A method for predicting optimal mission satellites for ultra-large-scale remote sensing constellations based on BP neural network

Abstract

The present invention is a method for predicting the optimal mission star for a very large-scale remote sensing constellation based on a BP neural network. The present invention relates to the technical field of mission star prediction. The present invention establishes a BP neural network, prepares training samples, trains the BP neural network, and predicts remote sensing mission stars based on the trained BP neural network. Although the BP network of the present invention cannot accurately select mission stars, when used for predicting the optimal star for a very large-scale constellation, it can ensure that the predicted star can actually perform the mission. On this basis, the prediction result of the optimal star is credible to a certain extent, and can effectively reduce the size of the solution space, reduce the amount of computation, and improve timeliness, so as to support the demand for predicting mission stars for very large-scale remote sensing constellations. The method has broad application prospects.

Description

Technical Field

[0001] The invention relates to the technical field of mission star prediction, and is a method for predicting the optimal mission star of a super-large-scale remote sensing constellation based on a BP neural network. Background Art

[0002] With the rapid deployment and full-scale implementation of large-scale constellations, constellations are gradually increasing in size, often reaching thousands or tens of thousands of satellites. Ultra-large remote sensing constellations, with their large number of members and the volume of control data, are often used to handle time-sensitive missions, which are characterized by their sudden, complex, and time-sensitive nature. With the increasing number and widespread distribution of satellites, relying solely on manual control would become extremely burdensome, wasting significant material and human resources, significantly increasing control time, and making it impossible to cope with the rapidly changing space environment of the future.

[0003] Traditional multi-satellite mission planning for constellations primarily addresses the issue of limited observation resources, specifically oversubscription. However, ultra-large-scale constellations, with their relatively abundant observation resources, require selecting the most competitive satellites from a vast number of satellites. While existing orbit prediction and coverage calculation methods offer high accuracy, they suffer from high computational complexity, impacting mission real-time performance. Furthermore, predicting the optimal satellite requires complex computation of ongoing evaluation metrics. In recent years, researchers in the remote sensing mission field have proposed the concept of "task schedulability," leveraging historical scheduling data to predict the successful scheduling of remote sensing satellites (EOSs) before task assignment. Bal et al. (2015) used an integrated BP neural network to predict the schedulable satellites for single-satellite remote sensing missions, achieving an average prediction accuracy exceeding 85%. Li et al. (2013) designed a data-driven scheduling algorithm using a probabilistic prediction model constructed using a neural network. Zong et al. (2021) and Son (2015) extracted features such as task priority, duration, flexibility, subscription and conflict from historical EOS data, and used a BP neural network model with a variable hidden layer to formulate task plans. For the scheduling of multiple EOS, WAN Geta. (2011) used the Neuroevolution of Enhanced Topology (NEAT) algorithm to develop a reinforcement learning model for each satellite and proved the convergence and effectiveness of the model. Due et al. (2020) For the large-scale and time-consuming multi-AEOS scheduling problem, this paper proposed a data-driven parallel scheduling method consisting of a probabilistic prediction model, a task allocation strategy and a parallel scheduling method. Chen et al. (2022) proposed a real-time multi-satellite scheduling method consisting of a hierarchical prediction model based on machine learning and a heuristic local search algorithm, drawing on existing historical multi-EOS observation plans to generate high-quality initial solutions for the current scheduling scenario. However, although previous researchers have used neural network methods to effectively shorten the response time of large-scale scheduling problems, they are mainly aimed at situations with a small number of satellites and solving task scheduling problems with multiple observation targets. The prediction models established are only applicable to single satellites, and there is little discussion on ultra-large-scale constellations. There has been no research on directly using artificial neural networks to predict task satellites. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a method for predicting the optimal mission star of a super-large-scale remote sensing constellation based on a BP neural network.

[0005] The present invention provides a method for predicting the optimal mission star of a super-large-scale remote sensing constellation based on a BP neural network. The present invention provides the following technical solutions:

[0006] A method for predicting the optimal mission star of a super-large-scale remote sensing constellation based on a BP neural network comprises the following steps:

[0007] Step 1, establish BP neural network;

[0008] Step 2: prepare training samples and train the BP neural network;

[0009] Step 3: Based on the trained BP neural network, the remote sensing mission stars are predicted.

[0010] Preferably, the step 1 is specifically:

[0011] Determine network input and output by task completion indicator: response time T cos t j，k and observation time Tsee j，k , to predict the optimal mission star, the output of the mission star optimization BP network is clear, respectively, the remote sensing mission response time T cos t j，k and the duration of remote sensing satellite observation T see j，k ,The network input selects the observation situation of the low-orbit remote sensing satellite, which is determined by the initial orbit ,parameters, the position of the observation target and the ,illumination conditions.

