An intelligent method for determining relative advantages and disadvantages of electromagnetic spectrum situation
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- THE 54TH RESEARCH INSTITUTE OF CHINA ELECTRONICS TECHNOLOGY GROUP CORPORATION
- Filing Date
- 2026-03-30
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies struggle to efficiently evaluate the quality of electromagnetic spectrum situations in complex electromagnetic environments, particularly in situations involving nonlinear relationships. Traditional methods for intelligently assessing the quality of electromagnetic spectrum situations are often ineffective in accurately evaluating complex electromagnetic environments, especially given their lack of flexibility.
A method for intelligently determining the relative superiority or inferiority of electromagnetic spectrum situation based on deep neural networks is adopted. By constructing a judgment dataset based on expert experience and domain knowledge, and utilizing cross-comparison deep neural networks and numerical judgment logic modules, the mapping from electromagnetic spectrum situation indicators to quantitative judgment results is realized, and complex nonlinear relationships are automatically learned.
It enables accurate assessment of the electromagnetic spectrum situation in complex electromagnetic environments, improving the accuracy and flexibility of the assessment results. It can adapt to changes in situation indicators and meet the needs of electromagnetic spectrum management and game-theoretic decision-making.
Smart Images
Figure CN121959030B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the intersection of electromagnetic spectrum management and artificial intelligence, and more specifically, to an intelligent method for determining the relative superiority or inferiority of electromagnetic spectrum conditions. Background Technology
[0002] In complex electromagnetic environments, it is necessary to evaluate the relative merits of the electromagnetic spectrum in order to formulate electromagnetic spectrum management strategies and schemes, supporting game-theoretic decision-making and iterative optimization of schemes. How to compare multi-dimensional electromagnetic spectrum situation indicators, integrate expert experience and domain knowledge, and obtain a mapping from situation indicators to judgment results is an important research topic in the field of electromagnetic spectrum management.
[0003] In the patent "A Simulation Method and System for Intelligent Fusion and Dynamic Prediction of Electronic Warfare Intelligence" applied for by Beijing Fangzhou Technology Co., Ltd., with authorization announcement number CN120012609B, an electromagnetic spectrum situation fusion processing flow combining multiple learning methods is proposed for electromagnetic signals collected by multi-domain sensors. The process outputs threat level, type and intent. However, the detailed calculation or acquisition method of the weights in the evaluation system is not given, and the basis for setting the evaluation parameters is rather vague.
[0004] In Wang Shuai et al.'s "Research on Battlefield Electromagnetic Environment Complexity Assessment under Multi-level Indicators" published in *Equipment Environmental Engineering*, a combination of hierarchical analysis and gray-scale comprehensive evaluation was used to obtain the evaluation results. In Lei Yingjie et al.'s "Battlefield Situation Assessment and Analysis Method Based on Intuitive Fuzzy Decision-Making" published in *Acta Electronica Sinica*, intuitive fuzzy logic, the Delphi method, and hierarchical analysis were used to comprehensively assess the battlefield situation. Both methods have established complete fixed indicator systems and utility value calculation rules. However, the electromagnetic spectrum situation is complex, and the relationship between various indicators (such as the number of radiation sources, spectral density, and signal strength) and the final judgment of superiority or inferiority is often highly nonlinear. Traditional linear weighting or simple mathematical models are difficult to accurately fit this complex mapping relationship in electromagnetic spectrum situation assessment. Furthermore, this method requires pre-setting a fixed indicator system and evaluation rules. Once the compared situation indicators increase or change, the fixed indicator system and weights may no longer be applicable, requiring readjustment of the algorithm's internal structure and weights, resulting in poor flexibility. Summary of the Invention
[0005] In view of this, the present invention proposes an intelligent method for determining the relative superiority or inferiority of electromagnetic spectrum situations, which breaks through the mathematical mapping of complex determination relationships, realizes intelligent determination of relative superiority or inferiority, and has a clear structure that is easy to implement in engineering.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] A method for intelligently determining the relative superiority or inferiority of electromagnetic spectrum situations includes the following steps:
[0008] Step 1: Construct an electromagnetic spectrum situation assessment dataset based on expert experience and domain knowledge: Collect electromagnetic spectrum situation indicator data and assessment results; construct a sample set based on electromagnetic spectrum situation indicators, use the assessment results as a label set to form an electromagnetic spectrum situation assessment dataset, and divide the electromagnetic spectrum situation assessment dataset into a training dataset and a test dataset. The training dataset is used to train a deep neural network, while the test dataset is used to evaluate and save the deep neural network with the best assessment performance.
