Space competition situation threat level assessment method and system based on credibility weighting

Abstract

The present invention provides a method and system for assessing the threat level of space competition situation based on credibility weighting, which relates to the field of information technology, including constructing a threat feature vector based on historical orbital data and operating status, and using a deep learning model to obtain an initial target threat probability vector. Then, the credibility weighting coefficient is calculated based on multi-source situational awareness data, and the initial threat probability vector is corrected in combination with the characteristic weight coefficient. Finally, an integrated learning method is used to classify the target threat level, and the target priority is determined in combination with the target orbital parameters and the ant colony algorithm, and a threat target sequence and a situation threat assessment report are output. By fusing multi-source situational awareness data and a credibility weighting mechanism, the present invention can more accurately assess the threat level of space targets and effectively improve the reliability and efficiency of space situational awareness.

Description

Technical Field

[0001] The present invention relates to information technology, and in particular to a method and system for assessing the threat level of space competition situation based on credibility weighting. Background Art

[0002] With the rapid increase in the number of spacecraft in orbit and the increasing complexity of space activities, how to quickly and accurately assess the potential threats to space targets and prioritize high-threat targets has become a key issue that needs to be urgently addressed in the field of space security. Traditional threat assessment methods usually rely on expert experience and rule reasoning, which are difficult to adapt to the complex and changing space environment and are inefficient. The existing technologies mainly have the following defects and shortcomings:

[0003] It is difficult to effectively fuse multi-source heterogeneous data. Existing space situational awareness data comes from a variety of sensors such as radar, optics, and radio. These data types are different, the formats are not uniform, and there are noise and errors. How to effectively fuse these multi-source heterogeneous data and extract valuable threat feature information is a major challenge facing existing technologies.

[0004] The accuracy and generalization ability of threat assessment models are insufficient. Traditional threat assessment models are usually based on simple rules or statistical methods, which are difficult to capture complex space target behavior patterns. In addition, due to the constant changes in the types and behavior patterns of space targets, traditional models have poor generalization ability and are difficult to adapt to new threat types.

[0005] Lack of refined classification and prioritization of target threat levels. Existing threat assessment methods can usually only roughly classify targets, and it is difficult to distinguish the difference in threat levels between different targets. In addition, there is a lack of an effective mechanism for prioritizing high-threat targets, which makes it difficult to guide the effective allocation of resources and the deployment of defense measures. Summary of the invention

[0006] The embodiments of the present invention provide a method and system for assessing the threat level of space competition situation based on credibility weighting, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provide a method for assessing the threat level of space competition based on credibility weighting, including:

[0009] Based on historical orbital data and the operating status of on-orbit satellites, a threat feature vector including maneuvering features, payload features and target features is constructed, wherein the maneuvering features include orbital maneuvering frequency, speed change and orbit adjustment amplitude, the payload features include mission payload type, power change and signal characteristics, and the target features include target size, attitude change and orbit type; a hierarchical analysis method is used to determine the feature index weight coefficients of maneuvering features, payload features and target features, and the threat feature vector is input into the encoder-decoder deep learning model, a multi-head attention layer is set between the encoder and the decoder, and a generative adversarial network training method is used to obtain an initial target threat probability vector;

[0010] Based on radar, optical and radio measurements, multi-source situational awareness data is collected, and the collection time interval, data volatility and measurement accuracy are input into the data quality assessment network to calculate the timeliness coefficient, accuracy coefficient and integrity coefficient. The timeliness coefficient, accuracy coefficient and integrity coefficient are integrated by a deep belief network to obtain the credibility weighted coefficient of the multi-source situational awareness data; based on the credibility weighted coefficient and the characteristic index weight coefficient, a modified weight matrix is ​​formed, and the modified weight matrix and the initial target threat probability vector are input into the variational Bayesian network to obtain a modified target threat probability vector and a confidence interval;

[0011] According to the corrected target threat probability vector and confidence interval, a base classifier set including random forest, gradient boosting decision tree and support vector machine is constructed. The classification results of the base classifier set are fused using the stacking integration strategy. The threat level of space targets is divided into three levels: high risk, medium risk and low risk. The initial value of the target priority is determined based on the threat level classification result. The constraint conditions of threat propagation are constructed in combination with the target orbit parameters. The ant colony algorithm is used to solve the constraint conditions and update the target priority. The threat target sequence and situation threat assessment report are output according to the target priority.

[0012] In an optional embodiment,

[0013] The step of determining the characteristic index weight coefficients of the maneuvering characteristics, the load characteristics and the target characteristics by using the hierarchical analysis method, inputting the threat characteristic vector into the encoder-decoder deep learning model, setting a multi-head attention layer between the encoder and the decoder, and using the generative adversarial network training method to obtain the initial target threat probability vector includes:

[0014] Constructing a judgment matrix of a three-layer hierarchical structure, the three-layer hierarchical structure includes a target layer, a criterion layer and a solution layer, wherein the target layer is a threat assessment result, the criterion layer includes maneuverability characteristics, load characteristics and target characteristics, and the solution layer is a quantitative index corresponding to each characteristic in the criterion layer, normalizing the judgment matrix to obtain a criterion layer weight vector, and calculating the weight of each quantitative index of the solution layer based on the criterion layer weight vector to obtain a characteristic index weight coefficient;

[0015] The product of the threat feature vector and the feature index weight coefficient is input into an encoder constructed by a three-layer bidirectional recurrent neural network, and residual connections and layer normalization modules are set between each layer of the encoder, and the encoder outputs a hidden state sequence; a parallel attention mechanism is constructed on the hidden state sequence, and the hidden state sequence is linearly transformed to obtain a query matrix, a key matrix and a value matrix, and the attention weight is calculated based on the query matrix and the key matrix, and the attention weight is weighted and summed with the value matrix to obtain an attention head output, and multiple attention head outputs are spliced ​​to obtain an enhanced feature sequence, wherein the parallel attention mechanism includes a local attention head for capturing local temporal associations, a global attention head for establishing long-range dependencies, and a cross attention head for feature interaction;

[0016] A generative adversarial network is constructed, wherein the generator of the generative adversarial network generates a candidate threat probability vector through a fully connected network, the candidate threat probability vector is input into the discriminator of the generative adversarial network for discrimination, the cross entropy loss is calculated based on the discrimination value output by the discriminator, and the probability value constraint loss and the temporal consistency loss of the candidate threat probability vector are calculated at the same time, the cross entropy loss, the probability value constraint loss and the temporal consistency loss are used as training targets, and the generator and the discriminator are alternately trained until convergence; the enhanced feature sequence is input into the trained generative adversarial network to obtain an initial target threat probability vector.

[0017] In an optional embodiment,

[0018] The steps of calculating a cross entropy loss based on the discriminant value output by the discriminator, and simultaneously calculating a probability value constraint loss and a temporal consistency loss of the candidate threat probability vector, taking the cross entropy loss, the probability value constraint loss and the temporal consistency loss as training targets, and alternately training the generator and the discriminator until convergence include:

[0019] A cross entropy loss function is constructed based on the candidate threat probability vector and the historical real samples, the label value of the historical real sample is randomly perturbed to obtain a smooth label, the cross entropy loss is calculated based on the smooth label, a gradient norm constraint term of the discriminator output relative to the input is constructed as a gradient penalty term, and a weighted combination of the cross entropy loss function and the gradient penalty term is used as the discriminator loss function;

[0020] Performing probability constraint processing on the candidate threat probability vector to obtain a normalized probability vector, and constructing a probability value constraint loss function based on the normalized probability vector; constructing a temporal consistency loss function based on the normalized probability vector, wherein the temporal consistency loss function includes a probability difference constraint term at adjacent moments and a probability curve smoothing constraint term, introduces multiple time scales to construct a multi-scale temporal consistency loss, and performs a weighted combination on the losses of different time scales;

[0021] The discriminator loss function, the probability value constraint loss function and the temporal consistency loss function are used as the combined training loss function, the combined training loss function is input into the attention mechanism network, the temporal features are extracted based on the historical change trend of the combined training loss function, the correlation between the discriminator loss function, the probability value constraint loss function and the temporal consistency loss function is calculated through the self-attention layer to obtain the loss attention matrix, and an adaptive weight network is constructed based on the loss attention matrix, the adaptive weight network includes a fully connected layer and a gated recurrent unit, and is used to dynamically predict the optimal combined weight of the combined training loss function; the combined training loss function is weighted based on the predicted optimal combined weight to obtain the generator loss function;

[0022] Based on the generator loss function and the discriminator loss function, the network parameters of the discriminator and the generator are updated alternately until a preset training termination condition is met.

[0023] In an optional embodiment,

[0024] The steps of collecting multi-source situational awareness data based on radar, optical and radio measurements, inputting the collection time interval, data fluctuation rate and measurement accuracy into the data quality assessment network to calculate the timeliness coefficient, accuracy coefficient and integrity coefficient, and using the deep belief network to fuse the timeliness coefficient, accuracy coefficient and integrity coefficient to obtain the credibility weighted coefficient of the multi-source situational awareness data include:

[0025] The acquisition time interval is the time difference between two adjacent data acquisitions, the data fluctuation rate is the standard deviation of multiple consecutive measurement data, and the measurement accuracy is the measurement error index inherent in the equipment;

[0026] The collection time interval, data volatility and measurement accuracy are input into a data quality assessment network, which is a multi-layer perceptron structure and includes three hidden layers. The hidden layers use a ReLU activation function and residual connections are set between adjacent hidden layers. The output layer of the data quality assessment network includes a timeliness coefficient node, an accuracy coefficient node and an integrity coefficient node, and a sigmoid activation function is used to limit the output value within a preset interval. During the training process of the data quality assessment network, a batch stochastic gradient descent method is used for optimization, the mean square error is selected as the loss function, and a dynamic attenuation mechanism of the learning rate is introduced. A batch normalization layer is added to the middle layer of the data quality assessment network to convert the output of the middle layer into a standard normal distribution.

