A method and system for rapid detection of collision risk of large space targets

Abstract

The application discloses a large-scale space target collision risk rapid detection method and system. The method comprises the following steps: firstly, a binary classifier composed of a convolutional neural network and a feedforward neural network is used to rapidly infer a multi-epoch orbit feature sequence of a target pair, and a risk score is output; target pairs lower than a safety threshold are filtered, and the remaining high-risk candidate pairs are sorted and then sent to high-precision Monte Carlo simulation to obtain accurate collision probability and generate a hierarchical early warning. The application adopts a two-stage funnel architecture of deep learning and Monte Carlo simulation, adopts Focal Loss and positive sample weighting processing in training to deal with extremely unbalanced data, and realizes closed-loop incremental learning through simulation result backflow, which can not only significantly reduce the consumption of computing resources, maintain an extremely low false negative rate, but also decouple the inference speed from the catalog scale, and has the ability to expand to a super large-scale catalog.

Description

Technical Field

[0001] This invention relates to the field of space target collision early warning technology, and in particular to a method and system for rapid detection of large-scale space target collision risks. Background Technology

[0002] With the increasing frequency of human space activities, the number of cataloged targets in low Earth orbit has exceeded 35,000 and is still growing rapidly. In space target collision early warning, it is necessary to conduct collision risk assessments (all-vs-all) between any two cataloged targets. If high-precision Monte Carlo simulations are used to perform detailed calculations on each target pair, the computational complexity is O(n²), resulting in approximately 162 million potential collision pairs for 18,000 targets—a burden that current computing power cannot handle. Therefore, the industry generally adopts a pre-screening mechanism: first, a fast but coarse method is used to filter out the vast majority of safe target pairs, and then limited computing resources are concentrated on the detailed analysis of high-risk events.

[0003] Existing pre-screening technologies can be mainly classified into the following categories:

[0004] (1) Approach event pre-screening based on geometric filtering. For example, patent CN114327919B discloses a space target collision early warning method based on TLE elements and SGP4 model. It calculates the minimum approach distance between any two targets and uses a preset distance threshold (e.g., 5km) for geometric coarse screening. This method relies on a simplified two-body dynamics assumption and ignores the long-term effects of higher-order perturbations (such as changes in atmospheric drag, lunar and solar gravity) on the orbit, which can easily lead to the omission of real risks caused by perturbation offsets (false alarms). In addition, the distance threshold depends on manual setting and lacks adaptability. If the threshold is too large, it will lead to too many false alarms, and if the threshold is too small, it will increase the risk of false alarms.

[0005] (2) Rendezvous screening based on AI-assisted orbit propagation. In 2024, Petrov, Vanslette et al. proposed using neural networks to accelerate orbit propagation and then making distance threshold judgments based on predicted positions. Although this method accelerates single orbit extrapolation, pre-screening still requires performing AI orbit propagation + distance judgment on all target pairs, which cannot fundamentally change the O(n²) evaluation framework and still relies on distance thresholds, making it unable to learn hidden risk patterns.

[0006] (3) Observation scheduling based on traditional optimization methods. Patent CN120952470A uses an improved ant colony algorithm and branch and bound method to solve the multi-satellite collaborative observation scheduling problem. It belongs to the category of operations research optimization, does not involve machine learning, and is geared towards Earth observation scenarios rather than collision warning and pre-screening.

[0007] In summary, existing technologies have the following drawbacks: high risk of missed detection, strong threshold dependence and lack of adaptability, inability to utilize hidden correlations in historical data, and quadratic coupling of computational load with the number of targets, making it difficult to meet the real-time early warning needs of a full catalog of hundreds of thousands of targets in the future era of giant constellations. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes a rapid detection method and system for large-scale space target collision risks.

[0009] The main contents of this invention include:

[0010] A rapid method for detecting collision risks of large-scale space targets includes the following steps:

[0011] S1: Obtain the orbital root count of all on-orbit targets in the catalog database and generate a full list of target pairs;

[0012] S2: Input the orbital feature sequence of each target pair in the full target pair list into a pre-trained deep learning binary classifier. The deep learning binary classifier outputs the risk score of the corresponding target pair, which represents the probability that the corresponding target pair has a collision risk.