[0012] Preferably, the initial orbital inclination i is determined j，0 、Initial perigee argument ω j，0 、Initial ascending node right ascension Ω j，0 、Initial true anomaly f j，0 Considering the observation of the target point by remote sensing satellite j, we must first calculate its sub-satellite point position. The sub-satellite point position is determined by the satellite orbit position. At a certain time t, the latitude and longitude of the sub-satellite point [α j，t ,λ j，t ] is calculated by the following formula:

[0013] sinα j，t =sin(ω j，t +f j，t )sin i j，t

[0014]

[0015] Among them, α G，j，t is the Greenwich mean sidereal time corresponding to the current moment;

[0016] Determine the initial orbital semi-major axis a j，0 and the initial orbital eccentricity e j，0 The coverage of the sensor is described by the angle between the sight vector of the sensor field of view boundary and the satellite position vector in the geocentric inertial coordinate system. The earth is a standard sphere with a radius of R. E , set all remote sensing satellite sensors to have a conical field of view, with a semi-cone angle of η jAssume that the distance from the satellite to the Earth's center is r t , calculated by the following formula:

[0017]

[0018]

[0019] Among them, β j，t is the geocentric angle corresponding to the great circle arc between the intersection of the sensor line of sight and the earth's surface and the sub-satellite point of the remote sensing satellite. j，t ≤η means that the remote sensing satellite has geometric visibility to the target at the current moment;

[0020] Determine the latitude α of the mission target point k , target point longitude λ k , according to the latitude and longitude of the mission target [α k ,λ k ]、Satellite subsatellite position [α j，t ,λ j，t l can be calculated j，t , and then determine whether the remote sensing satellite has geometric visibility with the target;

[0021] Determine the time when the task is issued (Julian date jd) k , using the mission issuance time as the initial time t=0, using the DE430 ephemeris published by the Jet Propulsion Laboratory of the United States, calculate the sun's position at any time t in the next hour; based on the sun's position, the remote sensing satellite position and the target position, it can be calculated whether the remote sensing payload meets the lighting conditions.

[0022] Preferably, the hidden layer structure of the neural network needs to be determined, and the optimal task star is predicted. The network parameter range of the candidate BP network is determined using the empirical formula and the optimal network selection method, and ten groups of optimal and suboptimal networks are selected in parallel to fit the prediction results to ensure the accuracy of the prediction results.

[0023] The activation function used is the classic sigmoid function, which is expressed by the following formula:

[0024]

[0025] Preferably, the step 2 is specifically as follows:

[0026] Prepare training samples, generate a low-orbit remote sensing satellite constellation of 28,800 satellites based on the distribution characteristics of existing low-orbit remote sensing satellites, and use January 1, 2024 as the constellation control base time, and the task issuance time is jd kRandomly generated within one day before and after the benchmark time, the approximate area of ​​the time-sensitive remote sensing mission distribution is determined based on the distribution of the generated ultra-large-scale remote sensing constellation. Within this area, the time-sensitive mission target point [α k ,λ k ] Randomly generate and calculate time-sensitive tasks in jd k After the time is issued, the initial orbit parameter is a j，0 、e j，0 、i j，0 、ω j，0 ,Ω j，0 、f j，0 The remote sensing satellite j will detect the target [α k ,λ k ]'s visible situation V j，k If it is visible within the next hour, V j，k =1, otherwise 0;

[0027] For the task response time prediction network and the observation duration prediction network, the training sample is regarded as an array containing a pair of network input and network output. The input is:

[0028] a j，0 、e j，0 、i j，0 、ω j，0 ,Ω j，0 、f j，0 , α k ,λ k and jd k There are 9 parameters in total, and the output is the response time T cos t j，k and observation time T see j，k ,Considering that the prediction of time parameters is relatively ,complex, multiple groups of different mission stars corresponding to ,different observation tasks are selected, with a total of 6238 groups of samples for ,training and analysis.

[0029] Preferably, a fault-tolerant computing mechanism is used to ensure the accuracy of the optimal task star prediction. When optimizing the task star, the prediction results of each sub-network are fully considered, and the prediction results of all networks are averaged to reduce the prediction error.

[0030] Preferably, the step 3 is specifically:

[0031] An optimal mission star prediction network was established. The network randomly divided 6,238 samples into two groups: training data and test data, accounting for 80% and 20% of the total samples, respectively. During the training process of the mission star prediction network, it was found that the prediction accuracy of the dual hidden layer was higher than that of the single hidden layer. After training all candidate dual hidden layer networks, the network performance was analyzed and compared.

[0032] A BP neural network-based system for predicting optimal mission stars in a large-scale remote sensing constellation, comprising:

[0033] A network building module, wherein the network building module builds a BP neural network;

[0034] A sample module, wherein the sample module prepares training samples and trains the BP neural network;

[0035] The prediction module predicts the remote sensing mission stars based on the trained BP neural network.