[0009] Step 2: Construct an electromagnetic spectrum situation assessment model: consisting of a cross-comparison deep neural network and a numerical judgment logic module; design the cross-comparison deep neural network to map the electromagnetic spectrum situation into quantitative judgment values, and then use the numerical judgment logic module to realize the mapping from the quantitative judgment values to the superiority / inferiority logic.
[0010] Step 3: Train a cross-contrast deep neural network using an electromagnetic spectrum situation assessment dataset: Use the training dataset to fit the parameters of the cross-contrast deep neural network, and obtain the judgment accuracy of the neural network on the test dataset. Save the parameters of the cross-contrast deep neural network with the highest judgment accuracy as the optimal cross-contrast deep neural network parameters.
[0011] Step 4: Input the electromagnetic spectrum situation indicators into the electromagnetic spectrum situation quality judgment model and judge the quality of the input electromagnetic spectrum situation: Load the optimal cross-comparison deep neural network parameters, input the electromagnetic spectrum situation indicators into the cross-comparison deep neural network, obtain the quantitative judgment value, and obtain the quality judgment through the numerical judgment logic module.
[0012] Furthermore, in step 1, an electromagnetic spectrum situation assessment dataset based on expert experience and domain knowledge is constructed, specifically as follows:
[0013] Step 1-1: Construct an electromagnetic spectrum situation assessment sample dataset Electromagnetic spectrum situation assessment sample dataset There are N pairs of electromagnetic spectrum situation indicators. Each pair contains two sets of electromagnetic spectrum situation indicators, corresponding to two electromagnetic spectrum situations to be judged as superior or inferior, for a total of 2N electromagnetic spectrum situations. Each set of electromagnetic spectrum situation indicators contains M indicators, so each pair contains 2M indicators. The first M indicators correspond to the first electromagnetic spectrum situation, and the last M indicators correspond to the second electromagnetic spectrum situation. Therefore, the sample representation of the nth pair of electromagnetic spectrum situation indicators is as follows:
[0014]
[0015] in This represents the Mth index in the nth electromagnetic spectrum situational index pair; n = 1, 2, 3, ..., N;
[0016] Electromagnetic spectrum situation assessment sample dataset Represented as:
[0017]
[0018] Step 1-2: Construct an electromagnetic spectrum situation assessment label dataset Experts assess each pair of electromagnetic spectrum situation indicators based on their experience and domain knowledge. For the nth pair of electromagnetic spectrum situation indicators: if the first electromagnetic spectrum situation is better than or equal to the second electromagnetic spectrum situation, then the tag data... =1; if the first electromagnetic spectrum situation is worse than the second electromagnetic spectrum situation, then the tag data The value is 0; after judging each of the N electromagnetic spectrum situation indicators, the judgment results are summarized to obtain the electromagnetic spectrum situation judgment label dataset. for:
[0019]
[0020] Steps 1-3: Electromagnetic Spectrum Situation Assessment Sample Dataset Electromagnetic spectrum situation determination label dataset Together they constitute the electromagnetic spectrum situation assessment dataset: according to proportion The data is split into segments to form a training sample dataset. Test sample dataset Training label dataset and test label dataset The dimensions are respectively , , and ; .
[0021] Furthermore, in step 2, an electromagnetic spectrum situation assessment model is constructed, specifically as follows:
[0022] Step 2-1: Construct a cross-comparison deep neural network: The input dimension is 2M, corresponding to the electromagnetic spectrum situation index pairs of samples, and the output dimension is 1, representing the quantification judgment value;
[0023] The cross-contrast deep neural network consists of an input layer, a feature extraction layer 1, a feature extraction layer 2, and an output layer; wherein, the feature extraction layer 1 includes feature extraction layer 1-1 and feature extraction layer 1-2, and the feature extraction layer 2 includes feature extraction layer 2-1 and feature extraction layer 2-2;
[0024] The input layer is constructed from a fully connected layer named input_layer, with parameters in_features and out_features. The input data is processed into ipt.