[0013] The orbit feature sequence refers to an ordered set of one or more feature vectors extracted from the orbit data of the corresponding target pair to characterize their relative motion state.

[0014] Step S3: Compare the risk score with a preset safety threshold, mark target pairs with risk scores lower than the safety threshold as safe target pairs and exclude them, and mark target pairs with risk scores higher than or equal to the safety threshold as high-risk candidate pairs;

[0015] Step S4: Perform Monte Carlo simulation of the complete perturbation model on the high-risk candidate pairs to obtain the collision probability of each high-risk candidate pair;

[0016] Step S5: Generate a graded collision warning report based on the collision probability.

[0017] Preferably, the orbital feature sequence includes feature vectors extracted from multiple forecast epochs, and the feature vectors are stacked in chronological order to form a temporal feature matrix. The multiple forecast epochs include the current epoch and at least two future epochs.

[0018] Preferably, the feature vector for each forecast epoch includes at least a portion of the following features: semi-major axis difference, eccentricity difference, inclination difference, ascending node right ascension difference, perigee amplitude difference, mean perigee amplitude difference, radial position difference, tangential position difference, normal position difference, radial velocity difference, tangential velocity difference, normal velocity difference, minimum approach distance analytical estimate, and relative velocity modulus.

[0019] Preferably, the deep learning binary classifier includes:

[0020] A one-dimensional convolutional neural network encoder is used to extract local spatial features from an input sequence of orbital features.

[0021] A fully connected neural network classifier, connected after the one-dimensional convolutional neural network encoder, is used to output the risk score based on the extracted features;

[0022] The Sigmoid output layer is used to map the output to the [0,1] interval as a risk score.

[0023] Preferably, the deep learning binary classifier is trained in the following manner:

[0024] Select target pairs from the catalog database, and perform Monte Carlo simulation using the complete perturbation model for each selected target pair. Use whether the collision probability output by the simulation exceeds the preset risk threshold as the standard for labeling positive and negative samples.

[0025] Focal Loss is used as the loss function, and positive sample weight coefficients and focusing parameters are set.

[0026] The Adam optimizer was used to optimize the model parameters.

[0027] Preferably, the Monte Carlo simulation employs a complete perturbation model including higher-order gravity fields, atmospheric drag, solar radiation pressure, and third-body gravity, and the number of simulated target shots is no less than [number missing]. Second-rate.

[0028] Preferably, the method further includes a threshold tuning step: after the deep learning binary classifier is trained, multiple candidate security thresholds are traversed using an independent validation set, and the threshold with the highest security pair filtering rate that satisfies the condition of a false negative rate ≤ 0.1% is selected as the preset security threshold.

[0029] Preferably, before performing the Monte Carlo simulation in S4, the method further includes: sorting the high-risk candidate pairs in descending order according to their risk scores to form a high-risk priority queue; and performing the Monte Carlo simulation sequentially according to the order of the high-risk priority queue.

[0030] Preferably, the method further includes a closed-loop optimization step: the collision probability results obtained from the Monte Carlo simulation in step S4, together with the corresponding target pair orbital features, are fed back into the training dataset as new labeled data, and the deep learning binary classifier is incrementally fine-tuned periodically or triggered.

[0031] This invention also proposes a rapid detection system for large-scale space target collision risks, comprising:

[0032] The data acquisition module is configured to acquire the orbital root counts of all on-orbit targets in the catalog database and generate a full list of target pairs.

[0033] The deep learning pre-screening module is configured to input the orbital feature sequence of each target pair into a pre-trained deep learning binary classifier and output the risk score of the target pair.

[0034] The threshold filtering module is configured to compare the risk score with a preset safety threshold, mark target pairs with risk scores lower than the safety threshold as safe target pairs and exclude them, and mark target pairs with risk scores higher than or equal to the safety threshold as high-risk candidate pairs.

