accompanying drawings.

[0046] A star map recognition method based on the fusion of neural network and Hausdorff distance, comprising the following steps:

[0047] 1. Construct a navigation star feature database

[0048] 1.1 Screen out the qualified navigation stars from the SAO star catalog, regard the star pairs with an angular distance less than a certain angular distance threshold as double stars, and delete the companion star with a higher magnitude in the double stars. After processing, there are still 5046 navigation stars left in the star catalog.

[0049] 1.2 Let each navigation star become the main star in turn. Starting from the main star along its radial direction, select the n nearest neighboring stars to the main star according to the angular distance, and transform the main star and the n neighboring stars from the celestial coordinate system to a certain virtual image coordinate system.

[0050] 1.3 Starting from the radial direction of the navigation main star S0, the n neighboring stars arranged in ascending order of their distances from it are {S'1, S'2,..., S' n}, and the direction vector matrix of each neighboring star relative to S0 is Select the connection line between any neighboring star such as S'1 and the main star S0 as the starting side, with the counterclockwise direction as the positive direction. Taking S0 as the center, calculate the angle θ' between other neighboring stars and S'1 (m,1) , according to formula (1):

[0051]

[0052] where θ' (m,1)Denote the included angle between the neighboring star ranked as and the neighboring star ranked as 1 among neighboring stars. × represents the cross product between two vectors, and ∑ represents the summation of array elements.

[0053] 1.4 Rearrange the n neighboring stars in ascending order of θ' (m,1) in the counterclockwise order starting from S'1, that is, {S1, S2, …, S n}, and the corresponding direction vector matrix relative to S0 is SS n Shift the column vectors in SS

[0054] Calculate the proportional distance elements in the navigation star feature vectors, that is, divide the modulus of each vector in SS n _1 by the modulus of the corresponding vector in SS n ; Calculate the cosine value of the included angle between neighboring stars in the navigation star feature vectors, that is, perform a dot product after normalizing the corresponding column vectors in SS n _1 and SS n , and the obtained feature elements are in the corresponding relationship. Thus, the feature vectors of each navigation star are obtained, that is,

[0055] where

[0056] Taking the number of neighboring stars n = 3 as an example, the feature vectors of each navigation main star are shown in Table 1. The first column is the navigation star number, and the following 6 columns of each feature element correspond to p s respectively.

[0057] Table 1 Navigation star feature database

[0058]

[0059] 2. Construct the neighboring star index library

[0060] Record the numbers of the n neighboring stars in ascending order of distance from the main star to construct a preliminary neighboring star index library.

[0061] The initial neighboring star index library is shown in Table 2. Taking n = 3 as an example, the first column is the number of the main star S'0, and the following three columns are the numbers of the neighboring stars S'1, S'2, S'3 in ascending order of distance from the main star.

[0062] Table 2 Initial neighboring star index library

[0063]

[0064] The included angle θ between adjacent neighboring stars iSorted from small to large or from large to small, according to θ i The sorting method changes the order of the neighboring star asterisks, forming the final neighboring star index library, as shown in Table 3. Taking n = 3 as an example, the first column is the asterisk S0 of the main star, and the next three columns are the neighboring star asterisks S1, S2, S3 in ascending order of distance from the main star.

[0065] Table 3 Neighboring Star Index Library

[0066]

[0067]

[0068] 3. Build a self-organizing competitive neural network and integrate the calculation method of Hausdorff distance into the neurons

[0069] The neural network model adopted by this method is a self-organizing competitive neural network. This network has two layers. The first layer is the input layer, and the number of neurons is the same as the length of the navigation star feature vector, that is, 2n. The number of output nodes in the competition layer is the same as the total number of navigation stars to be classified, that is, N. In this method, the number of neurons in the output layer is 5046.

[0070] Integrate part of the Hausdorff distance into the calculation method of each neuron. Modify the original calculation method of the weight vector and the input vector of the neuron; in the competition layer of the self-organizing neural network, when the input vector is closest to the weight vector of the neuron, the neuron wins the competition. Change the measurement method of the closeness between the input vector and the neuron to part of the Hausdorff distance. The weight vector of a single neuron is w = [w1, w'1,..., w n , w' n , first calculate the matrix A according to formula (2);

[0071]

[0072] Take the minimum value of each row of the matrix A to form A' = [a'1, a'2,..., a' n-1 , a' n ;

[0073] Arrange the elements in A' from small to large to form B = [a1, a2,..., a n-1 , a n ;

[0074] Set the value k, 0 < k ≤ 1, and select the first elements in B, represents rounding down, that is

[0075] Sum the elements in C, that is ∑C;

[0076] Neuron activation and output; O = g(∑C + b), where b is the neuron threshold, g(·) is the activation function, and O is the output of the neuron;

[0077] 4. Training the neural network

[0078] Since each neuron in the competitive layer corresponds to a category of navigation stars, and the weight vector of each neuron is in one-to-one correspondence with the feature vector of the corresponding navigation star, the training task can be completed by simply assigning the feature database of the navigation stars to the weight matrix of the neural network.

