A method, device, equipment and medium for evaluating similarity of radiation source signals

Abstract

The present application discloses a radiation source signal similarity evaluation method, device, equipment and medium, which relate to the field of signal processing technology. First, the acquired radiation source signal is divided into similar sample pairs and dissimilar sample pairs according to the radiation source, which together constitute a training data set. The abnormal samples in the training data set are cleaned and labeled to obtain a processed training data set, which is input into a constructed dual twin network to calculate the dual loss function, and the neural network reverse gradient is propagated and the loss function is minimized. After multiple iterative training, a converged similarity comparison model is obtained, and finally two signal features to be compared are input into the similarity comparison model, and the normalized semantic space distance is obtained through calculation, so as to calculate and measure the similarity between the radiation source signal features. The dual twin network can fully extract the differences in signal features in high-dimensional space, which solves the difficulty of threshold setting in traditional similarity evaluation.

Description

Technical Field

[0001] The present application relates to the field of signal processing technology, and in particular to a method, device, equipment and medium for evaluating similarity of radiation source signals. Background Art

[0002] As electromagnetic radiation scenes become increasingly complex, the similarity measurement of radiation source signal characteristics is becoming increasingly important in areas such as signal association and recognition.

[0003] Traditional similarity evaluation methods are mainly based on distance measurement, which requires manual design and optimization of features, and the extracted features need to be manually aligned, and the calculation results need to be divided by setting appropriate thresholds. Feature processing and threshold setting require a lot of experience and skills. When faced with new tasks and data, it is difficult to make effective comparisons and judgments. There are disadvantages such as time-consuming, labor-intensive, and highly subjective. Although artificial intelligence technology has achieved remarkable results in the fields of image recognition and natural language processing, traditional models such as neural networks are mainly used for classification tasks, and their effects are not ideal in terms of similarity comparison. Summary of the invention

[0004] The purpose of this application is to provide a method, device, equipment and medium for evaluating the similarity of radiation source signals in order to overcome the existing technical defects. The normalized semantic space distance is calculated to calculate and measure the similarity between the radiation source signal features. The dual twin network can fully extract the differences in signal features in high-dimensional space, solving the difficulty of threshold setting in traditional similarity evaluation.

[0005] The purpose of this application is achieved through the following technical solutions:

[0006] In a first aspect, the present application proposes a method for evaluating the similarity of radiation source signals, the method comprising:

[0007] The acquired radiation source signals are divided into similar sample pairs and dissimilar sample pairs according to the radiation sources, and together they form a training data set;

[0008] The abnormal samples in the training data set are cleaned and labeled to obtain the processed training data set;

[0009] Input the processed training data set into the constructed dual Siamese network to calculate the dual loss function;

[0010] The dual loss function is propagated through the neural network reverse gradient and the loss function is minimized. After multiple iterative training, a converged similarity comparison model is obtained.

[0011] The two signal features to be compared in the training data set are input into the similarity comparison model, and the normalized semantic space distance is obtained through calculation.

[0012] In a possible implementation, the dual twin network includes two twin input layers, a feature extraction network, and a distance calculation layer. The step of inputting the processed training data set into the constructed dual twin network to calculate the dual loss function includes:

[0013] The two signal features to be compared in the processed training data set are input into two twin input layers respectively, and then input into the feature extraction network for mapping to obtain two feature vectors Embedding;

[0014] Input the two feature vectors Embedding into the distance calculation layer to obtain the semantic distance in the metric space;

[0015] The semantic distance in the metric space is normalized by sigmoid to obtain the contrast loss function and binary cross entropy loss;

[0016] The dual loss function is calculated based on the contrastive loss function and the binary cross entropy loss.

[0017] In a possible implementation, the feature extraction network includes two sub-networks, each sub-network includes three residual connection blocks and shares the same parameters.

[0018] In one possible implementation, the contrast loss function for: ,in is the number of samples, Is the label of whether two samples match, is the Euclidean distance between two sample features, To set the threshold.

[0019] In one possible implementation, the binary cross entropy loss for: ,in is the sample size, is the true label, The label value predicted by the neural network.