[0036] A computer-readable storage medium stores a computer program, which is executed by a processor to implement a method for predicting the optimal mission star of a super-large-scale remote sensing constellation based on a BP neural network.

[0037] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, a BP neural network-based method for predicting the optimal mission star of a super-large-scale remote sensing constellation is implemented.

[0038] The present invention has the following beneficial effects:

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] This paper, based on the rapid prediction of mission stars for ultra-large-scale constellations, proposes a new method for predicting optimal observation stars based on a BP neural network. First, the input and output of the prediction network are determined to ensure that the main factors affecting mission execution are fully considered. Subsequently, a constellation of 28,800 low-orbit remote sensing satellites is generated based on the existing distribution characteristics of low-orbit remote sensing satellites. Network training and testing samples are generated, completing the prediction network training with fault-tolerant and veto mechanisms. Finally, a comprehensive analysis and evaluation of the trained prediction network's performance in optimal star prediction are conducted. The results show that while the BP network cannot accurately select mission stars, when used for optimal star prediction in ultra-large-scale constellations, it can ensure that the predicted stars are actually capable of executing the mission. This provides a certain degree of confidence in the optimal star prediction results, effectively reducing the solution space, reducing computational complexity, and improving timeliness, thereby meeting the demand for mission star prediction in ultra-large-scale remote sensing constellations. The method has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 Schematic diagram of the covered field of view;

[0043] Figure 2 This is a distribution map of super-large-scale star clusters;

[0044] Figure 3 The star screening accuracy for the single hidden layer task with different numbers of neurons;

[0045] Figure 4 The accuracy of the training set for the task of star screening with different numbers of neurons and two hidden layers;

[0046] Figure 5 The prediction error of task response time for training sets with different network structures;

[0047] Figure 6 Distribution of prediction errors of the fault-tolerant network for the response time test set;

[0048] Figure 7 Distribution diagram of the prediction error of the optimal network for the response time test set;

[0049] Figure 8 Observe the prediction error of the training set for different network structures;

[0050] Figure 9 Observe the duration prediction error of the test set for different network structures;

[0051] Figure 10 The distribution diagram of the prediction error of the fault-tolerant network for the observation duration test set;

[0052] Figure 11 This is the distribution diagram of the optimal network prediction error of the observation duration test set. DETAILED DESCRIPTION

[0053] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0054] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0055] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0056] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0057] The present invention is described in detail below with reference to specific embodiments. Specific embodiment one:

[0059] according to Figures 1 to 11 As shown, the specific optimization technical solution adopted by the present invention to solve the above technical problems is: the present invention relates to a method for predicting the optimal mission star of a super-large-scale remote sensing constellation based on BP neural network.

[0060] A method for predicting the optimal mission star of a super-large-scale remote sensing constellation based on a BP neural network comprises the following steps:

[0061] Step 1, establish BP neural network;

[0062] Step 2: prepare training samples and train the BP neural network;

[0063] Step 3: Based on the trained BP neural network, the remote sensing mission stars are predicted. Specific embodiment two:

[0065] The difference between the second embodiment of the present invention and the first embodiment is that:

[0066] The step 1 is specifically as follows:

[0067] Determine network input and output by task completion indicator: response time T cos tj，k and observation time Tsee j，k , to predict the optimal mission star, the output of the mission star optimization BP network is clear, respectively, the remote sensing mission response time T cos t j，k and the duration of remote sensing satellite observation T see j，k ,The network input selects the observation situation of the low-orbit remote sensing satellite, which is determined by the initial orbit ,parameters, the position of the observation target and the ,illumination conditions. Specific embodiment three:

[0069] The only difference between the third embodiment of the present invention and the second embodiment is that:

[0070] First, based on the characteristics of remote sensing mission management of large-scale remote sensing constellations, the following assumptions are made: Assumption 1: When predicting the remote sensing mission situation of a single satellite, its attitude maneuverability is ignored and only the impact of the orbital position on the observation range is considered.

[0071] Assumption 2: The remote sensing payloads on each satellite in the constellation are fixed to the satellite's coordinate system and do not rotate or yaw. This means the centerline of the camera's field of view always points toward the subsatellite point and remains constant. It is assumed that all remote sensing satellite sensors have a conical field of view with a semi-cone angle of 7° = 15°.

[0072] Assumption 3: This paper considers a remote sensing satellite constellation primarily composed of low-orbit satellites, whose orbital periods are relatively short. Therefore, it is assumed that when a remote sensing mission occurs, each satellite in the constellation will complete one orbital period within the next hour. The mission satellite prediction problem can be simplified to simply determining whether the satellite will be visible to the mission target within the next hour. Therefore, after comprehensive considerations, the input parameters for the mission satellite screening BP network and the mission satellite selection BP network are determined to be the same, with a total of nine parameters summarized as network inputs.