[0079] 5. Star chart recognition

[0080] 5.1 Screening the observed stars; after extracting the centroids of the star points, the star points with the centroid distance less than or equal to 4 pixels are regarded as double stars, and the dimmer one of the double stars is locked and removed.

[0081] 5.2 Extracting the feature vectors of the observed stars; sort the observed stars in ascending order of the distance from the main point, and take each observed star as the observed main star in turn. From the radial direction of the observed main star, take the n nearest neighboring stars to the observed main star, and calculate the distance b from each of the n observed neighboring stars to the observed main star i ;

[0082] Along the circumferential direction with the counterclockwise direction as positive, with the observed main star as the center, calculate the ratio of the distances from adjacent neighboring stars to the main star At the same time, calculate the cosine value of the angle between adjacent observed neighboring stars Using the feature vector of this observed star, that is

[0083] 5.3 Establishing the observed star neighboring star index library; record the serial numbers of the observed main star and its neighboring stars. This serial number represents the order of the observed stars from the nearest to the farthest from the main point. Sort the serial numbers of the neighboring stars in the first time according to the distance from the observed neighboring stars to the observed main star from the nearest to the farthest; with the counterclockwise direction as positive, sort the serial numbers of the observed neighboring stars in ascending or descending order according to the size of, and form the observed star neighboring star index library.

[0084] 5.4 Matching the observed main star; input the feature vector of the observed star into the trained neural network. Each neuron in the competitive layer calculates the partial Hausdorff distance between the input vector and its own weight vector. The neuron with the smallest partial Hausdorff distance between the weight vector and the input vector wins the competition; since the serial number of the neuron is in one-to-one correspondence with the navigation star number, query the neuron with the output of 1, and it can be considered that the observed star has found its corresponding navigation star.

[0085] 5.5 Observation of neighboring star matching; after the observation of the primary star is identified, query the n navigation neighboring stars corresponding to this star in the navigation star neighboring star index library, and correspond one by one with the n neighboring stars of this star in the observation star neighboring star index library, that is, the observed neighboring stars are matched to the corresponding navigation neighboring stars.

[0086] Embodiment

[0087] In the initial stage of the design of star sensors and star map recognition algorithms, in order to determine the best recognition parameters, the performance of the algorithms is generally evaluated preliminarily by means of pure digital simulation. In the present invention, through simulation, with MATLAB 2022a as the simulation platform, simulated star maps are first generated, and noises that may occur in actual situations are added to the simulated star maps to test the robustness of the algorithms to the corresponding noises.

[0088] The relevant parameter settings for star map simulation in the present invention are shown in Table 4.

[0089] Table 4 Star map simulation parameters

[0090]

[0091] Using the Monte Carlo method, 100,000 star maps are randomly simulated and generated in the entire celestial sphere. Under the condition of no interference factors, the simulated star maps are recognized by this method, and the recognition rate is 100%, and there is no misrecognition situation.

[0092] In order to select the best recognition parameters, the present invention selects different recognition parameter modes to verify the performance of the star map recognition algorithm under the interference factors of parameter deviation, star point position deviation, star magnitude deviation, and pseudo-star magnitude in the star sensor. The selected recognition parameters include the number of neighboring stars n, and k participating in the calculation of the partial Hausdorff distance, 0 < k ≤ 1; the value of n can be n = 3, 4, 5, 6, and the value of k is k = 0.8, 1;

[0093] According to the values of n and k, there are a total of 8 recognition parameter modes, as shown in Table 5;

[0094] Table 5 Recognition parameter modes

[0095]

[0096] Due to the defects in the installation and adjustment process of the star sensor, as well as the vibration and radiation suffered during the launch and operation of the carrier, a certain degree of internal parameter drift will occur, such as key parameters like focal length. In order to simulate this situation, the focal length change range is set to [-2, 2], and the change amplitude each time is 1 mm. Under the 8 modes in Table 5, the influence of focal length change on the algorithm performance is evaluated respectively, as Figure 2 shown. It can be seen from Figure 2 that the best parameter recognition mode is Mode5. At this time, the focal length change will not affect the recognition rate, and the recognition rate of the algorithm is 100%;