[0027] A deep belief network is constructed, wherein the deep belief network includes multiple restricted Boltzmann machine layers, the timeliness coefficient, the accuracy coefficient, the integrity coefficient and their first-order differences are input into the input layer of the deep belief network, and a sigmoid activation function is set at the output layer of the deep belief network; the dynamic importance weights of the timeliness coefficient, the accuracy coefficient and the integrity coefficient are calculated based on the attention mechanism in the deep belief network; the timeliness coefficient, the accuracy coefficient and the integrity coefficient are weightedly combined with the corresponding dynamic importance weights, and the credibility weighted coefficient of the multi-source situation awareness data is obtained through the sigmoid activation function.

[0028] In an optional embodiment,

[0029] The steps of forming a modified weight matrix based on the credibility weight coefficient and the characteristic index weight coefficient, inputting the modified weight matrix and the initial target threat probability vector into a variational Bayesian network, and obtaining a modified target threat probability vector and a confidence interval include:

[0030] Performing an outer product operation on the credibility weight coefficient vector and the characteristic index weight coefficient vector and performing L1 normalization on the operation result to obtain a modified weight matrix; constructing a variational Bayesian network including a latent variable layer and an observation variable layer, wherein the latent variable layer is used to model the approximate posterior distribution of the true threat probability, and the observation variable layer receives the modified weight matrix and the initial target threat probability vector as input, and adopts a diagonal Gaussian distribution as a variational distribution family in the variational Bayesian network;

[0031] The modified weight matrix is ​​multiplied by the initial threat probability vector to obtain a weighted feature vector, and a variational inference process is performed based on the weighted feature vector by minimizing a variational lower bound including a likelihood function term and a KL divergence term; in the variational inference process, a stochastic gradient variational Bayesian method is used to iteratively update the mean parameter and the standard deviation parameter of the latent variable, and the mean parameter and the standard deviation parameter are combined with a standard normal distribution sampling value through a parameter reconstruction method for gradient estimation to obtain an approximate posterior distribution of the true threat probability;

[0032] The value range of the approximate posterior distribution is divided into multiple probability subspaces, and importance sampling is performed in each probability subspace based on the probability value of the approximate posterior distribution corresponding to the subspace to obtain a sample set; the probability density function of the true threat probability is reconstructed for the sample set using a Gaussian kernel density estimation method with an adaptive bandwidth, wherein the adaptive bandwidth is determined by the sample standard deviation and the number of samples, and the quantiles at different confidence levels are calculated based on the reconstructed probability density function by a numerical integration method to obtain the upper and lower bounds of the confidence interval of the true threat probability;

[0033] A dynamic time window is constructed. In each dynamic time window, the approximate posterior distribution is recalculated based on the mean parameter and standard deviation parameter updated by the latest iteration. The bandwidth parameter of the Gaussian kernel function and the upper and lower bounds of the confidence interval of the true threat probability are updated. The boundaries of the confidence interval are obtained by solving the integral equation using the binary search method. The corrected target threat probability vector is output according to the mean parameter updated by the latest iteration, and the corresponding confidence interval is output based on the upper and lower bounds of the confidence interval updated by the latest iteration.

[0034] In an optional embodiment,

[0035] According to the corrected target threat probability vector and confidence interval, a base classifier set including random forest, gradient boosting decision tree and support vector machine is constructed, and the classification results of the base classifier set are fused by stacking integration strategy, and the threat level of the space target is divided into three levels: high risk, medium risk and low risk. The initial value of the target priority is determined based on the threat level classification result, and the constraint condition of threat propagation is constructed in combination with the target orbit parameter. The ant colony algorithm is used to solve the constraint condition and update the target priority. The steps of outputting the threat target sequence and the situation threat assessment report according to the target priority include:

[0036] The corrected target threat probability vector is combined with the confidence interval to form a multidimensional feature vector, and a base classifier set including a random forest classifier, a gradient boosted decision tree classifier, and a support vector machine classifier is constructed based on the multidimensional feature vector. The random forest classifier uses the Gini coefficient as a node splitting criterion and constructs a decision tree through bootstrap sampling. The gradient boosted decision tree classifier uses a logarithmic loss function and is optimized through the Newton-Raphson method. The support vector machine classifier uses a radial basis kernel function, and each base classifier outputs a probability distribution vector of the target belonging to each threat level;

[0037] The probability distribution vectors output by the base classifiers are concatenated to form a meta-feature vector, a lightweight gradient boosting machine is used as a secondary learner to stack and integrate the meta-feature vectors, a focal loss function is introduced to correct the category imbalance problem, and the category weights and focus parameters are determined through cross-validation to divide the threat levels of space targets into high risk, medium risk and low risk; based on the threat level division result and the corrected target threat probability vector, a priority scoring function including a threat level indication function and a threat probability is constructed, and an initial value of the target priority is obtained through normalization processing;

[0038] Based on the target orbit parameters, a simplified perturbation orbit propagation model is used to calculate the orbit intersection probability, and a threat propagation constraint condition set including a spatial distance function and a time window constraint is constructed. The spatial distance function is calculated based on the target position vector, and the time window constraint is determined by the threat judgment threshold. The target priority initial value is used as the initial pheromone distribution, and a heuristic function is constructed based on the threat propagation constraint condition set. In the iterative process of the ant colony algorithm, the state transition probability is calculated by combining the pheromone concentration and the heuristic function, and the pheromone distribution is updated by pheromone attenuation and the optimal solution increment to obtain the optimized target priority.

[0039] The threat targets are sorted according to the optimized target priorities, and a threat target sequence and a situation threat assessment report containing basic target information, threat assessment indicators, priority analysis and situation evolution prediction are generated.

[0040] In an optional embodiment,

[0041] The probability distribution vectors output by the base classifiers are concatenated to form a meta-feature vector, a lightweight gradient boosting machine is used as a secondary learner to stack and integrate the meta-feature vectors, a focal loss function is introduced to correct the class imbalance problem, and the class weights and focus parameters are determined by cross-validation. The steps of classifying the threat level of space targets into high risk, medium risk and low risk include:

[0042] A base classifier set consisting of a random forest classifier, a gradient boosting decision tree classifier, and a support vector machine classifier is used to classify and predict the space target, and each classifier outputs a first probability distribution vector that the space target belongs to three threat levels: high risk, medium risk, and low risk; the first probability distribution vectors are spliced ​​and combined to construct a multidimensional element feature vector containing the prediction probability information of all base classifiers;

[0043] A lightweight gradient boosting machine is used as a secondary learner to perform stacked ensemble learning on the multidimensional element feature vector to obtain a second probability distribution vector. The lightweight gradient boosting machine configures model parameters by controlling the complexity of the tree, the learning rate, the feature random sampling ratio, the data random sampling ratio and the minimum number of leaf node samples. A focus loss function is introduced to correct the category imbalance problem of the threat level in the second probability distribution vector. The focus loss function includes a category weight factor and a focus parameter.

[0044] The category weight factor and the focus parameter are optimized by a cross-validation method, a weighted F1 score is used as an evaluation index for the optimization of the category weight factor, and a harmonic mean of a weighted precision and a recall rate is used as an evaluation index for the optimization of the focus parameter, to obtain an optimized focus loss function;

[0045] The lightweight gradient boosting machine is trained based on the optimized focal loss function to obtain a final third probability distribution vector, and multi-category judgment is performed on the third probability distribution vector by setting a probability threshold to generate a final threat level classification result of the space target.

[0046] According to a second aspect of the embodiments of the present invention,

[0047] Provide a space competition situation threat level assessment system based on credibility weighting, including:

[0048] The first unit is used to construct a threat feature vector including maneuvering features, load features and target features based on historical orbit data and the operating status of on-orbit satellites, wherein the maneuvering features include orbital maneuvering frequency, speed change and orbit adjustment amplitude, the load features include mission load type, power change and signal characteristics, and the target features include target size, attitude change and orbit type; a hierarchical analysis method is used to determine the feature index weight coefficients of maneuvering features, load features and target features, the threat feature vector is input into an encoder-decoder deep learning model, a multi-head attention layer is set between the encoder and the decoder, and a generative adversarial network training method is used to obtain an initial target threat probability vector;

[0049] The second unit is used to collect multi-source situational awareness data based on radar, optical and radio measurements, input the collection time interval, data volatility and measurement accuracy into the data quality assessment network to calculate the timeliness coefficient, accuracy coefficient and integrity coefficient, and use the deep belief network to fuse the timeliness coefficient, accuracy coefficient and integrity coefficient to obtain the credibility weighted coefficient of the multi-source situational awareness data; based on the credibility weighted coefficient and the characteristic index weight coefficient, a modified weight matrix is ​​formed, and the modified weight matrix and the initial target threat probability vector are input into the variational Bayesian network to obtain a modified target threat probability vector and a confidence interval;

[0050] The third unit is used to construct a base classifier set including random forest, gradient boosting decision tree and support vector machine according to the corrected target threat probability vector and confidence interval, and adopts a stacking integration strategy to fuse the classification results of the base classifier set, so as to divide the threat level of space targets into three levels: high risk, medium risk and low risk. The initial value of the target priority is determined based on the threat level classification result, and the constraint conditions of threat propagation are constructed in combination with the target orbit parameters. The ant colony algorithm is used to solve the constraint conditions and update the target priority, and the threat target sequence and situation threat assessment report are output according to the target priority.