[0020] In a second aspect, the present application proposes a radiation source signal similarity evaluation device, the device comprising:

[0021] A sample pair division module is used to divide the acquired radiation source signals into similar sample pairs and dissimilar sample pairs according to the radiation sources, which together form a training data set;

[0022] A cleaning and labeling module is used to clean and label abnormal samples in the training data set to obtain the processed training data set;

[0023] A first computing module is used to input the processed training data set into the constructed dual Siamese network to calculate the dual loss function;

[0024] The training module is used to propagate the dual loss function through the reverse gradient of the neural network and minimize the loss function. After multiple iterations of training, a converged similarity comparison model is obtained;

[0025] The second calculation module is used to input two signal features to be compared in the training data set into the similarity comparison model, and obtain the normalized semantic space distance through calculation.

[0026] In a possible implementation, the first calculation module includes:

[0027] The input module is used to input the two signal features to be compared in the processed training data set into two twin input layers respectively, and then input them into the feature extraction network for mapping to obtain two feature vectors Embedding;

[0028] The semantic distance generation module is used to input two feature vector embeddings into the distance calculation layer to obtain the semantic distance in the metric space;

[0029] Normalization module, which is used to obtain the contrast loss function and binary cross entropy loss by normalizing the semantic distance in the metric space through sigmoid operation;

[0030] The calculation submodule is used to calculate the dual loss function based on the contrast loss function and the binary cross entropy loss.

[0031] In a possible implementation, the feature extraction network includes two sub-networks, each sub-network includes three residual connection blocks and shares the same parameters.

[0032] In a third aspect, the present application further proposes a computer device, comprising a processor and a memory, wherein the memory stores a computer program, and the computer program is loaded and executed by the processor to implement the radiation source signal similarity evaluation method as described in any one of the first aspects.

[0033] In a fourth aspect, the present application further proposes a computer-readable storage medium, in which a computer program is stored. The computer program is loaded and executed by a processor to implement the radiation source signal similarity evaluation method as described in any one of the first aspects.

[0034] The above main scheme of the present application and its further options can be freely combined to form multiple schemes, all of which are schemes that can be adopted and claimed for protection in the present application; and in the present application, (non-conflicting options) options and other options can also be freely combined. After understanding the scheme of the present application, those skilled in the art can understand that there are many combinations based on the existing technology and common knowledge, all of which are technical schemes to be protected by the present application, and they are not exhaustively listed here.

[0035] The present application discloses a radiation source signal similarity evaluation method, device, equipment and medium, which relate to the field of signal processing technology. Firstly, the acquired radiation source signal is divided into similar sample pairs and dissimilar sample pairs according to the radiation source, which together constitute a training data set. The abnormal samples in the training data set are cleaned and annotated to obtain a processed training data set, which is input into a constructed dual twin network to calculate the dual loss function, and the reverse gradient of the neural network is propagated and the loss function is minimized. After multiple iterative training, a converged similarity comparison model is obtained, and finally two signal features to be compared are input into the similarity comparison model, and the normalized semantic space distance is obtained through calculation, so as to calculate and measure the similarity between the radiation source signal features. The dual twin network can fully extract the differences in signal features in high-dimensional space, which solves the difficulty of threshold setting in traditional similarity evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 A flow chart of a radiation source signal similarity evaluation method proposed in an embodiment of the present application is shown.

[0038] Figure 2 The dual twin network in the embodiment of the present application is shown.

[0039] Figure 3a A comparison chart of the accuracy of the three loss functions proposed in the embodiments of the present application is shown.

[0040] Figure 3b A loss comparison diagram of the three loss functions proposed in the embodiments of the present application is shown.

[0041] Figure 4a A graph showing the change in accuracy during the dual twin network training process proposed in an embodiment of the present application is shown.

[0042] Figure 4bA graph showing the change in the loss function during the dual twin network training process proposed in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0043] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0044] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of this application.

[0045] In the prior art, the traditional similarity evaluation method is mainly based on the distance metric similarity evaluation method, which requires manual design and optimization of features, and the extracted features need to be manually aligned, and the calculation results need to be divided by setting appropriate thresholds. Feature processing and threshold setting require a lot of experience and skills. When faced with new tasks and data, it is difficult to make effective comparisons and judgments. There are disadvantages such as time-consuming, labor-intensive, and highly subjective. Although artificial intelligence technology has achieved remarkable results in the fields of image recognition and natural language processing, traditional models such as neural networks are mainly used for classification tasks, and their effects are not ideal in terms of similarity comparison.

[0046] Therefore, in order to solve the above-mentioned technical problems, the embodiments of the present application propose a method, device, equipment and medium for evaluating the similarity of radiation source signals. Since the traditional method relies on manually designed features or fixed thresholds when processing signal feature similarity judgments, this limits the ability to recognize complex patterns. The present application adopts a deep twin network, which is an advanced method based on neural networks, which extracts more discriminative feature representations from original signals through automatic learning. The features of the two received radiation source signals are taken as input, which means that the network can complete the extraction from the original signal to high-dimensional features and the calculation of the similarity between features in a continuous process. This end-to-end learning framework reduces the reliance on intermediate processing steps and improves efficiency and accuracy.