[0035] The high-precision simulation module is configured to perform Monte Carlo simulation on the high-risk candidate pairs to obtain the collision probability of each high-risk candidate pair;

[0036] The early warning report generation module is configured to generate graded collision early warning reports based on the collision probability.

[0037] Preferably, the system further includes a queue sorting module, configured to sort the high-risk candidate pairs in descending order of risk score before the high-precision simulation module is executed, forming a high-risk priority queue; the high-precision simulation module is executed sequentially according to the order of the high-risk priority queue.

[0038] Preferably, it also includes a closed-loop feedback module, configured to feed back the collision probability results output by the high-precision simulation module, together with the corresponding target pair orbital features, as new labeled data into the training dataset; and an incremental learning module, configured to periodically or triggerically use the updated training dataset to incrementally fine-tune the deep learning binary classifier.

[0039] Compared with existing technologies, this application has the following advantages: It automatically learns collision risk patterns from massive historical simulation data through deep learning, enabling it to capture complex nonlinear hazardous scenarios that are difficult to identify using geometric filtering, thus greatly reducing the false negative rate; it adopts a funnel architecture of "AI pre-screening + fine screening," performing expensive Monte Carlo simulations only on a very small proportion of high-risk candidate pairs after pre-screening, significantly reducing computational resource consumption; furthermore, it continuously updates the training set and performs incremental fine-tuning using Monte Carlo simulation results through a closed-loop optimization mechanism, enabling the pre-screening model to adapt to dynamic changes in the orbital environment and the continuous growth of cataloged targets; finally, it systematically adopts a conservative strategy through Focal Loss, positive sample weighting, and threshold optimization, preferring to label more false alarms rather than miss real high-risk events, further ensuring safety. Attached Figure Description

[0040] Figure 1 This is a block diagram of the module structure of the system of the present invention;

[0041] Figure 2 This is an overall flowchart of the method of the present invention;

[0042] Figure 3 This is a schematic diagram illustrating the construction of multi-epoch time-series features according to the present invention. Detailed Implementation

[0043] The technical solution protected by this invention will be described in detail below with reference to the accompanying drawings.

[0044] Example 1

[0045] This invention provides a rapid method for detecting collision risks of large-scale space targets. This method decomposes the collision risk assessment of all target pairs into two stages: "deep learning pre-screening" and "Monte Carlo fine screening". The deep learning binary classifier quickly filters out most safe target pairs, and only performs high-precision simulation on a small number of high-risk candidate pairs, thereby achieving a balance between computational efficiency and safety.

[0046] S1: Get the full list of target pairs.

[0047] Obtain the orbital roots of all on-orbit targets from a space target cataloging database (such as the TLE roots provided by Space-Track). If the total number of cataloged targets is N, the generated list of full target pairs will contain M = N(N-1) / 2 unordered pairs. This list serves as input for subsequent pre-screening.

[0048] S2: Construct the orbital feature sequence and input it into a deep learning binary classifier.

[0049] (1) For each target pair, multi-epoch feature extraction is performed first. The specific method is as follows:

[0050] Select K forecast epochs (K≥3, e.g., the current epoch and at least two future epochs at equal intervals), and use the SGP4 model to extrapolate the TLE roots of each target to each epoch, obtaining the position and velocity vectors at each epoch. At each epoch, calculate the following eigenvectors:

[0051] (a) Differences in the six orbital elements: semi-major axis difference Δa, eccentricity difference Δe, inclination difference Δi, right ascension difference of ascending node ΔΩ, perigee argument difference Δω, and mean perigee angle difference ΔM;

[0052] (b) Components of relative position in the RTN coordinate system: radial difference ΔR, tangential difference ΔT, normal difference ΔN; Components of relative velocity in the RTN coordinate system: ΔV_R, ΔV_T, ΔV_N;

[0053] (c) Analytical estimate of minimum approach distance d_min (MOID approximation based on the two-body assumption); relative velocity modulus |ΔV|.