[0051] According to a third aspect of the embodiments of the present invention,

[0052] An electronic device is provided, comprising:

[0053] processor;

[0054] a memory for storing processor-executable instructions;

[0055] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0056] A fourth aspect of the embodiments of the present invention is:

[0057] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0058] The present invention combines multi-source situational awareness data and uses deep learning models and variational Bayesian networks for correction, which can more accurately assess the threat probability of space targets, reduce the misjudgment rate, and provide confidence intervals to enhance the credibility of the assessment results. The present invention calculates the timeliness, accuracy, and integrity coefficients through a data quality assessment network, and fuses them to obtain a credibility weighted coefficient, which can effectively reduce the impact of data noise and errors and improve the reliability and effectiveness of situational awareness data. The present invention fuses multiple base classifiers through a stacking integration strategy to achieve a more accurate threat level division, and combines target orbit parameters and ant colony algorithms for priority sorting, which can help users quickly identify high-risk targets, optimize threat disposal processes, and improve response efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flowchart of a method for assessing the threat level of space competition situation based on credibility weighting according to an embodiment of the present invention;

[0060] Figure 2 The figure is a schematic diagram of the structure of a space competition situation threat level assessment system based on credibility weighting according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0063] Figure 1 FIG. 1 is a flow chart of a method for assessing the threat level of space competition situation based on credibility weighting according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] S1. Based on historical orbital data and the operating status of on-orbit satellites, a threat feature vector including maneuvering features, payload features and target features is constructed, wherein the maneuvering features include orbital maneuvering frequency, velocity change and orbit adjustment amplitude, the payload features include mission payload type, power change and signal characteristics, and the target features include target size, attitude change and orbit type; a hierarchical analysis method is used to determine the feature index weight coefficients of maneuvering features, payload features and target features, and the threat feature vector is input into the encoder-decoder deep learning model, a multi-head attention layer is set between the encoder and the decoder, and a generative adversarial network training method is used to obtain an initial target threat probability vector;

[0065] S2. Collect multi-source situational awareness data based on radar, optical and radio measurements, input the collection time interval, data volatility and measurement accuracy into the data quality assessment network to calculate the timeliness coefficient, accuracy coefficient and integrity coefficient, and use the deep belief network to fuse the timeliness coefficient, accuracy coefficient and integrity coefficient to obtain the credibility weighted coefficient of the multi-source situational awareness data; based on the credibility weighted coefficient and the characteristic index weight coefficient, a revised weight matrix is ​​formed, and the revised weight matrix and the initial target threat probability vector are input into the variational Bayesian network to obtain the revised target threat probability vector and confidence interval;

[0066] S3. According to the corrected target threat probability vector and confidence interval, a base classifier set including random forest, gradient boosting decision tree and support vector machine is constructed, and the classification results of the base classifier set are fused by stacking integration strategy. The threat level of space targets is divided into three levels: high risk, medium risk and low risk. The initial value of target priority is determined based on the threat level classification result, and the constraint conditions of threat propagation are constructed in combination with the target orbit parameters. The ant colony algorithm is used to solve the constraint conditions and update the target priority. The threat target sequence and situation threat assessment report are output according to the target priority.

[0067] In an optional embodiment,

[0068] The step of determining the weight coefficients of the characteristic indicators of the maneuvering characteristics, the load characteristics and the target characteristics by using the hierarchical analysis method, inputting the threat characteristic vector into the encoder-decoder deep learning model, setting a multi-head attention layer between the encoder and the decoder, and using the generative adversarial network training method to obtain the initial target threat probability vector includes:

[0069] Constructing a judgment matrix of a three-layer hierarchical structure, the three-layer hierarchical structure includes a target layer, a criterion layer and a solution layer, wherein the target layer is a threat assessment result, the criterion layer includes maneuverability characteristics, load characteristics and target characteristics, and the solution layer is a quantitative index corresponding to each characteristic in the criterion layer, normalizing the judgment matrix to obtain a criterion layer weight vector, and calculating the weight of each quantitative index of the solution layer based on the criterion layer weight vector to obtain a characteristic index weight coefficient;

[0070] The product of the threat feature vector and the feature index weight coefficient is input into an encoder constructed by a three-layer bidirectional recurrent neural network, and residual connections and layer normalization modules are set between each layer of the encoder, and the encoder outputs a hidden state sequence; a parallel attention mechanism is constructed on the hidden state sequence, and the hidden state sequence is linearly transformed to obtain a query matrix, a key matrix and a value matrix, and the attention weight is calculated based on the query matrix and the key matrix, and the attention weight is weighted and summed with the value matrix to obtain an attention head output, and multiple attention head outputs are spliced ​​to obtain an enhanced feature sequence, wherein the parallel attention mechanism includes a local attention head for capturing local temporal associations, a global attention head for establishing long-range dependencies, and a cross attention head for feature interaction;

[0071] A generative adversarial network is constructed, wherein the generator of the generative adversarial network generates a candidate threat probability vector through a fully connected network, the candidate threat probability vector is input into the discriminator of the generative adversarial network for discrimination, the cross entropy loss is calculated based on the discrimination value output by the discriminator, and the probability value constraint loss and the temporal consistency loss of the candidate threat probability vector are calculated at the same time, the cross entropy loss, the probability value constraint loss and the temporal consistency loss are used as training targets, and the generator and the discriminator are alternately trained until convergence; the enhanced feature sequence is input into the trained generative adversarial network to obtain an initial target threat probability vector.

[0072] Exemplarily, first, a hierarchical model of threat assessment is constructed. The model is divided into three layers: target layer, criterion layer and solution layer. The target layer represents the final threat assessment result. The criterion layer includes three main factors that affect the target threat: maneuvering characteristics, load characteristics and target characteristics. The solution layer is a specific quantitative indicator for each feature in the criterion layer. For example, the solution layer under the maneuvering characteristics may include indicators such as speed, acceleration, and turning radius; the solution layer under the load characteristics may include indicators such as weapon type, number of weapons, and remaining fuel; the solution layer under the target characteristics may include indicators such as target type, target importance, and target distance.

[0073] Next, construct the judgment matrix and determine the weight coefficient of the characteristic index. Compare the elements of each layer in pairs to construct the judgment matrix. For example, at the criterion layer, it is necessary to compare the relative importance of the maneuvering feature, load feature and target feature, and express them with numerical values ​​to form a 3x3 matrix. Then, normalize the judgment matrix to obtain the weight vector of each feature at the criterion layer. For example, the weight of the maneuvering feature is 0.4, the weight of the load feature is 0.3, and the weight of the target feature is 0.3. According to the weight vector of the criterion layer, calculate the weight of each quantitative index at the scheme layer, and finally obtain the weight coefficient of the characteristic index. For example, the weight coefficient of the speed index is the weight of the maneuvering feature 0.4 multiplied by the weight of the speed index in the maneuvering feature, assuming it is 0.5, and the final weight coefficient of the speed index is 0.2. And so on, get the weight coefficients of all quantitative indicators. Assuming that the speed of target A is 100, the acceleration is 20, the target type is a fighter, the importance is high, and the distance is 500, multiply these values ​​by the corresponding weight coefficients, and you will get the weighted values ​​of each feature of target A.

[0074] Then, the threat feature vector is input into the encoder for feature extraction. The weighted values ​​of each feature of target A are combined into a threat feature vector. This vector is input into an encoder constructed by a three-layer bidirectional recurrent neural network. Residual connections and layer normalization modules are set between each layer of the encoder to improve the training effect and stability of the model. The encoder converts the input threat feature vector into a hidden state sequence, which contains a more advanced feature representation.

[0075] A multi-head attention layer is set between the encoder and decoder to enhance feature representation. A parallel attention mechanism is constructed on the hidden state sequence, which includes three types of attention heads: local attention head, global attention head, and cross attention head. The local attention head focuses on the correlation between adjacent moments in the feature vector; the global attention head focuses on the dependency of all moments in the feature vector; the cross attention head focuses on the interactive information between different features. The output of each attention head is obtained by calculating the query matrix, key matrix, and value matrix, and performing weighted summation. The outputs of multiple attention heads are spliced ​​together to obtain an enhanced feature sequence, which integrates different types of attention information and can more comprehensively represent the target threat.

[0076] Construct a generative adversarial network for threat probability prediction. The generative adversarial network consists of two parts: the generator and the discriminator. The generator generates a candidate threat probability vector through a fully connected network. The discriminator determines whether the vector is real or generated. By alternately training the generator and the discriminator, the model parameters are continuously optimized until the model converges. During the training process, cross entropy loss, probability value constraint loss, and temporal consistency loss are used as training targets. The probability value constraint loss ensures that the generated probability value is between 0 and 1. The temporal consistency loss ensures that the threat probability changes smoothly at adjacent moments.

[0077] Finally, the enhanced feature sequence is input into the trained generative adversarial network to obtain the initial target threat probability vector, which represents the threat level of the target at the current moment.

[0078] The present invention combines the hierarchical analysis method with the deep learning model, which can more effectively extract and utilize threat feature information, thereby improving the accuracy of threat assessment. The introduction of the multi-head attention mechanism and the generative adversarial network enables the model to better handle noise and uncertainty, enhances the robustness of the model, and can automatically learn feature representation and threat probability without human intervention, thereby improving the evaluation efficiency.