[0025] Feature extraction layer 1-1 and feature extraction layer 1-2 are both constructed from fully connected layers, named extract11_layer and extract12_layer respectively, with parameters in_features and out_features respectively. After inputting ipt into extract11_layer and extract12_layer respectively, ex11 and ex12 are obtained.
[0026] Feature extraction layer 2-1 and feature extraction layer 2-2 are both constructed from fully connected layers, named extract21_layer and extract22_layer respectively, with parameters in_features and out_features respectively; the result of subtracting ex12 from ex11 is input into extract21_layer to obtain ex21, and the result of adding ex22 to ex21 is input into extract22_layer to obtain ex22;
[0027] The output layer is constructed from a fully connected layer named output_layer, with parameters in_features and out_features. ex21 and ex22 are concatenated to form ex3, which conforms to the in_features dimension of the output layer. ex3 is input into output_layer to obtain opt, which is used as the quantization decision value.
[0028] Where in_features is the number of input features and out_features is the number of output features. The in_features value of the input layer is 2M, the out_features value of the input layer is equal to the in_features of feature extraction layers 1-1 and 1-2, the out_features value of feature extraction layers 1-1 and 1-2 is equal to the in_features of feature extraction layers 2-1 and 2-2, the in_features value of the output layer is twice the out_features of feature extraction layers 2-1 and 2-2, and the out_features value of the output layer is 1. The parameter values of feature extraction layers 1-1 and 1-2 are the same, and the parameter values of feature extraction layers 2-1 and 2-2 are the same.
[0029] Step 2-2: Construct a numerical judgment logic module: If the quantization judgment value obtained by the cross-comparison deep neural network is greater than or equal to 0, then the quantization judgment result is recorded as 1, indicating that the first electromagnetic spectrum situation is better than or equal to the second electromagnetic spectrum situation; if the quantization judgment value obtained by the cross-comparison deep neural network is less than 0, then the quantization judgment result is recorded as 0, indicating that the first electromagnetic spectrum situation is worse than the second electromagnetic spectrum situation.
[0030] Steps 2-3: An electromagnetic spectrum situation assessment model is constructed by a cross-comparison deep neural network and a numerical judgment logic module. The quantitative judgment values obtained from the cross-comparison deep neural network are input into the numerical judgment logic module to obtain the assessment of superiority or inferiority.
[0031] Furthermore, in step 3, a cross-comparison deep neural network is trained using the electromagnetic spectrum situation assessment dataset. The specific method is as follows:
[0032] Step 3-1: Define the optimal consistency variable, with an initial value of 0;
[0033] Step 3-2: Train a cross-comparison deep neural network using multiple batches of data from the training sample dataset. and training label dataset Data from group B is randomly selected to form a batch training sample dataset. and batch training label dataset The batch training sample dataset has dimensions of B×2M and B×1 respectively. The input is fed into a cross-comparison deep neural network to obtain a quantized decision value set. The dimension is B×1; and the corresponding quantitative judgment result set is obtained through the numerical judgment logic module. Batch training label dataset is calculated based on binary cross-entropy with Logits. And quantitative judgment result set The error is calculated, and backpropagation is performed based on the calculated error to optimize the neural network parameters; then, based on the training sample dataset... and training label dataset The remaining data in the dataset continues to be randomly sampled from group B for training, until the training sample dataset is reached. and training label dataset Once all data has been trained, proceed to step 3-3;
[0034] Step 3-3: Set the cross-comparison deep neural network to evaluation mode and use the test sample dataset. Inputting the data into a cross-comparison deep neural network yields a set of quantized decision values. The quantitative judgment result set is obtained through the numerical judgment logic module. Quantify the judgment result set With test label dataset Check each digit individually, count the number of equal digits, and divide by the total number. Achieve consistency;
[0035] Steps 3-4: If the consistency score is greater than the optimal consistency score, replace the optimal consistency score with the consistency score and save the current cross-comparison deep neural network parameters as the optimal network parameters; otherwise, do nothing.
[0036] Steps 3-5: Steps 3-2 to 3-4 are repeated E times. After the repetition ends, proceed to step 4.