[0054] (2) Normalize the above features (e.g., Z-score standardization), and then stack the feature vectors of K epochs in chronological order to form a time-series feature matrix of size K×D (D is the feature dimension, usually 30~50). This matrix is ​​the "orbit feature sequence".

[0055] (3) Input the orbital feature sequence into a pre-trained deep learning binary classifier. This classifier consists of a one-dimensional convolutional neural network (CNN) encoder and a fully connected neural network (FNN) classifier connected in series. The CNN encoder contains 3 to 5 one-dimensional convolutional layers with kernel sizes of 3 or 5 and channel numbers of 64, 128, and 256 respectively. Each convolutional layer is followed by batch normalization and the rectified linear unit (ReLU) activation function. The output of the last convolutional layer is followed by a global average pooling layer. The FNN classifier contains 2 to 3 fully connected layers with hidden units of 256 and 128 respectively. The ReLU activation function is used, and the output layer has one neuron. The Sigmoid function is used to map the output to the [0,1] interval to obtain the risk score. . The closer it is to 1, the greater the likelihood that the target poses a collision risk.

[0056] S3: Threshold filtering.

[0057] Risk score With the preset safety threshold Compare. If If the target pair is not found in the previous step, it is marked as a safe target pair and excluded, and will not proceed to the next stage of screening; if If the pair is high-risk, it is marked as a high-risk candidate pair.

[0058] S4: Ranking of high-risk candidate pairs and high-precision Monte Carlo simulation.

[0059] All high-risk candidate pairs are sorted in descending order of risk score to form a high-risk priority queue. Then, Monte Carlo simulations using a complete perturbation model are performed on each target pair in the queue. The simulation model includes at least: a higher-order Earth gravity field (e.g., EGM2008 70×70 order), atmospheric drag (NRLMSISE-00 model), solar radiation pressure (considering Earth's shadow), and gravitational forces from the Sun, Moon, and other third bodies. Initial position / velocity uncertainties are set according to the TLE nominal accuracy (position error 1 km, velocity error 1 m / s, 3σ). The trajectory extrapolation for each firing exercise terminates at the predicted closest approach time (TCA), which is determined by iteratively refining the distance between targets over time (using three-point interpolation). The number of simulated firing exercises is no less than [number missing]. Next (standard mode) or (High-precision mode). The proportion of samples falling into the colliding sphere (the sum of the radii of the two targets) is statistically analyzed to obtain an estimate of the collision probability Pc. For extremely low Pc values ​​(below 10),... -5 In scenarios where importance sampling is required, the importance sampling technique can be used to improve estimation accuracy with a limited number of firing attempts.

[0060] S5: Generate tiered early warning reports.

[0061] Based on the Pc value, an early warning report is generated according to a preset grading standard. For example: Yellow warning. Orange alert Red alert The report includes the target identifier (NORAD catalog number), the expected closest approach time (TCA, accurate to the second), the estimated collision probability Pc, the minimum approach distance, the orbital state uncertainty covariance ellipsoid parameters of the two targets, and the alarm level indicator.

[0062] Furthermore, the training process of the deep learning binary classifier is as follows:

[0063] (1) Training data construction: Select a sufficient number of target pairs from the catalog database, such as through stratified sampling. For each selected target pair, perform a high-precision Monte Carlo simulation with the same configuration as S4 to obtain the collision probability Pc. Use a preset risk threshold (e.g., Using Pc > threshold as the standard, samples are labeled as positive samples (risky), and the rest are labeled as negative samples (safe). Since real collision events are extremely sparse, the ratio of positive to negative samples in the training set is usually on the order of 1:1000.

[0064] (2) Feature extraction: For each training sample, extract the multi-epoch time series feature matrix according to the method described in S2.