[0079] In an optional embodiment,

[0080] The steps of calculating a cross entropy loss based on the discriminant value output by the discriminator, and simultaneously calculating a probability value constraint loss and a temporal consistency loss of the candidate threat probability vector, taking the cross entropy loss, the probability value constraint loss and the temporal consistency loss as training targets, and alternately training the generator and the discriminator until convergence include:

[0081] A cross entropy loss function is constructed based on the candidate threat probability vector and the historical real samples, the label value of the historical real sample is randomly perturbed to obtain a smooth label, the cross entropy loss is calculated based on the smooth label, a gradient norm constraint term of the discriminator output relative to the input is constructed as a gradient penalty term, and a weighted combination of the cross entropy loss function and the gradient penalty term is used as the discriminator loss function;

[0082] Performing probability constraint processing on the candidate threat probability vector to obtain a normalized probability vector, and constructing a probability value constraint loss function based on the normalized probability vector; constructing a temporal consistency loss function based on the normalized probability vector, wherein the temporal consistency loss function includes a probability difference constraint term at adjacent moments and a probability curve smoothing constraint term, introduces multiple time scales to construct a multi-scale temporal consistency loss, and performs a weighted combination on the losses of different time scales;

[0083] The discriminator loss function, the probability value constraint loss function and the temporal consistency loss function are used as the combined training loss function, the combined training loss function is input into the attention mechanism network, the temporal features are extracted based on the historical change trend of the combined training loss function, the correlation between the discriminator loss function, the probability value constraint loss function and the temporal consistency loss function is calculated through the self-attention layer to obtain the loss attention matrix, and an adaptive weight network is constructed based on the loss attention matrix, the adaptive weight network includes a fully connected layer and a gated recurrent unit, and is used to dynamically predict the optimal combination weight of the combined training loss function; the combined training loss function is weighted based on the predicted optimal combination weight to obtain the generator loss function;

[0084] Based on the generator loss function and the discriminator loss function, the network parameters of the discriminator and the generator are updated alternately until a preset training termination condition is met.

[0085] For example, first, historical threat data is collected, including various types of threat events and their corresponding real threat level labels. These data will be used to train the generator and discriminator. For example, data on cyber attacks in the past year are collected, including information such as attack type, attack target, attack duration, and threat level labels evaluated by experts, such as low, medium, and high.

[0086] Next, initialize the generator and discriminator networks. The input of the generator network can be random noise or other feature vectors, and the output is a candidate threat probability vector, which represents the probability of different threat levels. The input of the discriminator network is the candidate threat probability vector or the label value of the real sample, and the output is a scalar value, which represents the probability that the input is a real sample.

[0087] Then, start training the generator and discriminator alternately. When training the discriminator, use historical real samples and candidate threat probability vectors generated by the generator as input. To prevent overfitting, randomly perturb the label values ​​of real samples, such as randomly adding or subtracting a small value from the label value within a certain range to obtain a smooth label. Calculate the cross entropy loss between the discriminator output and the smooth label. To further improve the robustness of the model, add a gradient penalty term to limit the gradient norm of the discriminator output relative to the input. The weighted combination of the cross entropy loss and the gradient penalty term is used as the discriminator loss function. For example, the weight of the cross entropy loss is 0.8 and the weight of the gradient penalty term is 0.2.

[0088] When training the generator, the generated candidate threat probability vector is first subjected to probability constraint processing, such as normalization to ensure that the sum of all probability values ​​is 1, and a normalized probability vector is obtained. Then, a probability value constraint loss function is constructed based on the normalized probability vector, such as calculating the distance between the probability vector and the uniform distribution, to encourage the generator to generate a more uniform probability distribution. At the same time, a temporal consistency loss function is constructed, which contains probability difference constraints at adjacent moments and probability curve smoothing constraints, such as calculating the Euclidean distance between probability vectors at adjacent moments and the second-order derivative of the probability vector in the time dimension. In order to capture the temporal characteristics of different time scales, multiple time scales are introduced, such as short-term, medium-term, and long-term, and the losses of different time scales are weighted and combined.

[0089] The discriminator loss function, probability value constraint loss function and temporal consistency loss function are used as the combined training loss function. The combined training loss function is input into the attention mechanism network. The network first extracts the temporal features based on the historical change trend of the combined training loss function. Then, the correlation between the discriminator loss function, the probability value constraint loss function and the temporal consistency loss function is calculated through the self-attention layer to obtain the loss attention matrix. An adaptive weight network is constructed based on the loss attention matrix, which includes a fully connected layer and a gated recurrent unit to dynamically predict the optimal combined weight of the combined training loss function. For example, in the early stage of training, more emphasis may be placed on the discriminator loss, while in the later stage of training, more emphasis may be placed on the temporal consistency loss.

[0090] Finally, the combined training loss function is weighted based on the predicted optimal combination weights to obtain the generator loss function. Based on the generator loss function and the discriminator loss function, the network parameters of the discriminator and the generator are updated alternately until the preset training termination conditions are met, such as reaching the maximum number of training rounds or the loss function converges.

[0091] By introducing probability value constraints and temporal consistency losses, the present invention can generate a threat probability vector that is more in line with the actual situation, thereby improving the accuracy of threat assessment. By alternately training the generator and the discriminator and using gradient penalty terms, the model can be effectively prevented from overfitting and the stability of the model can be improved. By introducing an attention mechanism network and an adaptive weight network, the weights of different loss functions can be dynamically adjusted, so that the model can better adapt to different data and scenarios.

[0092] In an optional embodiment,

[0093] The steps of collecting multi-source situational awareness data based on radar, optical and radio measurements, inputting the collection time interval, data fluctuation rate and measurement accuracy into the data quality assessment network to calculate the timeliness coefficient, accuracy coefficient and integrity coefficient, and using the deep belief network to fuse the timeliness coefficient, accuracy coefficient and integrity coefficient to obtain the credibility weighted coefficient of the multi-source situational awareness data include:

[0094] The acquisition time interval is the time difference between two adjacent data acquisitions, the data fluctuation rate is the standard deviation of multiple consecutive measurement data, and the measurement accuracy is the measurement error index inherent in the equipment;

[0095] The collection time interval, data volatility and measurement accuracy are input into a data quality assessment network, which is a multi-layer perceptron structure and includes three hidden layers. The hidden layers use a ReLU activation function and residual connections are set between adjacent hidden layers. The output layer of the data quality assessment network includes a timeliness coefficient node, an accuracy coefficient node and an integrity coefficient node, and a sigmoid activation function is used to limit the output value within a preset interval. During the training process of the data quality assessment network, a batch stochastic gradient descent method is used for optimization, the mean square error is selected as the loss function, and a dynamic attenuation mechanism of the learning rate is introduced. A batch normalization layer is added to the middle layer of the data quality assessment network to convert the output of the middle layer into a standard normal distribution.

[0096] A deep belief network is constructed, wherein the deep belief network includes multiple restricted Boltzmann machine layers, the timeliness coefficient, the accuracy coefficient, the integrity coefficient and their first-order differences are input into the input layer of the deep belief network, and a sigmoid activation function is set at the output layer of the deep belief network; the dynamic importance weights of the timeliness coefficient, the accuracy coefficient and the integrity coefficient are calculated based on the attention mechanism in the deep belief network; the timeliness coefficient, the accuracy coefficient and the integrity coefficient are weightedly combined with the corresponding dynamic importance weights, and the credibility weighted coefficient of the multi-source situation awareness data is obtained through the sigmoid activation function.

[0097] Exemplarily, first, situational awareness data is collected from multiple sensors such as radar, optics, and radio. The time of each data collection is recorded, and the time difference between two adjacent data collections is calculated as the collection time interval. For example, the radar collects the position data of target A at 10:00:00, and collects the position data of target A again at 10:00:05, then the collection time interval is 5 seconds. At the same time, the standard deviation of the continuous measurement data is calculated to characterize the data volatility. For example, the speed of target A is measured five times in a row, which is 100 m / s, 102 m / s, 98 m / s, 101 m / s, and 99 m / s, respectively. The speed volatility is the standard deviation of these speed values, which is about 1.58 m / s. In addition, the measurement accuracy of each sensor is obtained. The accuracy value is provided by the equipment manufacturer and represents the inherent measurement error index of the equipment. For example, the ranging accuracy of the radar is 10 meters, and the angle measurement accuracy of the optical camera is 0.1 degrees.

[0098] Next, the collection time interval, data volatility and measurement accuracy are input into the data quality assessment network. The network adopts a multi-layer perceptron structure and contains three hidden layers. Each hidden layer uses the ReLU activation function, and residual connections are set between adjacent hidden layers to enhance the learning ability of the network. The output layer of the network contains three nodes, which output the timeliness coefficient, accuracy coefficient and integrity coefficient respectively. The output layer uses the sigmoid activation function to limit the values ​​of these coefficients between 0 and 1. In the process of training the data quality assessment network, the batch stochastic gradient descent method is used for optimization, the mean square error is selected as the loss function, and the learning rate dynamic attenuation mechanism is introduced to improve the training efficiency and the generalization ability of the model. In order to avoid the disappearance or explosion of the gradient, a batch normalization layer is added to the middle layer of the network to convert the output of the middle layer into a standard normal distribution. For example, the input collection time interval is 5 seconds, the data volatility is 1.58 meters / second, and the measurement accuracy is 10 meters. After calculation by the data quality assessment network, the output timeliness coefficient is 0.9, the accuracy coefficient is 0.8, and the integrity coefficient is 0.7.