[0037] Furthermore, in step 4, the electromagnetic spectrum situation indicators are input into the electromagnetic spectrum situation quality assessment model to determine the quality of the input electromagnetic spectrum situation. The specific method is as follows:
[0038] Step 4-1: Load the optimal parameters for the cross-comparison deep neural network;
[0039] Step 4-2: Construct the index set of the electromagnetic spectrum situation to be judged into index pairs, input them into the electromagnetic spectrum situation superiority and inferiority judgment model, and obtain the electromagnetic spectrum situation superiority and inferiority judgment result.
[0040] Due to the adoption of the above technical solution, the beneficial effects of this invention compared with the prior art are as follows:
[0041] First, based on deep neural networks and combining domain knowledge and expert experience, a dataset for judging merits and demerits is constructed, and a mapping from electromagnetic spectrum situation indicators to quantitative judgment results is obtained through comparison. Compared with methods that manually construct formula mappings and evaluation systems, this invention can automatically learn complex nonlinear relationships from data. Its ability to fit complex nonlinear functions can transform complex logic into quantitative judgment values.
[0042] Then, a cross-comparison deep neural network capable of extracting the superiority or inferiority features of the electromagnetic spectrum situation was designed. By designing a difference feature extraction processing structure, the two input electromagnetic spectrum situation indicators are added and subtracted to deeply extract the difference features of the electromagnetic spectrum situation. This structure can effectively support the formation of intelligent superiority or inferiority judgment capabilities. Attached Figure Description
[0043] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.
[0044] Figure 2 This is a loss curve for cross-comparison deep neural network training.
[0045] Figure 3 This is a consistency curve during the training process of a cross-compare deep neural network. Detailed Implementation
[0046] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0047] A method for intelligently determining the relative superiority or inferiority of electromagnetic spectrum situations, such as Figure 1 As shown, it includes the following steps:
[0048] Step 1 involves constructing an electromagnetic spectrum situation assessment dataset based on expert experience and domain knowledge.
[0049] Step 1-1: Construct an electromagnetic spectrum situation assessment sample dataset Electromagnetic spectrum situation assessment sample dataset There are N pairs of electromagnetic spectrum situation indicators, N=1000. Each pair contains two sets of electromagnetic spectrum situation indicators, corresponding to two electromagnetic spectrum situations to be judged as superior or inferior, for a total of 2N electromagnetic spectrum situation indicators. Each set of electromagnetic spectrum situation indicators contains M indicators, M=5, so each pair contains 2M indicators. The electromagnetic spectrum situation indicators include signal strength, spectrum occupancy, spectrum utilization, spatial distribution, and signal-to-noise ratio. The first M indicators correspond to the first electromagnetic spectrum situation, and the last M indicators correspond to the second electromagnetic spectrum situation. Therefore, the sample of the nth pair of electromagnetic spectrum situation indicators can be represented as:
[0050]
[0051] in This represents the Mth index in the nth electromagnetic spectrum situational index pair; n = 1, 2, 3, ..., N;
[0052] Electromagnetic spectrum situation assessment sample dataset It can be represented as:
[0053]
[0054] Step 1-2: Construct an electromagnetic spectrum situation assessment label dataset Experts assess each pair of electromagnetic spectrum situation indicators based on their experience and domain knowledge, and then make a judgment on the nth pair of electromagnetic spectrum situation indicators. If the first electromagnetic spectrum situation is better than or equal to the second electromagnetic spectrum situation, then the tag data... The value is 1. If the first electromagnetic spectrum situation is worse than the second electromagnetic spectrum situation, then the tag data... The value is 0. After judging each of the N electromagnetic spectrum situation indicators, the judgment results are summarized to obtain the electromagnetic spectrum situation judgment label dataset. for:
[0055]
[0056] Steps 1-3: Electromagnetic Spectrum Situation Assessment Sample Dataset Electromagnetic spectrum situation determination label dataset Together, they constitute the electromagnetic spectrum situation assessment dataset. According to proportion... The data is split and set to 0.9 to form the training sample dataset. Test sample dataset Training label dataset and test label dataset The dimensions are respectively , , and .