[0065] (3) Loss function: Focal Loss is adopted, and its expression is:

[0066]

[0067] in, The probability predicted by the model (for positive samples) For negative samples ), This is the weighting coefficient for positive samples (0.75 is used in an extremely imbalanced scenario where the ratio of positive to negative samples in the training set is approximately 1:1000, to give higher-risk samples greater weight; this differs from the default value of 0.25 used in the original Focal Loss paper for object detection tasks). This is the focusing parameter (usually set to 2.0). This loss function can reduce the contribution of a large number of easily classified negative samples, forcing the model to focus on difficult-to-classify boundary samples, thereby strengthening its ability to identify positive samples (which are at risk).

[0068] (4) Optimizer: The Adam optimizer is used, with an initial learning rate of 0.001, a batch size set according to the GPU memory (e.g., 10000), 50 training rounds, and early stopping on the validation set.

[0069] After training, threshold tuning is performed using a separate validation set, consisting of target pairs that were not used in training and their Monte Carlo labels. The specific steps include:

[0070] Iterate through the candidate thresholds τ, for example, 0.0001, 0.0002, 0.0005, 0.001, 0.002, 0.005, 0.01, and calculate the false negative rate (FNR) for each τ. The ratio) and Safety Filtering Rate (SFR, i.e., true safety and The ratio). Select τ, which maximizes SFR under the condition that FNR ≤ 0.1%, as the default safety threshold.

[0071] Furthermore, during system operation, the collision probability results obtained from each Monte Carlo simulation, along with the corresponding target pair orbital features, are fed back into the training dataset as new labeled data. The deep learning binary classifier is then periodically (e.g., weekly) or triggered incrementally fine-tuned using the accumulated new data (using a small learning rate, such as...). The model is trained in a small number of epochs; optionally, an Elastic Weight Consolidation (EWC) mechanism is employed, which further prevents catastrophic forgetting by adding a penalty term for historically important weights to the loss function. This mechanism enables the model to adapt to dynamic changes in the orbital environment (such as atmospheric density fluctuations caused by solar activity cycles) and the continuous growth of cataloged targets.

[0072] Corresponding to the above method, this invention also provides a rapid detection system for large-scale space target collision risks, comprising: a data acquisition module for reading a catalog database and generating a full list of target pairs; a feature calculation module for performing multi-epoch orbit extrapolation and feature extraction; a deep learning pre-screening module for loading a pre-trained model and outputting a risk score; a threshold filtering and sorting module for eliminating safe pairs and sorting high-risk candidate pairs; a high-precision Monte Carlo simulation module for performing target calculations using a complete perturbation model; a warning report generation module; and a training data management module for storing simulation results and triggering incremental model training. These modules can be deployed on a CPU+GPU heterogeneous computing platform, where feature calculation and deep learning inference are executed in parallel on the GPU, and Monte Carlo simulation is run in a distributed manner on a CPU cluster.

[0073] Example 2

[0074] This embodiment uses a low-orbit space target cataloging database as a specific application scenario and provides a detailed numerical implementation method.

[0075] I. Cataloging Scale and Initial Calculations

[0076] The catalog database contains approximately 18,000 on-orbit targets (standard TLE format). A complete list of target pairs is generated: total number of target pairs M = 18000 × 17999 / 2 ≈ 162 million.

[0077] II. Training Phase Configuration

[0078] (1) Training sample selection: 10 million pairs were selected from the full number of target pairs using stratified sampling (stratified by track height and inclination angle) as training samples.

[0079] (2) Label generation: High-precision Monte Carlo simulation was performed on each training sample. Simulation configuration: EGM2008 gravity field (70×70 order), NRLMSISE-00 atmospheric density model, solar radiation pressure model (considering the Earth's shadow), and gravitational force of the Sun, Moon, and other third bodies were used; number of shots = The initial position / velocity uncertainty is set to a position error of 1 km and a velocity error of 1 m / s (3σ) based on the TLE nominal accuracy. The collision probability is calculated using the simulation output. As a standard for labeling positive samples, the final training set contained approximately 10,000 positive samples and 9.99 million negative samples, with a positive-to-negative ratio of approximately 1:1000.