[0097] Affected by image noise, optical system distortion, and the systematic error of the centroid positioning algorithm, the true star position deviates from the ideal star position. To simulate this deviation of the star position, noise that follows a Gaussian distribution with a mean of 0 and a standard deviation varying from 0 to 7 pixels, with a change amplitude of 1 pixel each time, is added to the star positions in the simulated star map. The influence of the star position deviation on the algorithm is evaluated under the 8 modes in Table 5, as Figure 3 shown. The optimal parameter identification mode is Mode6. When the standard deviation of the star is 7 pixels, the recognition rate of the algorithm is 99.75%;

[0098] Affected by the spectral characteristics of stars, the brightness characteristics of stars, and the performance of optoelectronic detection devices, there is inaccuracy in the measurement of the brightness of stars by the optoelectronic detection device of the star sensor, that is, there is magnitude noise. During the generation process of the simulated star map, noise that follows a Gaussian distribution with a mean of 0 and a standard deviation varying from 0 to 1 Mv, with a change amplitude of 0.2 Mv each time, is added to the magnitude. The influence of the star position deviation on the algorithm is evaluated under the 8 modes in Table 5, as Figure 4 shown. Since the magnitude information is not utilized when constructing the feature mode of the present invention, as can be seen from Figure 4 it, the optimal parameter identification mode is Mode5. At this time, the magnitude deviation does not affect the recognition rate, and the recognition rate of the algorithm is 100%;

[0099] Pseudo-stars can be divided into false stars and missing stars. False stars are due to planets, space debris, engine high-brightness substances, etc., and their images are formed on the photosensitive surface of the star sensor, making it difficult to separate them from the true star targets. When the algorithm processes them, they will also be regarded as observed stars, thus affecting the recognition performance of the algorithm; Missing stars are due to the performance of the detector and the influence of some space obstacles, resulting in stars that should have been captured by the detector not being captured, causing starry sky gaps in some positions. For false stars, 0 to 5 stars with magnitudes of 3 to 6 Mv are randomly added to the simulated star map; for missing stars, during the generation process of the star map simulation, 0 to 5 stars are randomly deleted; The influence of false stars and missing stars on the algorithm is evaluated under the 8 modes in Table 5. The influence of false stars on the recognition performance of the algorithm under different modes is as Figure 5 shown. The optimal parameter identification mode is Mode5. When the number of false stars is 5, the recognition rate of the algorithm is 99.86%; The influence of missing stars on the recognition performance of the algorithm under different modes is as Figure 6 shown. The optimal parameter identification mode is Mode5. When the number of missing stars is 5, the recognition rate of the algorithm is 99.16%;

[0100] It should be noted that the above description of the preferred embodiment is relatively detailed, and it should not be considered as a limitation to the protection scope of the present invention. Under the inspiration of the present invention, those of ordinary skill in the art can also make several improvements and refinements without departing from the protection scope defined by the claims of the present invention. These improvements and refinements should also be regarded as the protection scope of the present invention. The scope of protection claimed by the present invention shall be subject to the appended claims.