[0057] In step 2, an electromagnetic spectrum situation assessment model is constructed.
[0058] Step 2-1: Construct a cross-comparison deep neural network. The input dimension is 2M, corresponding to electromagnetic spectrum situation indicators for samples, and the output dimension is 1, representing the quantified judgment value.
[0059] The cross-contrast deep neural network consists of an input layer, a feature extraction layer 1, a feature extraction layer 2, and an output layer. Feature extraction layer 1 comprises feature extraction layers 1-1 and 1-2, and feature extraction layer 2 comprises feature extraction layers 2-1 and 2-2.
[0060] The input layer is constructed from a fully connected layer named input_layer, with parameters in_features and out_features, which are 10 and 16 respectively. The input data is processed into ipt.
[0061] Feature extraction layers 1-1 and 1-2 are both constructed from fully connected layers, named extract11_layer and extract12_layer respectively, with parameters in_features and out_features of 16 and 32 respectively. After inputting ipt into extract11_layer and extract12_layer respectively, ex11 and ex12 are obtained.
[0062] Feature extraction layers 2-1 and 2-2 are both constructed from fully connected layers, named extract21_layer and extract22_layer, respectively, with parameters in_features and out_features of 32 and 16, respectively. The result of subtracting ex12 from ex11 is input into extract21_layer to obtain ex21, and the result of adding ex22 to ex21 is input into extract22_layer to obtain ex22.
[0063] The output layer is constructed from a fully connected layer named output_layer, with parameters in_features and out_features, which are 32 and 1 respectively. ex21 and ex22 are concatenated along the first dimension to form ex3, making it conform to the in_features dimension of the output layer. The output layer is input to output_layer to obtain opt, which is used as the quantization decision value.
[0064] Step 2-2: Construct the numerical judgment logic module. If the quantization judgment value obtained by the cross-comparison deep neural network is greater than or equal to 0, the quantization judgment result is recorded as 1, indicating that the first electromagnetic spectrum situation is better than or equal to the second electromagnetic spectrum situation; if the quantization evaluation value obtained by the cross-comparison deep neural network is less than 0, the quantization judgment result is recorded as 0, indicating that the first electromagnetic spectrum situation is worse than the second electromagnetic spectrum situation.
[0065] Steps 2-3: An electromagnetic spectrum situation assessment model is constructed using a cross-comparison deep neural network and a numerical judgment logic module. The quantified judgment values obtained from the cross-comparison deep neural network are input into the numerical judgment logic module to obtain the assessment of superiority or inferiority.
[0066] Step 3 involves training a cross-contrast deep neural network using the dataset. The specific method is as follows:
[0067] Step 3-1: Define the optimal consistency variable, with an initial value of 0.
[0068] Step 3-2: Train a cross-comparison deep neural network using multiple batches of data. (From the training sample dataset...) and training label dataset B data points are randomly selected from the dataset, where B=64, to form a batch training sample dataset. and batch training label dataset The dimensions are B×2M and B×1, respectively. The batch training sample dataset... The quantization determination value is obtained from the input network. The dimension is B×1, and the corresponding quantitative judgment result set is obtained through the numerical judgment logic module. Batch training label dataset is calculated based on binary cross-entropy with Logits. And quantitative judgment result set The error is calculated, and backpropagation is performed based on the calculated error to optimize the neural network parameters; then, based on the training sample dataset... and training label dataset The remaining data in the dataset continues to be randomly sampled from group B for training, until the training sample dataset is reached. and training label dataset Once all data has been trained, proceed to step 3-3;
[0069] Step 3-3: Set the cross-comparison deep neural network to evaluation mode and use the test sample dataset. Inputting the data into a cross-comparison deep neural network yields a set of quantized decision values. The quantitative judgment result set is obtained through the numerical judgment logic module. Quantify the judgment result set With test label dataset Check each digit individually, count the number of equal digits, and divide by the total number. Achieve consistency.
[0070] Steps 3-4: If the consistency score is greater than the optimal consistency score, the optimal consistency score is replaced with the consistency score, and the parameters of the cross-comparison deep neural network are saved as the optimal network parameters.