[0080] (3) Feature extraction: Five epochs were selected: the current epoch, T+6h, T+12h, T+18h, and T+24h. The position / velocity of each epoch was obtained by extrapolation using SGP4. For each epoch, 40-dimensional features were extracted: the difference of the six orbital roots (6-dimensional), the relative position of the RTN (3-dimensional), the relative velocity of the RTN (3-dimensional), the minimum approach distance (1-dimensional), the relative velocity magnitude (1-dimensional), and the first-order difference of the above features in adjacent epochs (26-dimensional). After Z-score normalization, a 5×40 temporal feature matrix was formed.

[0081] (4) Model architecture: The input layer receives a 5×40 matrix; the first layer is a Conv1D (kernel size 3, output channels 64, ReLU, same padding); the second layer is a Conv1D (kernel size 3, output channels 128, ReLU, same padding); the third layer is a Conv1D (kernel size 3, output channels 256, ReLU, same padding); the global average pooling layer outputs 256 dimensions; the first fully connected layer (256 neurons, ReLU, Dropout=0.3); the second fully connected layer (128 neurons, ReLU); the output layer (1 neuron, Sigmoid). The total number of parameters is approximately 800,000.

[0082] (5) Training hyperparameters: loss function Focal Loss (γ=2.0, α=0.75); optimizer Adam (learning rate=0.001); batch size=10000; number of training epochs=50; validation set is another 2 million pairs (independent sampling). Number of early stopping epochs=10 (to monitor the false negative rate of the validation set).

[0083] (6) Threshold tuning: Iterate through the validation set for τ = 0.0001, 0.0002, 0.0005, 0.001, 0.002, 0.005, 0.01. Calculate the FNR and SFR for each τ. The results are shown in the table below:

[0084] 0.0001 0.02% 97.5% 0.0002 0.04% 98.1% 0.0005 0.06% 98.8% 0.001 0.09% 99.1% 0.002 0.32% 99.5% 0.005 1.2% 99.8% 0.01 3.5% 99.9%

[0085] The default safety threshold is τ=0.001, which satisfies FNR ≤ 0.1% and has the largest SFR.

[0086] III. Reasoning Stage

[0087] For the full set of 162 million target pairs, feature extraction was performed sequentially (following the exact same process as the training phase), and then the data was fed into the trained model for forward inference. On a single NVIDIA A100 GPU, with a batch size of 10,000, the total inference time was approximately 1.2 seconds (162 million / 10,000 × approximately 0.074 ms per batch forward time). The risk score for each target pair was then obtained. .

[0088] Apply a threshold τ=0.001 for filtering: Target pairs with a value <0.001 were marked as safe and excluded, totaling approximately 160.4 million pairs (filtering rate ≈99.0%); approximately 1.6 million high-risk candidate pairs remained. ).

[0089] 1.6 million high-risk candidate pairs were categorized as follows: Sort in descending order to form a high-risk queue.

[0090] IV. Fine Screening Stage

[0091] The same Monte Carlo simulation configuration as the training phase (complete perturbation model) is used. (Single target firing), but only on target pairs in the high-risk queue. To speed up the process, 256 CPU cores are used for parallel computation, with an average time of 0.5 seconds per target pair. Total screening time: 1.6 million × 0.5 seconds / 256 ≈ 3125 seconds ≈ 52 minutes. In practice, this can be compressed to within 30 minutes by further increasing the parallelism (e.g., using GPU clusters or more CPU cores).

[0092] Final statistics: Among 1.6 million high-risk candidate pairs, Pc > The target pairs are approximately 135, where Pc > Approximately 40 pairs, Pc > Approximately 9 pairs. The system automatically generates tiered early warning reports.

[0093] V. Performance Evaluation

[0094] Evaluate model performance on an independent test set (an additional 2 million pairs, never used in training and validation):

[0095] False negative rate (FNR) = (True high risk and (Number of cases <0.001) / (Total number of actual high-risk cases) = 0.048%;

[0096] Safety to Filtration Rate (SFR) = (True Safety and (Quantity <0.001) / (Total number of actual safe items) = 99.12%.