[0073] Determine the initial orbital inclination i j，0 、Initial perigee argument ω j，0 、Initial ascending node right ascension Ω j，0 、Initial true anomaly f j，0 Considering the observation of the target point by remote sensing satellite j, we must first calculate its sub-satellite point position. The sub-satellite point position is determined by the satellite orbit position. At a certain time t, the latitude and longitude of the sub-satellite point [α j，t ,λ j，t ] is calculated by the following formula:

[0074] sinα j，t =sin(ω j，t +f j，t )sin i j，t

[0075]

[0076] Among them, α G，j，t is the Greenwich mean sidereal time corresponding to the current moment;

[0077] Determine the initial orbital semi-major axis a j，0 and the initial orbital eccentricity e j，0 The coverage of the sensor is described by the angle between the sight vector of the sensor field of view boundary and the satellite position vector in the geocentric inertial coordinate system. The earth is a standard sphere with a radius of R. E , set all remote sensing satellite sensors to have a conical field of view, with a semi-cone angle of η j Assume that the distance from the satellite to the Earth's center is r t , calculated by the following formula:

[0078]

[0079]

[0080] Among them, β j，t is the geocentric angle corresponding to the great circle arc between the intersection of the sensor line of sight and the earth's surface and the sub-satellite point of the remote sensing satellite. j，t ≤η means that the remote sensing satellite has geometric visibility to the target at the current moment;

[0081] Determine the latitude α of the mission target point k , target point longitude λ k , according to the latitude and longitude of the mission target [α k ,λ k ]、Satellite subsatellite position [α j，t ,λ j，t ] then β can be calculated j，t , and then determine whether the remote sensing satellite has geometric visibility with the target;

[0082] Determine the time when the task is issued (Julian date jd) k , using the mission issuance time as the initial time t=0, using the DE430 ephemeris published by the Jet Propulsion Laboratory of the United States, calculate the sun's position at any time t in the next hour; based on the sun's position, the remote sensing satellite position and the target position, it can be calculated whether the remote sensing payload meets the lighting conditions. Specific embodiment four:

[0084] The only difference between the fourth embodiment of the present invention and the third embodiment is that:

[0085] It is necessary to determine the hidden layer structure of the neural network, predict the optimal task star, use the empirical formula and the optimal network selection method to determine the network parameter range of the candidate BP network, and select ten sets of parallel optimal and suboptimal networks to fit the prediction results to ensure the accuracy of the prediction results;

[0086] The activation function used is the classic sigmoid function, which is expressed by the following formula:

[0087] Specific embodiment five:

[0089] The only difference between the fifth embodiment of the present invention and the fourth embodiment is that:

[0090] Prepare training samples, generate a low-orbit remote sensing satellite constellation of 28,800 satellites based on the distribution characteristics of existing low-orbit remote sensing satellites, and use January 1, 2024 as the constellation control base time, and the task issuance time is jd k Randomly generated within one day before and after the benchmark time, the approximate area of ​​the time-sensitive remote sensing mission distribution is determined based on the distribution of the generated ultra-large-scale remote sensing constellation. Within this area, the time-sensitive mission target point [α k ,λ k ] Randomly generate and calculate time-sensitive tasks in jd k After the time is issued, the initial orbit parameter is a j，0 、e j，0 、i j，0 、ω j，0 ,Ω j，0 、f j，0 The remote sensing satellite j will detect the target [α k ,λ k ]'s visible situation V j，k If it is visible within the next hour, V j，k =1, otherwise 0;

[0091] For the task response time prediction network and the observation duration prediction network, the training sample is regarded as an array containing a pair of network input and network output. The input is:

[0092] a j，0 、e j，0 、i j，0 、ω j，0 ,Ω j，0 、f j，0 , α k ,λ k and jd k There are 9 parameters in total, and the output is the response time T cos t j，k and observation time T see j，k ,Considering that the prediction of time parameters is relatively ,complex, multiple groups of different mission stars corresponding to ,different observation tasks are selected, with a total of 6238 groups of samples for ,training and analysis. Specific embodiment six:

[0094] The only difference between the sixth embodiment of the present invention and the fifth embodiment is that:

[0095] A fault-tolerant computing mechanism is used to ensure the accuracy of the prediction of the optimal task star. When optimizing the task star, the prediction results of each sub-network are fully considered, and the prediction results of all networks are averaged to reduce the prediction error. Specific embodiment seven:

[0097] The only difference between the seventh embodiment of the present invention and the sixth embodiment is that:

[0098] The step 3 is specifically as follows:

[0099] An optimal mission star prediction network was established. The network randomly divided 6,238 samples into two groups: training data and test data, accounting for 80% and 20% of the total samples, respectively. During the training process of the mission star prediction network, it was found that the prediction accuracy of the dual hidden layer was higher than that of the single hidden layer. After training all candidate dual hidden layer networks, the network performance was analyzed and compared. Specific embodiment eight:

[0101] The only difference between the eighth embodiment of the present invention and the seventh embodiment is that:

[0102] The present invention provides a BP neural network-based system for predicting optimal mission stars in a super-large-scale remote sensing constellation. The system comprises:

[0103] A network building module, wherein the network building module builds a BP neural network;

[0104] A sample module, wherein the sample module prepares training samples and trains the BP neural network;

[0105] The prediction module predicts the remote sensing mission stars based on the trained BP neural network. Specific embodiment nine:

[0107] The only difference between the ninth embodiment of the present invention and the eighth embodiment is that:

[0108] The present invention provides a computer-readable storage medium having a computer program stored thereon. The program is executed by a processor to implement a method for predicting optimal mission stars of a super-large-scale remote sensing constellation based on a BP neural network. Specific embodiment ten:

[0110] The only difference between the tenth embodiment of the present invention and the ninth embodiment is that:

[0111] The present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, a method for predicting the optimal mission star of a super-large-scale remote sensing constellation based on a BP neural network is implemented.

[0112] Ultra-large constellations will number over 10,000 satellites, generating enormous amounts of control data. These constellations are often used to handle sudden, complex, and time-sensitive missions. To meet these time-sensitive requirements, a fast and relatively accurate mission satellite prediction method is required. Traditional methods select mission satellites by traversing the constellation's satellites and recursively inferring their orbits. While accurate, these methods are computationally expensive and cannot guarantee timely decisions for ultra-large constellations. This paper proposes a new mission satellite prediction method that does not rely on orbit recursion. A back-propagation (BP) neural network is used to directly fit the relationship between satellite orbital parameters, mission launch time, and target location, and the performance of the observation mission. Since there are only 1,192 remote sensing satellites currently in orbit, this paper generates a constellation of 28,800 low-orbit remote sensing satellites based on the distribution characteristics of existing low-orbit remote sensing satellites. This is used to generate network training and testing samples. Considering that residual errors can still occur in the BP neural network even after all parameters have been optimized, a fault-tolerant mechanism and a veto mechanism are employed to ensure the reliability of mission satellite selection. The results show that the predicted mission star's actual executable probability and the mission indicator prediction error meet the requirements to a certain extent for mission star screening and prediction of the optimal mission star (optimal response time and observation duration). In addition, because it does not involve orbit recursion, the BP network has high computational efficiency in mission star prediction, showing good application prospects. Specific embodiment eleven:

[0114] The only difference between the eleventh embodiment of the present invention and the tenth embodiment is that:

[0115] The BP neural network is the most commonly used algorithmic model in artificial neural networks and has a relatively comprehensive theoretical foundation. Composed of multi-layer perceptrons, BP neural networks, through the forward propagation of input signals and the reverse diversion of error information, undergo an iterative learning and training process to establish an intelligent network model specifically for predicting or making decisions regarding signal data with fuzzy structural parameters. This model mimics the human brain's processing of external stimuli. BP neural networks offer unique advantages in processing complex signal data, particularly their strong self-correction capabilities and fault tolerance for extreme signals.

[0116] Forward propagation of signals, backward propagation of errors, learning and training, and convergence. During forward propagation, external input signals enter the network from the input layer according to the order of neuron connections, then pass through the hidden layer to the output layer. When the output value of the neural network differs significantly from the desired output, a second stage of training is required. During backward propagation, error signals that exceed the accuracy requirements are propagated backward from the output end layer by layer, shunting the errors to the neurons in each layer and adjusting the connection weights of each layer. The first two processes alternate and adjust each other, so that the weights and thresholds of each neuron in the network are continuously adjusted, ultimately meeting the desired output requirements. This marks the end of the BP neural network learning and training phase and the achievement of convergence. The ultimate goal of training is to minimize the difference between the actual output of the neural network and the target output, thereby achieving network convergence.

[0117] When dealing with nonlinear, high-dimensional problems, neural networks do not require accurate structural parameters for the input and output functions themselves. Learning and training can be used to acquire the inherent relationships between them. Furthermore, when fed data outside the training dataset, the neural network can generate accurate output values. Tang and Gong (2023) used neural networks to determine the relationship between the initial state of a vertical landing phase and its fuel consumption. Han et al. (2023) used BP neural networks to locate space debris impact sources. Pérezeta et al. (2014) used neural networks to predict atmospheric density values ​​for a spacecraft's future orbit. Existing research has shown that BP networks with a single hidden layer can fit nonlinear functions by properly configuring the number of neurons and network parameters. Therefore, using BP neural networks to fit the nonlinear mapping between space object orbital parameters and remote sensing missions has a solid theoretical foundation. (Li et al. (2023)) can be applied to solving the problem of satellite prediction for ultra-large-scale remote sensing satellite constellations.