[0071] Steps 3-5: Steps 3-2 to 3-4 are repeated E times, where E=1000. The loss curve during the cross-comparison deep neural network training process is shown below. Figure 2 As shown, the consistency curve during the training process of the cross-comparison deep neural network is as follows: Figure 3 As shown. After the repetition ends, proceed to step 4.
[0072] In step 4, the electromagnetic spectrum situation indicators are input into the electromagnetic spectrum situation quality judgment model to determine the quality of the input electromagnetic spectrum situation.
[0073] Step 4-1: Load the optimal parameters for the cross-comparison deep neural network;
[0074] Step 4-2: Construct the index set of the electromagnetic spectrum situation to be determined into index pairs, a total of 100 pairs, as shown below: [[0.9736173 0.9550846 0.8913594 0.8429173 0.26555642 0.07916606 0.8768143 0.70865756 0.45641232 0.7243086 ] [0.9197325 0.6107675 0.4789023 0.39617398 0.23088303 0.6033127 0.9010833 0.37507868 0.5299655 0.694047]
[0077] ... [0.6673682 0.04109485 0.79285717 0.76314294 0.92629457 0.1052649 0.31684172 0.0592408 0.5966528 0.12419324] [0.73509127 0.35511827 0.88209873 0.01241088 0.2519372 0.21657415 0.33122692 0.8336373 0.51358837 0.8589025]]
[0080] The indicators to be judged are input into the electromagnetic spectrum situation assessment model to obtain the electromagnetic spectrum situation assessment results. The intelligent assessment results are shown below: [1. 0. 1. 0. 0. 0. 0. 1. 0. 0. 0. 1. 0. 1. 0. 0. 1. 1. 1. 0. 1. 0. 0.1. 0. 1. 1. 0. 0. 1. 1. 1. 1. 0. 0. 1. 1. 0. 1. 0. 1. 0. 1. 1. 0. 0. 1.0. 1. 0. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 0. 0. 0. 0. 1. 0. 1. 1. 1. 1. 0. 0. 1. 1. 1. 1. 0. 0. 1. 1. 1. 0. 1. 0. 1. 1. 0. 1. 1. 0. 1. 0. 1. 1. 1. 1. 0. 1. 1. 0. 1. 0. 0. 1. 0. 0. 0. 1.]
[0082] The actual judgment results are as follows: [1. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 1. 0. 1. 1. 0. 1. 1. 0. 1. 0. 1.1. 0. 1. 1. 0. 0. 1. 0. 0. 1. 0. 0. 1. 1. 0. 1. 0. 1. 0. 1. 0. 1. 0. 0. 0. 1.0. 1. 0. 1. 0. 0. 0. 1.0. 1. 0. 1. 1. 1. 1. 1. 1. 1. 1. 1. 0. 0. 0. 0. 1. 0. 1. 1. 1. 1. 0. 0. 1. 1. 0. 1. 1. 0. 1. 1. 0. 1. 0. 1. 1. 1. 1. 0. 1. 1. 0. 1. 0. 0. 1. 0. 0. 0. 1.]
[0084] The consistency rate between the predicted results and the actual judgment results is 94%, which is highly consistent with the judgment results formed by experts. It can be seen that the method of this patent can effectively realize the intelligent judgment of the relative superiority or inferiority of the electromagnetic spectrum situation.