[0097] The end-to-end processing time (including feature extraction, model inference, sorting, and Monte Carlo screening) is approximately 31 minutes with a single GPU and a 256-core CPU configuration, which is much shorter than the typical orbital update cycle of 8 hours.

[0098] VI. Closed-loop incremental learning effect

[0099] The system ran continuously for one month, with incremental fine-tuning performed weekly. Each fine-tuning used Monte Carlo simulation results generated within the past 7 days (approximately 2 million pairs, including newly discovered hazardous events). A learning rate of 1e was used during fine-tuning. -4 The model was trained for 5 epochs. After four fine-tuning iterations, the false negative rate further decreased to 0.031% on the newly collected test set that month, while the safety pair filtering rate remained at 99.2%. This demonstrates that the closed-loop mechanism effectively improves the model's adaptability to changes in the orbital environment.

[0100] VII. Comparative Testing

[0101] To verify the efficiency advantage of this invention, it was compared with pure Monte Carlo full-scale simulation (without pre-screening) and traditional geometric filtering pre-screening methods. Test environment: NVIDIA A100 GPU + 256-core CPU, catalog size 18,000 targets, forecast duration 7 days. The results are shown in the table below:

[0102] Pure Monte Carlo full-scale simulation 162 million Unable to complete 0% Geometric filtering (5km threshold) pre-screening Approximately 15 million Approximately 8 hours Approximately 5.2% This invention (AI pre-screening + fine screening) Approximately 1.6 million Approximately 31 minutes 0.048%

[0103] As can be seen, this invention reduces the computation time from 8 hours for geometric filtering pre-screening to 31 minutes while ensuring an extremely low false negative rate (0.048%), and at the same time reduces the false negative rate by two orders of magnitude.

[0104] Those skilled in the art will understand that the above values ​​are merely examples. The following parameters can be adjusted according to actual engineering needs:

[0105] Safety threshold τ: For a more conservative scenario, it can be set to 0.0001 (lower false negative rate, but increased fine screening load); for a more aggressive scenario, it can be set to 0.005 (more computing power saved, but higher false negative rate).

[0106] Monte Carlo firing count: For high-value targets (such as manned spacecraft), this can be increased to [number missing]. For rapid screening, the number of tests can be reduced to 5,000.

[0107] Number of forecast epochs: can be increased to 7 or 9 to capture more subtle long-term trends, or reduced to 3 to improve calculation speed.

[0108] The number of CNN layers and channels can be adjusted between 2 and 5 layers and 64 and 512 channels, depending on the feature complexity and GPU computing power.

[0109] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A rapid detection method for collision risk of large-scale space targets, characterized in that, Includes the following steps: S1: Obtain the orbital root count of all on-orbit targets in the catalog database and generate a full list of target pairs; S2: Input the orbital feature sequence of each target pair in the full target pair list into a pre-trained deep learning binary classifier. The deep learning binary classifier outputs the risk score of the corresponding target pair, which represents the probability that the corresponding target pair has a collision risk. The orbit feature sequence refers to an ordered set of one or more feature vectors extracted from the orbit data of the corresponding target pair to characterize their relative motion state. Step S3: Compare the risk score with a preset safety threshold, mark target pairs with risk scores lower than the safety threshold as safe target pairs and exclude them, and mark target pairs with risk scores higher than or equal to the safety threshold as high-risk candidate pairs; Step S4: Perform Monte Carlo simulation of the complete perturbation model on the high-risk candidate pairs to obtain the collision probability of each high-risk candidate pair; Step S5: Generate a graded collision warning report based on the collision probability.

2. The method for rapid detection of collision risk of large-scale space targets according to claim 1, characterized in that, The orbital feature sequence includes feature vectors extracted from multiple forecast epochs, and the feature vectors are stacked in chronological order to form a temporal feature matrix. The multiple forecast epochs include the current epoch and at least two future epochs.