[0099] Then, a deep belief network is constructed, which contains multiple restricted Boltzmann machine layers. The timeliness coefficient, accuracy coefficient, integrity coefficient and their first-order differences (for example, the difference between the timeliness coefficient at the current moment and the timeliness coefficient at the previous moment) output by the data quality assessment network are input into the input layer of the deep belief network. The output layer of the deep belief network is also set with a sigmoid activation function. In the deep belief network, the dynamic importance weights of the timeliness coefficient, accuracy coefficient and integrity coefficient are calculated based on the attention mechanism. For example, at a specific moment, if the timeliness of the data is very important, the dynamic importance weight of the timeliness coefficient will be higher. Finally, the timeliness coefficient, accuracy coefficient and integrity coefficient are weighted and combined with their corresponding dynamic importance weights, and then the final credibility weighted coefficient of the multi-source situation awareness data is obtained through the sigmoid activation function, and the value range of this coefficient is also between 0 and 1. For example, if the timeliness coefficient is 0.9, its dynamic importance weight is 0.8; if the accuracy coefficient is 0.8, its dynamic importance weight is 0.7; if the integrity coefficient is 0.7, its dynamic importance weight is 0.6. Then the final credibility weight coefficient may be 0.85.

[0100] By fusing multi-source data and evaluating data quality, the present invention can effectively reduce the uncertainty and error of a single data source and improve the reliability of situational awareness results. By considering the timeliness of the data, the latest and most timely information can be used first, thereby enhancing the real-time nature of situational awareness and responding to changing environments more quickly. By evaluating the accuracy and completeness of the data, low-quality data can be identified and filtered out, thereby improving the accuracy of situational awareness results and better supporting decision-making.

[0101] In an optional embodiment,

[0102] The steps of forming a modified weight matrix based on the credibility weight coefficient and the characteristic index weight coefficient, inputting the modified weight matrix and the initial target threat probability vector into a variational Bayesian network, and obtaining a modified target threat probability vector and a confidence interval include:

[0103] Performing an outer product operation on the credibility weight coefficient vector and the characteristic index weight coefficient vector and performing L1 normalization on the operation result to obtain a modified weight matrix; constructing a variational Bayesian network including a latent variable layer and an observation variable layer, wherein the latent variable layer is used to model the approximate posterior distribution of the true threat probability, and the observation variable layer receives the modified weight matrix and the initial target threat probability vector as input, and adopts a diagonal Gaussian distribution as a variational distribution family in the variational Bayesian network;

[0104] The modified weight matrix is ​​multiplied by the initial threat probability vector to obtain a weighted feature vector, and a variational inference process is performed based on the weighted feature vector by minimizing a variational lower bound including a likelihood function term and a KL divergence term; in the variational inference process, a stochastic gradient variational Bayesian method is used to iteratively update the mean parameter and the standard deviation parameter of the latent variable, and the mean parameter and the standard deviation parameter are combined with a standard normal distribution sampling value through a parameter reconstruction method for gradient estimation to obtain an approximate posterior distribution of the true threat probability;

[0105] The value range of the approximate posterior distribution is divided into multiple probability subspaces, and importance sampling is performed in each probability subspace based on the probability value of the approximate posterior distribution corresponding to the subspace to obtain a sample set; the probability density function of the true threat probability is reconstructed for the sample set using a Gaussian kernel density estimation method with an adaptive bandwidth, wherein the adaptive bandwidth is determined by the sample standard deviation and the number of samples, and the quantiles at different confidence levels are calculated based on the reconstructed probability density function by a numerical integration method to obtain the upper and lower bounds of the confidence interval of the true threat probability;

[0106] A dynamic time window is constructed. In each dynamic time window, the approximate posterior distribution is recalculated based on the mean parameter and standard deviation parameter updated by the latest iteration. The bandwidth parameter of the Gaussian kernel function and the upper and lower bounds of the confidence interval of the true threat probability are updated. The boundaries of the confidence interval are obtained by solving the integral equation using the binary search method. The corrected target threat probability vector is output according to the mean parameter updated by the latest iteration, and the corresponding confidence interval is output based on the upper and lower bounds of the confidence interval updated by the latest iteration.

[0107] Exemplarily, first, obtain the expert's evaluation of the reliability of different information sources to form a credibility weighted coefficient vector. For example, assuming there are three information sources, and experts evaluate their reliability as 0.8, 0.5, and 0.7, respectively, then the credibility weighted coefficient vector is [0.8, 0.5, 0.7]. At the same time, obtain the expert's evaluation of the importance of different feature indicators to form a feature indicator weight coefficient vector. Assuming there are two feature indicators, and experts evaluate their importance as 0.6 and 0.4, respectively, then the feature indicator weight coefficient vector is [0.6, 0.4]. Perform a matrix multiplication-like operation on these two vectors, and divide the result by the sum of all elements to obtain a modified weight matrix. For example, the modified weight matrix in the above example is [[0.48 / 1.58, 0.32 / 1.58], [0.3 / 1.58, 0.2 / 1.58], [0.42 / 1.58, 0.28 / 1.58]].

[0108] Next, a variational Bayesian network is constructed. This network consists of two layers: a hidden variable layer and an observed variable layer. The hidden variable layer is used to simulate the distribution of the true threat probability, and the observed variable layer receives the modified weight matrix and the initial target threat probability vector as input. Assume that the initial target threat probability vector is [0.3, 0.6, 0.5]. The modified weight matrix and the initial target threat probability vector are operated similar to matrix multiplication to obtain the weighted feature vector.

[0109] Then, using the variational inference method, the mean and standard deviation parameters of the latent variables are iteratively updated by minimizing an objective function that includes a likelihood function term and a KL divergence term. In this process, a parameter reconstruction method is used to combine the mean and standard deviation parameters with values ​​sampled from a standard normal distribution for gradient estimation. After multiple iterations, an approximate posterior distribution of the true threat probability is obtained.

[0110] The value range of the approximate posterior distribution is divided into multiple small intervals. In each interval, importance sampling is performed based on the probability value of the approximate posterior distribution corresponding to the interval to obtain a sample set. Using these sample sets, the Gaussian kernel density estimation method with adaptive bandwidth is used to reconstruct the probability density function of the true threat probability. The adaptive bandwidth is determined according to the sample standard deviation and the number of samples. Based on the reconstructed probability density function, the quantiles at different confidence levels are calculated by the numerical integration method to obtain the upper and lower bounds of the confidence interval of the true threat probability.

[0111] Construct a dynamic time window. In each time window, recalculate the approximate posterior distribution based on the mean and standard deviation parameters updated by the latest iteration, update the bandwidth parameter of the Gaussian kernel function and the upper and lower bounds of the confidence interval of the true threat probability. The boundaries of the confidence interval are obtained by solving the integral equation using the binary search method. Output the corrected target threat probability vector based on the mean parameter updated by the latest iteration, and output the corresponding confidence interval based on the upper and lower bounds of the confidence interval updated by the latest iteration.

[0112] The present invention combines expert experience and data-driven methods to correct the initial target threat probability, effectively reducing the influence of subjective factors and improving the accuracy of threat assessment. Through the variational Bayesian network, not only can the corrected target threat probability vector be output, but also the corresponding confidence interval can be given, which quantifies the uncertainty of the assessment result and enhances the reliability of threat assessment. By adopting a dynamic time window mechanism, the target threat probability and confidence interval can be updated in real time according to the latest information, timely reflecting the changes in the target threat situation and improving the real-time performance of threat assessment.

[0113] In an optional embodiment,

[0114] According to the corrected target threat probability vector and confidence interval, a base classifier set including random forest, gradient boosting decision tree and support vector machine is constructed, and the classification results of the base classifier set are fused by stacking integration strategy, and the threat level of the space target is divided into three levels: high risk, medium risk and low risk. The initial value of the target priority is determined based on the threat level classification result, and the constraint condition of threat propagation is constructed in combination with the target orbit parameter. The ant colony algorithm is used to solve the constraint condition and update the target priority. The steps of outputting the threat target sequence and the situation threat assessment report according to the target priority include:

[0115] The corrected target threat probability vector is combined with the confidence interval to form a multidimensional feature vector, and a base classifier set including a random forest classifier, a gradient boosted decision tree classifier, and a support vector machine classifier is constructed based on the multidimensional feature vector. The random forest classifier uses the Gini coefficient as a node splitting criterion and constructs a decision tree through bootstrap sampling. The gradient boosted decision tree classifier uses a logarithmic loss function and is optimized through the Newton-Raphson method. The support vector machine classifier uses a radial basis kernel function, and each base classifier outputs a probability distribution vector of the target belonging to each threat level;

[0116] The probability distribution vectors output by the base classifiers are concatenated to form a meta-feature vector, a lightweight gradient boosting machine is used as a secondary learner to stack and integrate the meta-feature vectors, a focal loss function is introduced to correct the category imbalance problem, and the category weights and focus parameters are determined through cross-validation to divide the threat levels of space targets into high risk, medium risk and low risk; based on the threat level division result and the corrected target threat probability vector, a priority scoring function including a threat level indication function and a threat probability is constructed, and an initial value of the target priority is obtained through normalization processing;

[0117] Based on the target orbit parameters, a simplified perturbation orbit propagation model is used to calculate the orbit intersection probability, and a threat propagation constraint condition set including a spatial distance function and a time window constraint is constructed. The spatial distance function is calculated based on the target position vector, and the time window constraint is determined by the threat judgment threshold. The target priority initial value is used as the initial pheromone distribution, and a heuristic function is constructed based on the threat propagation constraint condition set. In the iterative process of the ant colony algorithm, the state transition probability is calculated by combining the pheromone concentration and the heuristic function, and the pheromone distribution is updated by pheromone attenuation and the optimal solution increment to obtain the optimized target priority.