Claims
1. A method for intelligently determining the relative superiority or inferiority of electromagnetic spectrum situations, characterized in that, Includes the following steps: Step 1: Construct an electromagnetic spectrum situation assessment dataset based on expert experience and domain knowledge: collect electromagnetic spectrum situation indicator data and assessment results; A sample set is constructed based on electromagnetic spectrum situation indicators, and the judgment results are used as a label set to form an electromagnetic spectrum situation judgment dataset. The electromagnetic spectrum situation judgment dataset is divided into a training dataset and a test dataset. The training dataset is used to train a deep neural network, while the test dataset is used to evaluate and save the deep neural network with the best judgment effect. The electromagnetic spectrum situation indicators include signal strength, spectrum occupancy rate, spectrum utilization rate, spatial distribution, and signal-to-noise ratio. Step 2: Construct an electromagnetic spectrum situation assessment model: This model consists of a cross-comparison deep neural network and a numerical judgment logic module. The cross-comparison deep neural network maps the electromagnetic spectrum situation to quantified judgment values, and the numerical judgment logic module then maps these quantified values to the assessment logic. Specifically: Step 2-1: Construct a cross-comparison deep neural network: The input dimension is 2M, corresponding to the electromagnetic spectrum situation index pairs of samples, and the output dimension is 1, representing the quantification judgment value; The cross-contrast deep neural network consists of an input layer, feature extraction layer 1, feature extraction layer 2, and an output layer. Step 2-2: Construct a numerical judgment logic module: If the quantization judgment value obtained by the cross-comparison deep neural network is greater than or equal to 0, then the quantization judgment result is recorded as 1, indicating that the first electromagnetic spectrum situation is better than or equal to the second electromagnetic spectrum situation; if the quantization judgment value obtained by the cross-comparison deep neural network is less than 0, then the quantization judgment result is recorded as 0, indicating that the first electromagnetic spectrum situation is worse than the second electromagnetic spectrum situation. Steps 2-3: The electromagnetic spectrum situation superiority and inferiority judgment model is constructed by the cross-comparison deep neural network and the numerical judgment logic module. The quantitative judgment value obtained by the cross-comparison deep neural network is input into the numerical judgment logic module to obtain the superiority and inferiority judgment. Step 3: Train a cross-contrast deep neural network using an electromagnetic spectrum situation assessment dataset: Use the training dataset to fit the parameters of the cross-contrast deep neural network, and obtain the judgment accuracy of the neural network on the test dataset. Save the parameters of the cross-contrast deep neural network with the highest judgment accuracy as the optimal cross-contrast deep neural network parameters. Step 4: Input the electromagnetic spectrum situation indicators into the electromagnetic spectrum situation quality judgment model and judge the quality of the input electromagnetic spectrum situation: Load the optimal cross-comparison deep neural network parameters, input the electromagnetic spectrum situation indicators into the cross-comparison deep neural network to obtain the quantitative judgment value, and obtain the quality judgment through the numerical judgment logic module.
2. The intelligent method for determining the relative superiority or inferiority of electromagnetic spectrum situations according to claim 1, characterized in that, Step 1 involves constructing an electromagnetic spectrum situation assessment dataset based on expert experience and domain knowledge. The specific method is as follows: Step 1-1: Construct an electromagnetic spectrum situation assessment sample dataset Electromagnetic spectrum situation assessment sample dataset There are N pairs of electromagnetic spectrum situation indicators. Each pair contains two sets of electromagnetic spectrum situation indicators, corresponding to two electromagnetic spectrum situations to be judged as superior or inferior, for a total of 2N electromagnetic spectrum situations. Each set of electromagnetic spectrum situation indicators contains M indicators, so each pair contains 2M indicators. The first M indicators correspond to the first electromagnetic spectrum situation, and the last M indicators correspond to the second electromagnetic spectrum situation. Therefore, the sample representation of the nth pair of electromagnetic spectrum situation indicators is as follows: ； in This represents the Mth index in the nth electromagnetic spectrum situational index pair; n = 1, 2, 3, ..., N; Electromagnetic spectrum situation assessment sample dataset Represented as: ； Step 1-2: Construct an electromagnetic spectrum situation assessment label dataset Experts assess each pair of electromagnetic spectrum situation indicators based on their experience and domain knowledge. For the nth pair of electromagnetic spectrum situation indicators: if the first electromagnetic spectrum situation is better than or equal to the second electromagnetic spectrum situation, then the tag data... =1; if the first electromagnetic spectrum situation is worse than the second electromagnetic spectrum situation, then the tag data The value is 0; after judging each of the N electromagnetic spectrum situation indicators, the judgment results are summarized to obtain the electromagnetic spectrum situation judgment label dataset. for: ； Steps 1-3: Electromagnetic Spectrum Situation Assessment Sample Dataset Electromagnetic spectrum situation determination label dataset Together they constitute the electromagnetic spectrum situation assessment dataset: according to proportion The data is split into segments to form a training sample dataset. Test sample dataset Training label dataset and test label dataset The dimensions are respectively , , and ; .