3. The method for rapid detection of collision risk of large-scale space targets according to claim 2, characterized in that, The eigenvectors for each forecast epoch include at least one of the following features: semi-major axis difference, eccentricity difference, inclination difference, ascending node right ascension difference, perihelion amplitude difference, mean perihelion amplitude difference, radial position difference, tangential position difference, normal position difference, radial velocity difference, tangential velocity difference, normal velocity difference, minimum approach distance analytical estimate, and relative velocity modulus.

4. The method for rapid detection of collision risk of large-scale space targets according to claim 1, characterized in that, The deep learning binary classifier includes: A one-dimensional convolutional neural network encoder is used to extract local spatial features from an input sequence of orbital features. A fully connected neural network classifier, connected after the one-dimensional convolutional neural network encoder, is used to output the risk score based on the extracted features; The Sigmoid output layer is used to map the output to the [0,1] interval as a risk score.

5. The method for rapid detection of collision risk of large-scale space targets according to claim 1, characterized in that, The deep learning binary classifier is trained in the following way: Select target pairs from the catalog database, and perform Monte Carlo simulation using the complete perturbation model for each selected target pair. Use whether the collision probability output by the simulation exceeds the preset risk threshold as the standard for labeling positive and negative samples. Focal Loss is used as the loss function, and positive sample weight coefficients and focusing parameters are set. The Adam optimizer was used to optimize the model parameters.

6. The method for rapid detection of collision risk of large-scale space targets according to claim 5, characterized in that, The Monte Carlo simulation employs a complete perturbation model including higher-order gravity fields, atmospheric drag, solar radiation pressure, and third-body gravity, and the number of simulated target shots is no less than [number missing]. Second-rate.

7. The method for rapid detection of collision risk of large-scale space targets according to claim 1, characterized in that, It also includes a threshold tuning step: after the deep learning binary classifier is trained, multiple candidate security thresholds are traversed using an independent validation set, and the threshold with the highest security pair filtering rate that satisfies the condition of a false negative rate ≤ 0.1% is selected as the preset security threshold.

8. The method for rapid detection of collision risk of large-scale space targets according to claim 1, characterized in that, Before executing the Monte Carlo simulation in S4, the process further includes: sorting the high-risk candidate pairs in descending order of their risk scores to form a high-risk priority queue; and executing the Monte Carlo simulation sequentially according to the order of the high-risk priority queue.

9. The method for rapid detection of collision risk of large-scale space targets according to claim 1, characterized in that, It also includes a closed-loop optimization step: the collision probability results obtained from the Monte Carlo simulation in step S4, together with the corresponding target pair orbital features, are fed back into the training dataset as new labeled data, and the deep learning binary classifier is incrementally fine-tuned periodically or triggered.

10. A rapid detection system for collision risk of large-scale space targets, characterized in that, include: The data acquisition module is configured to acquire the orbital root counts of all on-orbit targets in the catalog database and generate a full list of target pairs. The deep learning pre-screening module is configured to input the orbital feature sequence of each target pair into a pre-trained deep learning binary classifier and output the risk score of the target pair. The threshold filtering module is configured to compare the risk score with a preset safety threshold, mark target pairs with risk scores lower than the safety threshold as safe target pairs and exclude them, and mark target pairs with risk scores higher than or equal to the safety threshold as high-risk candidate pairs. The high-precision simulation module is configured to perform Monte Carlo simulation on the high-risk candidate pairs to obtain the collision probability of each high-risk candidate pair; The early warning report generation module is configured to generate graded collision early warning reports based on the collision probability.

11. A rapid detection system for large-scale space target collision risk according to claim 10, characterized in that, It also includes a queue sorting module, configured to sort the high-risk candidate pairs in descending order of risk score before the high-precision simulation module is executed, forming a high-risk priority queue; the high-precision simulation module is executed sequentially according to the order of the high-risk priority queue.

12. A rapid detection system for large-scale space target collision risk according to claim 10, characterized in that, It also includes a closed-loop feedback module, configured to feed back the collision probability results output by the high-precision simulation module, along with the corresponding target pair orbital features, as new labeled data into the training dataset; And an incremental learning module, configured to periodically or triggerically fine-tune the deep learning binary classifier using the updated training dataset.

Images