[0118] Training a BP network involves using samples to calibrate the network's weights and biases so that it correctly fits the expected input-output relationship. The basic BP algorithm consists of two steps: forward propagation of input information and backward propagation of output errors. As these two steps are repeated, the weights and biases of the neural network are continuously adjusted until the network output approaches the expected value. Even after optimizing all available network characteristics, residual errors often remain. To further reduce this error, we propose using a fault-tolerant computational mechanism to ensure the accuracy of task star selection. We operate not a single network, but a collection of networks, each trained on the same dataset. The basic idea is to use predictions from each subnetwork to determine the final prediction. Key to ensuring accurate task star selection is that each subnetwork has a veto power; if a negative result is produced, the task star is directly excluded from the set of feasible solutions. When selecting the task star from the collection, the predictions of each subnetwork are fully considered, and all predictions are averaged to reduce prediction error.

[0119] The Mission Star Rapid Screening Network randomly divided 1,563 samples into two groups: training data and test data, accounting for 80% and 20% of the total sample, respectively. The training data was used to train the BP neural network, while the test data was used to evaluate the network's generalization ability. After training all candidate networks, the performance of the networks was analyzed and compared.

[0120] When using training samples to evaluate network performance, the more neurons there are, the higher the accuracy of task star screening. However, when the number of neurons increases to a certain level, the accuracy no longer increases, and the accuracy remains at around 95% near the highest point. When using test samples to evaluate network performance, the prediction accuracy fluctuates less after the number of neurons increases to 15, and the accuracy basically remains above 85%. Therefore, when the number of neuron nodes in a single hidden layer is between 36 and 49, the accuracy and generalization ability of the task star screening network are

[0121] For a BP network with two hidden layers, the task star screening accuracy for different numbers of hidden layer neurons is shown. When evaluating network performance using training samples, a greater number of neurons in each hidden layer leads to a higher prediction success rate. This pattern is similar to that observed for a BP network with a single hidden layer. Furthermore, when the neural network has two hidden layers, the prediction accuracy obtained using training samples is higher. Increasing the number of neurons in each hidden layer increases the prediction error of the network for test samples. In other words, when the number of neurons is too large, overfitting occurs, resulting in reduced generalization ability of the network. The prediction performance of the network on both training and test samples is comprehensively considered to select the prediction network with the highest accuracy.

[0122] By integrating prediction networks with different network structures such as single hidden layer and double hidden layer, the ten prediction networks with the highest relative accuracy were selected as sub-networks of the voting mechanism to screen task stars with a veto mechanism. The prediction accuracy of the above 1563 samples was 95.27%, and the probability that the predicted task stars can actually execute tasks reached 98.52%. Compared with selecting the optimal BP neural network, although the prediction accuracy is not the highest, the probability that the predicted task stars can actually execute tasks is significantly improved, proving that the network can guarantee the mission success rate of the predicted task stars in response to the problem of relatively sufficient observation resources.

[0123] Optimal mission star prediction network

[0124] 6,238 samples were randomly divided into two groups: training data and test data, accounting for 80% and 20% of the total sample, respectively. During the training of the task star prediction network, it was found that the dual-hidden layer network achieved higher prediction accuracy than the single-hidden layer network. After training all candidate dual-hidden layer networks, the network performance was analyzed and compared. The absolute value of the task response time prediction error for different network structures is shown.

[0125] After comprehensively considering the prediction accuracy of training and test samples, 10 relatively optimal networks were selected as fault-tolerant subnetworks to predict the response time of the fault-tolerant task star. The average absolute value of the prediction error for the 6,238 samples was 13.56 seconds. For the test set samples, the distribution of the response time prediction error of the fault-tolerant network used in this invention was compared with the normal distribution, while the prediction error distribution of the optimal network was used.

[0126] For the test samples, the standard deviation of the response time error output by the fault-tolerant network is 21.79 seconds, while the standard deviation of the error output by the optimal network alone is 21.83 seconds. The error output by the fault-tolerant network is more concentrated around the mean. When using the fault-tolerant mechanism to predict response time, the probability that the output error falls within the 1-sigma interval is 80.21%. In other words, when using the fault-tolerant mechanism to predict the response time of a mission satellite within the next hour (3600 seconds), the confidence probability that the estimated error is within ±21.79 seconds is approximately 80.21%. Similar to the response time prediction method, the absolute value of the prediction error for different network structures is shown for the observation duration.

[0127] After comprehensively considering the prediction accuracy of training and test samples, 10 relatively optimal networks were selected as fault-tolerant subnetworks. The observation duration of mission satellites using the fault-tolerant mechanism was predicted. The average absolute value of the prediction error for these 6,238 samples was 2.43 seconds. For the test set samples, the distribution of the observation duration prediction error of the fault-tolerant network used in this invention was compared with a normal distribution, while the prediction error distribution of the optimal network was used.