3. The intelligent method for determining the relative superiority or inferiority of electromagnetic spectrum situations according to claim 1, characterized in that, The feature extraction layer 1 mentioned in step 2-1 includes feature extraction layer 1-1 and feature extraction layer 1-2, and the feature extraction layer 2 includes feature extraction layer 2-1 and feature extraction layer 2-2; The input layer is constructed from a fully connected layer named input_layer, with parameters in_features and out_features. The input data is processed into ipt. Feature extraction layer 1-1 and feature extraction layer 1-2 are both constructed from fully connected layers, named extract11_layer and extract12_layer respectively, with parameters in_features and out_features respectively. After inputting ipt into extract11_layer and extract12_layer respectively, ex11 and ex12 are obtained. Feature extraction layer 2-1 and feature extraction layer 2-2 are both constructed from fully connected layers, named extract21_layer and extract22_layer respectively, with parameters in_features and out_features respectively; the result of subtracting ex12 from ex11 is input into extract21_layer to obtain ex21, and the result of adding ex22 to ex21 is input into extract22_layer to obtain ex22; The output layer is constructed from a fully connected layer named output_layer, with parameters in_features and out_features. ex21 and ex22 are concatenated to form ex3, which conforms to the in_features dimension of the output layer. ex3 is input into output_layer to obtain opt, which is used as the quantization decision value. Where in_features is the number of input features and out_features is the number of output features. The in_features value of the input layer is 2M, the out_features value of the input layer is equal to the in_features of feature extraction layers 1-1 and 1-2, the out_features value of feature extraction layers 1-1 and 1-2 is equal to the in_features of feature extraction layers 2-1 and 2-2, the in_features value of the output layer is twice the out_features of feature extraction layers 2-1 and 2-2, and the out_features value of the output layer is 1. The parameter values of feature extraction layers 1-1 and 1-2 are the same, and the parameter values of feature extraction layers 2-1 and 2-2 are the same.
4. The intelligent method for determining the relative superiority or inferiority of electromagnetic spectrum situations according to claim 1, characterized in that, In step 3, a cross-comparison deep neural network is trained using the electromagnetic spectrum situation assessment dataset. The specific method is as follows: Step 3-1: Define the optimal consistency variable, with an initial value of 0; Step 3-2: Train a cross-comparison deep neural network using multiple batches of data from the training sample dataset. and training label dataset Data from group B is randomly selected to form a batch training sample dataset. and batch training label dataset The batch training sample dataset has dimensions of B×2M and B×1 respectively. The input is fed into a cross-comparison deep neural network to obtain a quantized decision value set. The dimension is B×1; and the corresponding quantitative judgment result set is obtained through the numerical judgment logic module. Batch training label dataset is calculated based on binary cross-entropy with Logits. And quantitative judgment result set The error is calculated, and backpropagation is performed based on the calculated error to optimize the neural network parameters; then, based on the training sample dataset... and training label dataset The remaining data in the dataset continues to be randomly sampled from group B for training, until the training sample dataset is reached. and training label dataset Once all data has been trained, proceed to step 3-3. Step 3-3: Set the cross-comparison deep neural network to evaluation mode and use the test sample dataset. Inputting the data into a cross-comparison deep neural network yields a set of quantized decision values. The quantitative judgment result set is obtained through the numerical judgment logic module. Quantify the judgment result set With test label dataset Check each digit individually, count the number of equal digits, and divide by the total number. Achieve consistency; Steps 3-4: If the consistency score is greater than the optimal consistency score, replace the optimal consistency score with the consistency score and save the current cross-comparison deep neural network parameters as the optimal network parameters; otherwise, do nothing. Steps 3-5: Steps 3-2 to 3-4 are repeated E times. After the repetition ends, proceed to step 4.
5. The intelligent method for determining the relative superiority or inferiority of electromagnetic spectrum situations according to claim 1, characterized in that, In step 4, the electromagnetic spectrum situation indicators are input into the electromagnetic spectrum situation assessment model to determine the quality of the input electromagnetic spectrum situation. The specific method is as follows: Step 4-1: Load the optimal parameters for the cross-comparison deep neural network; Step 4-2: Construct the index set of the electromagnetic spectrum situation to be judged into index pairs, input them into the electromagnetic spectrum situation superiority and inferiority judgment model, and obtain the electromagnetic spectrum situation superiority and inferiority judgment result.