[0118] The threat targets are sorted according to the optimized target priorities, and a threat target sequence and a situation threat assessment report containing basic target information, threat assessment indicators, priority analysis and situation evolution prediction are generated.

[0119] Exemplarily, obtain the modified target threat probability vector and confidence interval. Assume there are two targets, the threat probability of target A is 0.8, and the confidence interval is [0.7, 0.9]; the threat probability of target B is 0.6, and the confidence interval is [0.5, 0.7]. Combine the upper and lower bounds of the threat probability and the confidence interval to form a multidimensional feature vector. The feature vector of target A is [0.8, 0.7, 0.9], and the feature vector of target B is [0.6, 0.5, 0.7].

[0120] Construct a set of base classifiers based on multidimensional feature vectors. First, construct a random forest classifier. Set the number of decision trees to 100, use the Gini coefficient as the node splitting criterion, and use the bootstrap method for data sampling. Secondly, construct a gradient boosting decision tree classifier. Set the number of decision trees to 50, use the logarithmic loss function as the loss function, and use the Newton-Raphson method for optimization. Finally, construct a support vector machine classifier using the radial basis kernel function. The feature vectors of target A and target B are input into the three base classifiers respectively, and each base classifier outputs a probability distribution vector that the target belongs to the three levels of high risk, medium risk, and low risk. For example, the output of the random forest classifier for target A is [0.9, 0.05, 0.05], indicating that the probability of target A being judged as high risk, medium risk, and low risk is 0.9, 0.05, and 0.05, respectively.

[0121] The probability distribution vectors output by the base classifiers are concatenated to form a meta-feature vector. Assuming that the outputs of the three base classifiers for target A are [0.9, 0.05, 0.05], [0.8, 0.1, 0.1], and [0.95, 0.025, 0.025], the meta-feature vector of target A is obtained by concatenating these vectors: [0.9, 0.05, 0.05, 0.8, 0.1, 0.1, 0.95, 0.025, 0.025]. A lightweight gradient boosting machine is used as a secondary learner to stack and integrate the meta-feature vectors. The focal loss function is introduced to correct the class imbalance problem, and the class weights and focus parameters are determined by five-fold cross validation. Assume that target A is finally judged as high-risk and target B is judged as medium-risk.

[0122] Based on the threat level classification results and the corrected target threat probability vector, a priority scoring function is constructed. Assume that the indicator function of the threat level is: high risk is 3, medium risk is 2, and low risk is 1. The priority score of target A is 3×0.8=2.4, and the priority score of target B is 2×0.6=1.2. Normalize the priority scores to obtain the initial value of the target priority. Assume that the initial value of the priority of target A is 0.67, and the initial value of the priority of target B is 0.33.

[0123] The orbit intersection probability is calculated based on the target orbit parameters, and a set of threat propagation constraint conditions is constructed. Assume that the orbit intersection probability of target A and target B is 0.2. The spatial distance function is calculated based on the target position vector, and the time window constraint is determined by the threat judgment threshold. Assuming that the spatial distance between target A and target B is less than the threshold and they meet within the time window, the threat propagation constraint condition is satisfied.

[0124] The initial value of the target priority is used as the initial distribution of pheromones, and a heuristic function is constructed based on the set of threat propagation constraints. It is assumed that the heuristic function is inversely proportional to the probability of track intersection and the spatial distance. In the iterative process of the ant colony algorithm, the state transition probability is calculated by combining the pheromone concentration and the heuristic function, and the pheromone distribution is updated by pheromone attenuation and the optimal solution increment, and finally the optimized target priority is obtained. Assume that the priority of target A is updated to 0.75 and the priority of target B is updated to 0.25.

[0125] According to the optimized target priority, the threat targets are sorted and a threat target sequence and situation threat assessment report are generated. According to the priority sorting, target A has a higher priority than target B. The report contains the basic information of the target, threat assessment indicators, priority analysis, and situation evolution prediction.

[0126] Through integrated learning and ant colony algorithm optimization, the present invention can more accurately evaluate the threat level and priority of space targets. By considering the target orbit parameters and threat propagation constraints, it can more comprehensively perceive the space situation and predict the evolution of the situation. By sorting the threat targets, it can help decision makers quickly identify high-priority targets and formulate effective response strategies.

[0127] In an optional embodiment,

[0128] The probability distribution vectors output by the base classifiers are concatenated to form a meta-feature vector, a lightweight gradient boosting machine is used as a secondary learner to stack and integrate the meta-feature vectors, a focal loss function is introduced to correct the class imbalance problem, and the class weights and focus parameters are determined by cross-validation. The steps of classifying the threat level of space targets into high risk, medium risk and low risk include:

[0129] A base classifier set consisting of a random forest classifier, a gradient boosting decision tree classifier, and a support vector machine classifier is used to classify and predict the space target, and each classifier outputs a first probability distribution vector that the space target belongs to three threat levels: high risk, medium risk, and low risk; the first probability distribution vectors are spliced ​​and combined to construct a multidimensional element feature vector containing the prediction probability information of all base classifiers;

[0130] A lightweight gradient boosting machine is used as a secondary learner to perform stacked ensemble learning on the multidimensional element feature vector to obtain a second probability distribution vector. The lightweight gradient boosting machine configures model parameters by controlling the complexity of the tree, the learning rate, the feature random sampling ratio, the data random sampling ratio and the minimum number of leaf node samples. A focus loss function is introduced to correct the category imbalance problem of the threat level in the second probability distribution vector. The focus loss function includes a category weight factor and a focus parameter.

[0131] The category weight factor and the focus parameter are optimized by a cross-validation method, a weighted F1 score is used as an evaluation index for the optimization of the category weight factor, and a harmonic mean of a weighted precision and a recall rate is used as an evaluation index for the optimization of the focus parameter, to obtain an optimized focus loss function;

[0132] The lightweight gradient boosting machine is trained based on the optimized focal loss function to obtain a final third probability distribution vector, and multi-category judgment is performed on the third probability distribution vector by setting a probability threshold to generate a final threat level classification result of the space target.

[0133] Exemplarily, first, a set of base classifiers is constructed. Classifiers with excellent performance and differences are selected as base classifiers, such as random forest classifiers, gradient boosting decision tree classifiers, and support vector machine classifiers. Use historical space target data, including the orbital parameters, physical characteristics, functional characteristics, etc. of the target, to train the base classifiers. Suppose there are 1,000 space target samples, including 200 high-risk targets, 300 medium-risk targets, and 500 low-risk targets. Each base classifier predicts each space target and outputs a probability distribution vector that the target belongs to the three threat levels of high risk, medium risk, and low risk. For example, for a certain target, the probability distribution vector output by the random forest classifier is [0.2, 0.3, 0.5], the probability distribution vector output by the gradient boosting decision tree classifier is [0.1, 0.4, 0.5], and the probability distribution vector output by the support vector machine classifier is [0.3, 0.2, 0.5].

[0134] Next, construct a meta-feature vector. Concatenate the probability distribution vectors output by all base classifiers to construct a multi-dimensional meta-feature vector that contains the predicted probability information of all base classifiers. For example, after concatenating the above three probability distribution vectors, we get a 9-dimensional meta-feature vector [0.2, 0.3, 0.5, 0.1, 0.4, 0.5, 0.3, 0.2, 0.5]. This meta-feature vector integrates the prediction information of multiple base classifiers and can more comprehensively describe the threat level characteristics of space targets.

[0135] Then, stacked ensemble learning is performed. A lightweight gradient boosting machine is used as a secondary learner to perform stacked ensemble learning on the constructed meta-feature vector. The lightweight gradient boosting machine model is configured by controlling parameters such as the depth of the tree, the learning rate, the feature ratio used in each iteration, the data ratio used in each iteration, and the minimum number of samples contained in the leaf node to prevent overfitting and reduce computational complexity. The lightweight gradient boosting machine learns the mapping relationship between the meta-feature vector and the true threat level and outputs a second probability distribution vector. For example, for the above meta-feature vector, the second probability distribution vector output by the lightweight gradient boosting machine is [0.25, 0.35, 0.4].

[0136] Introducing the focus loss function. In order to solve the problem of imbalanced threat level categories, the focus loss function is introduced to correct the second probability distribution vector. The focus loss function contains a category weight factor and a focus parameter. The category weight factor is used to adjust the weights of different categories in the loss function, and the focus parameter is used to control the degree of attention the model pays to samples that are easy to classify.

[0137] Optimize the focus loss function parameters. Use the cross-validation method to optimize the category weight factor and focus parameters. Use the weighted F1 score as the evaluation indicator for category weight factor optimization, and the harmonic mean of weighted precision and weighted recall as the evaluation indicator for focus parameter optimization. Through cross-validation, find the category weight factor and focus parameter that make the evaluation indicators reach the optimal level, and obtain the optimized focus loss function.

[0138] Train the final model. Train the lightweight gradient boosting machine based on the optimized focal loss function to obtain the final third probability distribution vector. For example, for the above meta-feature vector, the third probability distribution vector output by the final model is [0.3, 0.3, 0.4].

[0139] Generate the final threat level classification result. By setting the probability threshold, the third probability distribution vector is judged in multiple categories to generate the final threat level classification result of the space target. For example, if the threshold is set to 0.35, the final threat level of the above target is low risk.