[0128] For the test samples, the average error in the observation duration output by the fault-tolerant network was 0.42 seconds, with a standard deviation of 7.14 seconds. However, the average error output by the optimal network alone was 0.50 seconds, with a standard deviation of 7.26 seconds. Therefore, the use of a fault-tolerant mechanism can improve the prediction accuracy of the optimal network. When using the fault-tolerant mechanism to predict observation durations, the probability that the output error falls within the 1-sigma interval is 94.87%. In other words, when using the fault-tolerant mechanism to predict the observation duration of a mission satellite, the confidence level is approximately 94.87% that the error is within ±7.14 seconds.

[0129] Obviously, for certain precision-critical applications, this level of estimation error is unacceptable. However, due to the abundance of observation resources in ultra-large-scale constellations, it is necessary to first perform a preliminary screening of mission satellites. Based on this screening, the optimal mission satellite is predicted based on different requirements (optimal response time, optimal observation duration). In this step, the algorithm's accuracy in predicting the optimal mission satellite is not required. However, it is important to ensure that the predicted mission satellite can actually perform the mission and that the prediction of the optimal mission satellite is reliable to a certain extent. Clearly, the method proposed in this paper can meet the prediction requirements of mission satellites in ultra-large-scale remote sensing constellations.

[0130] Based on the rapid screening of mission stars for ultra-large-scale constellations, this paper proposes a new method for mission star selection and optimal observation star prediction based on a BP neural network. First, the input and output of the prediction network are determined to ensure that the main factors affecting mission execution are fully considered. Subsequently, a 28,800-satellite low-orbit remote sensing satellite constellation is generated based on the existing distribution characteristics of low-orbit remote sensing satellites. Network training and testing samples are generated, completing the prediction network training with fault-tolerant and one-vote veto mechanisms. Finally, a comprehensive analysis and evaluation of the trained prediction network's performance in mission star selection and optimal star prediction are conducted. The results show that while the BP network cannot accurately select mission stars, when used for mission star selection and optimal star prediction for ultra-large-scale constellations, it can ensure that the predicted stars are actually capable of executing the mission. This provides a certain degree of reliability in the optimal star prediction results, effectively reducing the solution space, lowering the computational load, and improving timeliness, thus meeting the demand for mission star prediction for ultra-large-scale remote sensing constellations. The method has broad application prospects.

[0131] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in an appropriate manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples, unless otherwise clearly defined. In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, features defined as "first" and "second" may explicitly or implicitly include at least one of such features. In the description of the present invention, "N" means at least two, such as two, three, etc., unless otherwise clearly defined. Any process or method description in a flowchart or otherwise described herein can be understood to represent a module, segment, or portion of code comprising one or more executable instructions for implementing a custom logic function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed in a different order than shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which the embodiments of the present invention pertain. The logic and / or steps shown in a flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logic function, and can be embodied in any computer-readable medium for use by an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device), or for use in conjunction with such instruction execution systems, apparatuses, or devices. For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by an instruction execution system, apparatus, or device, or in conjunction with such instruction execution systems, apparatuses, or devices. More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection having one or N wirings (electronic devices), a portable computer disk cartridge (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and portable compact disc read-only memory (CDROM).In addition, the computer-readable medium may even be paper or other suitable medium on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, then editing, interpreting, or processing in other suitable ways as necessary, and then storing it in a computer memory. It should be understood that the various parts of the present invention can be implemented with hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented with software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented with hardware, as in another embodiment, any one of the following technologies known in the art or their combination can be used to implement: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0132] The above description is merely a preferred embodiment of a method for predicting optimal mission stars for a very large-scale remote sensing constellation based on a BP neural network. The scope of protection for a method for predicting optimal mission stars for a very large-scale remote sensing constellation based on a BP neural network is not limited to the aforementioned embodiment. All technical solutions based on this concept fall within the scope of protection of the present invention. It should be noted that improvements and variations that do not depart from the principles of the present invention, as readily apparent to those skilled in the art, should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the optimal mission star for a large-scale remote sensing constellation based on a BP neural network, characterized by: The following steps are involved: Step 1, establish BP neural network; Step 2: prepare training samples and train the BP neural network; Step 3: Based on the trained BP neural network, the remote sensing mission stars are predicted; The step 1 is specifically as follows: Determine network inputs and outputs, The input is: There are 9 parameters in total, among which, is the semi-major axis of the initial orbit, is the initial orbit eccentricity, Julian day is used to issue tasks. is the initial orbital inclination, is the initial argument of perigee, is the initial ascending node right ascension, is the initial true anomaly, is the latitude of the mission target point, is the longitude of the target point; through the task completion indicator: response time and observation time , to predict the optimal task star. The output of the task star prediction BP network is clear, which are the remote sensing task response time and the duration of continuous observation by remote sensing satellites ,The network input selects the observation situation of the low-orbit remote sensing satellite, which is determined by the initial orbit ,parameters, the position of the observation target and the ,illumination conditions.