Claims

1. A star map recognition method based on the fusion of neural network and Hausdorff distance, characterized in that the method comprises the following steps: 1.1 Building a navigation feature database includes the following steps: Select the navigation stars that meet the requirements from the SAO star catalog, let each navigation star become the main star in turn, take the main star as the center, start from its radial direction, select the n neighboring stars closest to the main star according to the angular distance, and transform the main star and the n neighboring stars from the celestial coordinate system to a virtual image coordinate system; Along its circumferential direction, select any neighboring star as the starting point, and calculate the distances from n neighboring stars to the main star in turn. , and calculate the ratio of the distance from the adjacent neighboring star to the main star , and calculate the cosine of the angle between adjacent satellites , which is used as the characteristic vector of the navigation star, namely ; 1.2 Build a neighboring star index library, Record the star numbers of n neighboring stars from near to far from the main star, build a preliminary neighbor star index library, and calculate the angles between adjacent neighbor stars. Arrange from small to large or from large to small, according to The sorting method changes the order of neighbor star numbers in the preliminary neighbor star index library to form the final neighbor star index library; 1.3 Build a neural network model, A self-organizing competitive neural network is used. The network has two layers. The first layer is the input layer. The number of neurons is the same as the length of the navigation star feature vector. The number of output nodes in the competition layer is the same as the total number of navigation stars to be classified. Part of the Hausdorff distance is integrated into the calculation method of each neuron. 1.4 Training the neural network, Each neuron in the competition layer corresponds to a category of navigation star, and the weight vector of each neuron corresponds one-to-one to the feature vector of the corresponding navigation star. Therefore, the training task is completed by assigning the feature database of the navigation star to the weight matrix of the neural network. 1.5 Star map recognition, including the following steps: 1.5.1 Screen the observed stars; after extracting the centroids of the star points, consider the star points with the centroid distance less than or equal to 4 pixels as double stars, and lock and remove the star with the lower brightness among the double stars; 1.5.2 Observation star feature vector extraction: Sort each observation star in ascending order according to the distance from the main point, and take each observation star as the main observation star in turn. From the radial direction of the main observation star, take the n nearest neighboring stars to the main observation star, and calculate the distances from the n observation neighboring stars to the main observation star in turn. ; Along the circumferential direction with the counterclockwise direction being positive, taking the observed primary star as the center, calculate the ratio of the distances from adjacent neighboring stars to the primary star , and at the same time calculate the cosine value of the angle between adjacent observed neighboring stars . Taking this as the characteristic vector of the observed star, that is ; 1.5.3 Establish an index library of neighboring stars of the observed stars; record the serial numbers of the observed main star and its neighboring stars. This serial number represents the order of the distances from the observed stars to the main point from near to far. Sort the serial numbers of the neighboring stars for the first time in the order of the distances from the observed neighboring stars to the observed main star from near to far; Taking counterclockwise as the positive direction, the serial numbers of the observed neighboring stars are arranged in ascending or descending order according to the magnitude of to form an observed star neighboring star index library; 1.5.4 Matching of the observed main star; input the feature vector of the observed star into the trained neural network. Each neuron in the competition layer calculates the partial Hausdorff distance between the input vector and its own weight vector. The neuron with the smallest partial Hausdorff distance between the weight vector and the input vector wins the competition; the serial number of the neuron corresponds one-to-one with the star number of the navigation star. Query the neuron with the output of 1 to find the navigation star corresponding to the observed star; 1.5.5 Matching of the observed neighboring stars; When the observed main star is identified, query the n navigation neighboring stars corresponding to this star in the navigation star neighboring star index library, which correspond one-to-one with the n neighboring stars of this star in the observed star neighboring star index library, that is, realize the matching of the observed neighboring stars with the corresponding navigation neighboring stars.

2. The method according to claim 1, wherein The screening of the navigation stars that meet the requirements from the SAO star catalog includes the following steps: Consider the star pairs with the angular distance less than a certain angular distance threshold as double stars, delete the companion star with the higher magnitude among the double stars, and the remaining stars in the star catalog are the navigation stars.

3. The method according to claim 1, wherein The feature extraction of the navigation stars includes the following steps: Starting from the radial direction of the navigation primary star and arranging n neighboring stars with increasing distances from it as , the direction vector matrix of each neighboring star relative to the primary star is . Arbitrarily select a neighboring star such as and use the connection line with the primary star as the starting side. With the counterclockwise direction as the positive direction and taking the primary star as the center, calculate the angles between other neighboring stars and according to the following formula: Among them, , , represents the cross product between two vectors, represents the summation of array elements; Arrange n neighboring stars in ascending order, representing the order of neighboring stars in the counterclockwise direction starting from That is, starting from The order of neighboring stars in the counterclockwise direction, namely The corresponding direction vector matrix relative to the primary star is . Shift the column vectors in one position to the left in a circular manner, that is ; Calculate the proportional distance elements in the navigation star feature vector, that is, divide the modulus of each vector in by the modulus of the corresponding vector in ; Calculate the cosine value of the adjacent star angle in the navigation star feature vector, that is, perform a dot product after normalizing the corresponding column vectors in and . Thus, the navigation star feature vector is obtained, that is: ; Among them, , .

4. The method according to claim 1 or 3, characterized in that, The method for extracting the feature vector of the observed stars is the same as that for extracting the feature vector of the navigation stars.

5. The method according to claim 1, characterized in that, Incorporating the partial Hausdorff into the neurons of the neural network includes the following steps: Modify the original calculation method of the weight vector of the neuron and the input vector; in the competitive layer of the self-organizing neural network, when the input vector is closest to the weight vector of the neuron, this neuron wins the competition, and change the measurement method of the closeness between the input vector and the neuron to the partial Hausdorff distance. The weight vector of a single neuron is , and its calculation method with the navigation star feature vector is as follows: Calculate matrix A; Take the minimum value of each row of matrix A to form ; Arrange the elements in in ascending order to form ; Set value , select the first elements in B, represents rounding down, that is ; Perform neuron activation and output according to the following formula; , where is the neuron threshold, is the activation function, is the output of the neuron.

6. The method according to claim 1, characterized in that The input of the feature vector of the observed star into the trained neural network includes the following steps: When the feature vector of the observed star is input into the self-organizing competitive neural network, the partial Hausdorff distance is used as the method for each neuron in the competition layer to calculate the activation value.

Images