[0140] The present invention can effectively improve the accuracy of space target threat level classification by integrating the prediction results of multiple base classifiers and combining the focal loss function to deal with the category imbalance problem. The cross-validation method is used to optimize the model parameters, which can enhance the generalization ability and robustness of the model so that it can still maintain a high classification performance when facing new space target data. The lightweight gradient boosting machine is used as a secondary learner, which can effectively reduce the computational complexity and improve the training and prediction efficiency of the model.

[0141] Figure 2FIG. 1 is a schematic diagram of a space competition situation threat level assessment system based on credibility weighting according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0142] The first unit is used to construct a threat feature vector including maneuvering features, load features and target features based on historical orbit data and the operating status of on-orbit satellites, wherein the maneuvering features include orbital maneuvering frequency, speed change and orbit adjustment amplitude, the load features include mission load type, power change and signal characteristics, and the target features include target size, attitude change and orbit type; a hierarchical analysis method is used to determine the feature index weight coefficients of maneuvering features, load features and target features, the threat feature vector is input into an encoder-decoder deep learning model, a multi-head attention layer is set between the encoder and the decoder, and a generative adversarial network training method is used to obtain an initial target threat probability vector;

[0143] The second unit is used to collect multi-source situational awareness data based on radar, optical and radio measurements, input the collection time interval, data volatility and measurement accuracy into the data quality assessment network to calculate the timeliness coefficient, accuracy coefficient and integrity coefficient, and use the deep belief network to fuse the timeliness coefficient, accuracy coefficient and integrity coefficient to obtain the credibility weighted coefficient of the multi-source situational awareness data; based on the credibility weighted coefficient and the characteristic index weight coefficient, a modified weight matrix is ​​formed, and the modified weight matrix and the initial target threat probability vector are input into the variational Bayesian network to obtain a modified target threat probability vector and a confidence interval;

[0144] The third unit is used to construct a base classifier set including random forest, gradient boosting decision tree and support vector machine according to the corrected target threat probability vector and confidence interval, and adopts a stacking integration strategy to fuse the classification results of the base classifier set, so as to divide the threat level of space targets into three levels: high risk, medium risk and low risk. The initial value of the target priority is determined based on the threat level classification result, and the constraint conditions of threat propagation are constructed in combination with the target orbit parameters. The ant colony algorithm is used to solve the constraint conditions and update the target priority, and the threat target sequence and situation threat assessment report are output according to the target priority.

[0145] According to a third aspect of the embodiments of the present invention,

[0146] An electronic device is provided, comprising:

[0147] processor;

[0148] a memory for storing processor-executable instructions;

[0149] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0150] A fourth aspect of the embodiments of the present invention is:

[0151] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0152] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A space competition situation threat level assessment method based on credibility weighting, characterized in that: include: Based on historical orbital data and the operating status of on-orbit satellites, a threat feature vector including maneuvering features, payload features and target features is constructed, wherein the maneuvering features include orbital maneuvering frequency, speed change and orbit adjustment amplitude, the payload features include mission payload type, power change and signal characteristics, and the target features include target size, attitude change and orbit type; a hierarchical analysis method is used to determine the feature index weight coefficients of maneuvering features, payload features and target features, and the threat feature vector is input into the encoder-decoder deep learning model, a multi-head attention layer is set between the encoder and the decoder, and a generative adversarial network training method is used to obtain an initial target threat probability vector; Based on radar, optical and radio measurements, multi-source situational awareness data is collected, and the collection time interval, data volatility and measurement accuracy are input into the data quality assessment network to calculate the timeliness coefficient, accuracy coefficient and integrity coefficient. The timeliness coefficient, accuracy coefficient and integrity coefficient are integrated by a deep belief network to obtain the credibility weighted coefficient of the multi-source situational awareness data; based on the credibility weighted coefficient and the characteristic index weight coefficient, a modified weight matrix is ​​formed, and the modified weight matrix and the initial target threat probability vector are input into the variational Bayesian network to obtain a modified target threat probability vector and a confidence interval; According to the corrected target threat probability vector and confidence interval, a base classifier set including random forest, gradient boosting decision tree and support vector machine is constructed. The classification results of the base classifier set are fused using the stacking integration strategy. The threat level of space targets is divided into three levels: high risk, medium risk and low risk. The initial value of the target priority is determined based on the threat level classification result. The constraint conditions of threat propagation are constructed in combination with the target orbit parameters. The ant colony algorithm is used to solve the constraint conditions and update the target priority. The threat target sequence and situation threat assessment report are output according to the target priority.

2. The method according to claim 1, characterized in that The step of determining the weight coefficients of the characteristic indicators of the maneuvering characteristics, the load characteristics and the target characteristics by using the hierarchical analysis method, inputting the threat characteristic vector into the encoder-decoder deep learning model, setting a multi-head attention layer between the encoder and the decoder, and using the generative adversarial network training method to obtain the initial target threat probability vector includes: Constructing a judgment matrix of a three-layer hierarchical structure, the three-layer hierarchical structure includes a target layer, a criterion layer and a solution layer, wherein the target layer is a threat assessment result, the criterion layer includes maneuverability characteristics, load characteristics and target characteristics, and the solution layer is a quantitative index corresponding to each characteristic in the criterion layer, normalizing the judgment matrix to obtain a criterion layer weight vector, and calculating the weight of each quantitative index of the solution layer based on the criterion layer weight vector to obtain a characteristic index weight coefficient; The product of the threat feature vector and the feature index weight coefficient is input into an encoder constructed by a three-layer bidirectional recurrent neural network, and residual connections and layer normalization modules are set between each layer of the encoder, and the encoder outputs a hidden state sequence; a parallel attention mechanism is constructed on the hidden state sequence, and the hidden state sequence is linearly transformed to obtain a query matrix, a key matrix and a value matrix, and the attention weight is calculated based on the query matrix and the key matrix, and the attention weight is weighted and summed with the value matrix to obtain an attention head output, and multiple attention head outputs are spliced ​​to obtain an enhanced feature sequence, wherein the parallel attention mechanism includes a local attention head for capturing local temporal associations, a global attention head for establishing long-range dependencies, and a cross attention head for feature interaction; A generative adversarial network is constructed, wherein the generator of the generative adversarial network generates a candidate threat probability vector through a fully connected network, the candidate threat probability vector is input into the discriminator of the generative adversarial network for discrimination, the cross entropy loss is calculated based on the discrimination value output by the discriminator, and the probability value constraint loss and the temporal consistency loss of the candidate threat probability vector are calculated at the same time, the cross entropy loss, the probability value constraint loss and the temporal consistency loss are used as training targets, and the generator and the discriminator are alternately trained until convergence; the enhanced feature sequence is input into the trained generative adversarial network to obtain an initial target threat probability vector.

3. The method according to claim 2, characterized in that The steps of calculating a cross entropy loss based on the discriminant value output by the discriminator, and simultaneously calculating a probability value constraint loss and a temporal consistency loss of the candidate threat probability vector, taking the cross entropy loss, the probability value constraint loss and the temporal consistency loss as training targets, and alternately training the generator and the discriminator until convergence include: A cross entropy loss function is constructed based on the candidate threat probability vector and the historical real samples, the label value of the historical real sample is randomly perturbed to obtain a smooth label, the cross entropy loss is calculated based on the smooth label, a gradient norm constraint term of the discriminator output relative to the input is constructed as a gradient penalty term, and a weighted combination of the cross entropy loss function and the gradient penalty term is used as the discriminator loss function; Performing probability constraint processing on the candidate threat probability vector to obtain a normalized probability vector, and constructing a probability value constraint loss function based on the normalized probability vector; constructing a temporal consistency loss function based on the normalized probability vector, wherein the temporal consistency loss function includes a probability difference constraint term at adjacent moments and a probability curve smoothing constraint term, introduces multiple time scales to construct a multi-scale temporal consistency loss, and performs a weighted combination on the losses of different time scales; The discriminator loss function, the probability value constraint loss function and the temporal consistency loss function are used as the combined training loss function, the combined training loss function is input into the attention mechanism network, the temporal features are extracted based on the historical change trend of the combined training loss function, the correlation between the discriminator loss function, the probability value constraint loss function and the temporal consistency loss function is calculated through the self-attention layer to obtain the loss attention matrix, and an adaptive weight network is constructed based on the loss attention matrix, the adaptive weight network includes a fully connected layer and a gated recurrent unit, and is used to dynamically predict the optimal combination weight of the combined training loss function; the combined training loss function is weighted based on the predicted optimal combination weight to obtain the generator loss function; Based on the generator loss function and the discriminator loss function, the network parameters of the discriminator and the generator are updated alternately until a preset training termination condition is met.