2. The method according to claim 1, wherein: Determine the initial orbital inclination , initial argument of perigee , initial ascending node right ascension , initial true anomaly Considering the observation of the target point by remote sensing satellite j, we must first calculate its sub-satellite point position. The sub-satellite point position is determined by the satellite orbit position. At a certain time t, the longitude and latitude of the sub-satellite point Calculated by the following formula: in, is the Greenwich mean sidereal time corresponding to the current moment; Determine the initial orbital semi-major axis and the initial orbital eccentricity The coverage of the sensor is described by the angle between the sight vector of the sensor field of view boundary and the satellite position vector in the geocentric inertial coordinate system. The earth is a standard sphere with a radius of , set all remote sensing satellite sensors to have a conical field of view, with a semi-cone angle of , assuming the satellite's geocentric distance is , calculated by the following formula: in, is the geocentric angle corresponding to the great circle arc between the intersection of the sensor line of sight and the earth's surface and the sub-satellite point of the remote sensing satellite. It means that the remote sensing satellite has geometric visibility to the target at the current moment; Determine the latitude of the mission target point , longitude of target point , according to the latitude and longitude of the mission target , Satellite subsatellite location Then we can calculate , and then determine whether the remote sensing satellite has geometric visibility with the target; Determine the Julian day when the task is issued , using the mission issuance time as the initial time t = 0, using the DE430 ephemeris published by the Jet Propulsion Laboratory of the United States, calculate the sun's position at any time t in the next hour; based on the sun's position, the remote sensing satellite position and the target position, it can be calculated whether the remote sensing payload meets the lighting conditions.

3. The method according to claim 2, wherein: The hidden layer structure of the neural network was determined, and the optimal task star was predicted. The network parameter range of the candidate BP network was determined using the empirical formula and the optimal network selection method. Ten sets of optimal and suboptimal networks were selected in parallel to fit the prediction results to ensure the accuracy of the prediction results. The activation function used is the classic sigmoid function, which is expressed as follows: 。 4. The method according to claim 3, wherein: The step 2 is specifically as follows: Prepare training samples, generate a low-orbit remote sensing satellite constellation of 28,800 satellites based on the distribution characteristics of existing low-orbit remote sensing satellites, and use January 1, 2024 as the constellation control base time and the task issuance time Randomly generated within one day before and after the benchmark time, the approximate area of ​​the time-sensitive remote sensing mission distribution is determined based on the distribution of the generated ultra-large-scale remote sensing constellation. The time-sensitive mission target points within this area Randomly generate and calculate time-sensitive tasks After the time is issued, the initial orbit parameters are Remote sensing satellite j will detect the target within the next hour Visible situation , if visible within the next hour , otherwise 0; For the task response time prediction network and the observation duration prediction network, the training sample is regarded as an array containing a pair of network input and network output, and the output is the response time and observation time ,Considering that the prediction of time parameters is relatively ,complex, multiple groups of different mission stars corresponding to ,different observation tasks are selected, with a total of 6238 groups of samples for ,training and analysis.

5. The method according to claim 4, wherein: A fault-tolerant computing mechanism is used to ensure the accuracy of the prediction of the optimal task star. When optimizing the task star, the prediction results of each sub-network are fully considered, and the prediction results of all networks are averaged to reduce the prediction error.

6. The method according to claim 5, wherein: The step 3 is specifically as follows: An optimal mission star prediction network was established. The network randomly divided 6,238 samples into two groups: training data and test data, accounting for 80% and 20% of the total samples, respectively. During the training process of the mission star prediction network, it was found that the double hidden layer had higher prediction accuracy than the single hidden layer. After training all candidate double hidden layer networks, the network performance was analyzed and compared.

7. A BP neural network-based system for predicting optimal mission stars in a large-scale remote sensing constellation, characterized by: The system comprises: A network building module, wherein the network building module builds a BP neural network; A sample module, wherein the sample module prepares training samples and trains the BP neural network; A prediction module, which predicts remote sensing mission stars based on a trained BP neural network; Identify network inputs and outputs, The input is: There are 9 parameters in total, among which, is the semi-major axis of the initial orbit, is the initial orbit eccentricity, Julian day is used to issue tasks. is the initial orbital inclination, is the initial argument of perigee, is the initial ascending node right ascension, is the initial true anomaly, is the latitude of the mission target point, is the longitude of the target point; through the task completion indicator: response time and observation time , to predict the optimal task star. The output of the task star prediction BP network is clear, which are the remote sensing task response time and the duration of continuous observation by remote sensing satellites ,The network input selects the observation situation of the low-orbit remote sensing satellite, which is determined by the initial orbit ,parameters, the position of the observation target and the ,illumination conditions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the method according to any one of claims 1 to 7.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

Images