4. The method according to claim 1, characterized in that: The steps of collecting multi-source situational awareness data based on radar, optical and radio measurements, inputting the collection time interval, data fluctuation rate and measurement accuracy into the data quality assessment network to calculate the timeliness coefficient, accuracy coefficient and integrity coefficient, and using the deep belief network to fuse the timeliness coefficient, accuracy coefficient and integrity coefficient to obtain the credibility weighted coefficient of the multi-source situational awareness data include: The acquisition time interval is the time difference between two adjacent data acquisitions, the data fluctuation rate is the standard deviation of multiple consecutive measurement data, and the measurement accuracy is the measurement error index inherent in the equipment; The collection time interval, data volatility and measurement accuracy are input into a data quality assessment network, which is a multi-layer perceptron structure and includes three hidden layers. The hidden layers use a ReLU activation function and residual connections are set between adjacent hidden layers. The output layer of the data quality assessment network includes a timeliness coefficient node, an accuracy coefficient node and an integrity coefficient node, and a sigmoid activation function is used to limit the output value within a preset interval. During the training process of the data quality assessment network, a batch stochastic gradient descent method is used for optimization, the mean square error is selected as the loss function, and a dynamic attenuation mechanism of the learning rate is introduced. A batch normalization layer is added to the middle layer of the data quality assessment network to convert the output of the middle layer into a standard normal distribution. A deep belief network is constructed, wherein the deep belief network includes multiple restricted Boltzmann machine layers, the timeliness coefficient, the accuracy coefficient, the integrity coefficient and their first-order differences are input into the input layer of the deep belief network, and a sigmoid activation function is set at the output layer of the deep belief network; the dynamic importance weights of the timeliness coefficient, the accuracy coefficient and the integrity coefficient are calculated based on the attention mechanism in the deep belief network; the timeliness coefficient, the accuracy coefficient and the integrity coefficient are weightedly combined with the corresponding dynamic importance weights, and the credibility weighted coefficient of the multi-source situation awareness data is obtained through the sigmoid activation function.

5. The method according to claim 1, characterized in that The steps of forming a modified weight matrix based on the credibility weight coefficient and the characteristic index weight coefficient, inputting the modified weight matrix and the initial target threat probability vector into a variational Bayesian network, and obtaining a modified target threat probability vector and a confidence interval include: Performing an outer product operation on the credibility weight coefficient vector and the characteristic index weight coefficient vector and performing L1 normalization on the operation result to obtain a modified weight matrix; constructing a variational Bayesian network including a latent variable layer and an observation variable layer, wherein the latent variable layer is used to model the approximate posterior distribution of the true threat probability, and the observation variable layer receives the modified weight matrix and the initial target threat probability vector as input, and adopts a diagonal Gaussian distribution as a variational distribution family in the variational Bayesian network; The modified weight matrix is ​​multiplied by the initial threat probability vector to obtain a weighted feature vector, and a variational inference process is performed based on the weighted feature vector by minimizing a variational lower bound including a likelihood function term and a KL divergence term; in the variational inference process, a stochastic gradient variational Bayesian method is used to iteratively update the mean parameter and the standard deviation parameter of the latent variable, and the mean parameter and the standard deviation parameter are combined with a standard normal distribution sampling value through a parameter reconstruction method for gradient estimation to obtain an approximate posterior distribution of the true threat probability; The value range of the approximate posterior distribution is divided into multiple probability subspaces, and importance sampling is performed in each probability subspace based on the probability value of the approximate posterior distribution corresponding to the subspace to obtain a sample set; the probability density function of the true threat probability is reconstructed for the sample set using a Gaussian kernel density estimation method with an adaptive bandwidth, wherein the adaptive bandwidth is determined by the sample standard deviation and the number of samples, and the quantiles at different confidence levels are calculated based on the reconstructed probability density function by a numerical integration method to obtain the upper and lower bounds of the confidence interval of the true threat probability; A dynamic time window is constructed. In each dynamic time window, the approximate posterior distribution is recalculated based on the mean parameter and standard deviation parameter updated by the latest iteration. The bandwidth parameter of the Gaussian kernel function and the upper and lower bounds of the confidence interval of the true threat probability are updated. The boundaries of the confidence interval are obtained by solving the integral equation using the binary search method. The corrected target threat probability vector is output according to the mean parameter updated by the latest iteration, and the corresponding confidence interval is output based on the upper and lower bounds of the confidence interval updated by the latest iteration.

6. The method according to claim 1, characterized in that According to the corrected target threat probability vector and confidence interval, a base classifier set including random forest, gradient boosting decision tree and support vector machine is constructed, and the classification results of the base classifier set are fused by stacking integration strategy, and the threat level of the space target is divided into three levels: high risk, medium risk and low risk. The initial value of the target priority is determined based on the threat level classification result, and the constraint condition of threat propagation is constructed in combination with the target orbit parameter. The ant colony algorithm is used to solve the constraint condition and update the target priority. The steps of outputting the threat target sequence and the situation threat assessment report according to the target priority include: The corrected target threat probability vector is combined with the confidence interval to form a multidimensional feature vector, and a base classifier set including a random forest classifier, a gradient boosted decision tree classifier, and a support vector machine classifier is constructed based on the multidimensional feature vector. The random forest classifier uses the Gini coefficient as a node splitting criterion and constructs a decision tree through bootstrap sampling. The gradient boosted decision tree classifier uses a logarithmic loss function and is optimized through the Newton-Raphson method. The support vector machine classifier uses a radial basis kernel function, and each base classifier outputs a probability distribution vector of the target belonging to each threat level; The probability distribution vectors output by the base classifiers are concatenated to form a meta-feature vector, a lightweight gradient boosting machine is used as a secondary learner to stack and integrate the meta-feature vectors, a focal loss function is introduced to correct the category imbalance problem, and the category weights and focus parameters are determined through cross-validation to divide the threat levels of space targets into high risk, medium risk and low risk; based on the threat level division result and the corrected target threat probability vector, a priority scoring function including a threat level indication function and a threat probability is constructed, and an initial value of the target priority is obtained through normalization processing; Based on the target orbit parameters, a simplified perturbation orbit propagation model is used to calculate the orbit intersection probability, and a threat propagation constraint condition set including a spatial distance function and a time window constraint is constructed. The spatial distance function is calculated based on the target position vector, and the time window constraint is determined by the threat judgment threshold. The target priority initial value is used as the initial pheromone distribution, and a heuristic function is constructed based on the threat propagation constraint condition set. In the iterative process of the ant colony algorithm, the state transition probability is calculated by combining the pheromone concentration and the heuristic function, and the pheromone distribution is updated by pheromone attenuation and the optimal solution increment to obtain the optimized target priority. The threat targets are sorted according to the optimized target priorities, and a threat target sequence and a situation threat assessment report containing basic target information, threat assessment indicators, priority analysis and situation evolution prediction are generated.

7. The method according to claim 6, characterized in that The probability distribution vectors output by the base classifiers are concatenated to form a meta-feature vector, a lightweight gradient boosting machine is used as a secondary learner to stack and integrate the meta-feature vectors, a focal loss function is introduced to correct the class imbalance problem, and the class weights and focus parameters are determined by cross-validation. The steps of classifying the threat level of space targets into high risk, medium risk and low risk include: A base classifier set consisting of a random forest classifier, a gradient boosting decision tree classifier, and a support vector machine classifier is used to classify and predict the space target, and each classifier outputs a first probability distribution vector that the space target belongs to three threat levels: high risk, medium risk, and low risk; the first probability distribution vectors are spliced ​​and combined to construct a multidimensional element feature vector containing the prediction probability information of all base classifiers; A lightweight gradient boosting machine is used as a secondary learner to perform stacked ensemble learning on the multidimensional element feature vector to obtain a second probability distribution vector. The lightweight gradient boosting machine configures model parameters by controlling the complexity of the tree, the learning rate, the feature random sampling ratio, the data random sampling ratio and the minimum number of leaf node samples. A focus loss function is introduced to correct the category imbalance problem of the threat level in the second probability distribution vector. The focus loss function includes a category weight factor and a focus parameter. The category weight factor and the focus parameter are optimized by a cross-validation method, a weighted F1 score is used as an evaluation index for the optimization of the category weight factor, and a harmonic mean of a weighted precision and a recall rate is used as an evaluation index for the optimization of the focus parameter, to obtain an optimized focus loss function; The lightweight gradient boosting machine is trained based on the optimized focal loss function to obtain a final third probability distribution vector, and multi-category judgment is performed on the third probability distribution vector by setting a probability threshold to generate a final threat level classification result of the space target.

8. A space competition situation threat level assessment system based on credibility weighting, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to construct a threat feature vector including maneuvering features, load features and target features based on historical orbit data and the operating status of on-orbit satellites, wherein the maneuvering features include orbital maneuvering frequency, speed change and orbit adjustment amplitude, the load features include mission load type, power change and signal characteristics, and the target features include target size, attitude change and orbit type; a hierarchical analysis method is used to determine the feature index weight coefficients of maneuvering features, load features and target features, the threat feature vector is input into an encoder-decoder deep learning model, a multi-head attention layer is set between the encoder and the decoder, and a generative adversarial network training method is used to obtain an initial target threat probability vector; The second unit is used to collect multi-source situational awareness data based on radar, optical and radio measurements, input the collection time interval, data volatility and measurement accuracy into the data quality assessment network to calculate the timeliness coefficient, accuracy coefficient and integrity coefficient, and use the deep belief network to fuse the timeliness coefficient, accuracy coefficient and integrity coefficient to obtain the credibility weighted coefficient of the multi-source situational awareness data; based on the credibility weighted coefficient and the characteristic index weight coefficient, a modified weight matrix is ​​formed, and the modified weight matrix and the initial target threat probability vector are input into the variational Bayesian network to obtain a modified target threat probability vector and a confidence interval; The third unit is used to construct a base classifier set including random forest, gradient boosting decision tree and support vector machine according to the corrected target threat probability vector and confidence interval, and adopts a stacking integration strategy to fuse the classification results of the base classifier set, so as to divide the threat level of space targets into three levels: high risk, medium risk and low risk. The initial value of the target priority is determined based on the threat level classification result, and the constraint conditions of threat propagation are constructed in combination with the target orbit parameters. The ant colony algorithm is used to solve the constraint conditions and update the target priority, and the threat target sequence and situation threat assessment report are output according to the target priority.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images