[0047] In addition, the two branches of the deep twin network in this application process each group of signals separately, and they share weights to ensure the consistency and comparability of feature learning. This process not only extracts the features of the signal, but also optimizes these features to make them more suitable for similarity measurement. The network outputs the normalized distance of the signal features in the semantic space, which is a numerical value reflecting the similarity of the two signals. Through normalization, all comparison results can be mapped to the same scale (such as between 0 and 1), which is convenient for direct comparison and interpretation.

[0048] Compared with traditional methods, this model does not require manual threshold setting to distinguish similar or dissimilar signals, reducing the errors caused by subjective judgment. By training on a large amount of labeled data, the network can automatically learn to distinguish the subtle differences between different signals and achieve more objective similarity judgment. The powerful learning ability of the deep learning model enables the model to learn directly from the original or slightly preprocessed signal features, reducing the dependence on external feature engineering and simplifying the processing flow, which is described in detail below.

[0049] Please refer to Figure 1 , Figure 1 A flow chart of a method for evaluating the similarity of radiation source signals proposed in an embodiment of the present application is shown, and the method includes:

[0050] Step S100: Divide the acquired radiation source signals into similar sample pairs and dissimilar sample pairs according to the radiation sources, and together form a training data set.

[0051] First, signal data is obtained from a variety of radiation sources. These data reflect the unique characteristics of different radiation sources. In order to better reveal the internal structure and differences of the signal, the original data is transformed, such as time-frequency transformation (such as short-time Fourier transform, wavelet transform) or high-order spectrum analysis. These transformations help to extract more discriminative features and improve the model's ability to recognize signal patterns.

[0052] Next, based on the transformed features, the samples are paired and labeled. If two samples come from the same radiation source, they are labeled as a "similar" sample pair and given a distance label of 0, which means that during the network optimization process, the model should learn to bring the feature representations of these samples closer to the same position in the semantic space. On the contrary, if the samples come from different radiation sources, they are labeled as "dissimilar" sample pairs and the distance label is 1, indicating that the model needs to represent these samples as far apart in the feature space.

[0053] Then, a series of similar and dissimilar data pairs are systematically generated to form a diverse and representative training dataset. Such a dataset can enable the model to learn extensive and discriminative features and improve the model's ability to generalize to unseen radiation source signals.

[0054] The distance between similar sample pairs in the normalized semantic space is as close to 0 as possible, while the distance between dissimilar sample pairs is close to 1. This design is directly related to the loss function of the model, that is, the network will minimize the distance between similar sample pairs and maximize the distance between dissimilar sample pairs by adjusting weights and biases, thereby achieving effective evaluation of the similarity of radiation source signal features.

[0055] Step S200: Clean and label abnormal samples in the training data set to obtain a processed training data set.

[0056] Since samples extracted directly from the original signal may contain noise, interference or low-quality data, this part of the data may interfere with model learning and reduce its performance. Therefore, according to system requirements and the characteristics of different radiation sources, signal samples with different qualities generated by the same radiation source are carefully screened. This step ensures the purity and representativeness of the training data.

[0057] By judging, we remove abnormal sample pairs from the training data set that will significantly affect the model training effect. These anomalies may come from measurement errors, environmental interference or other atypical signal performances. If they are included in the training, they may mislead the model learning and lead to a decrease in generalization ability.

[0058] In some cases, the similarity of sample pairs may not be an absolute 0 or 1, but a value somewhere in between. In this case, the distance labels of these sample pairs can be assigned continuous values ​​between 0 and 1 based on actual needs and expert evaluation. Such flexible labeling can reflect the continuity of signal changes in the real world, increase the model's ability to handle ambiguous situations, and improve its adaptability and accuracy in complex environments.

[0059] Step S300: input the processed training data set into the constructed dual twin network to calculate the dual loss function.

[0060] Figure 2 The dual twin network in an embodiment of the present application is shown, and the dual twin network includes two twin input layers, a feature extraction network and a distance calculation layer. The feature extraction network includes two sub-networks, each sub-network includes three residual connection blocks and shares the same parameters.

[0061] It consists of two mirror-symmetric deep neural networks that share the same weight parameter W and have completely consistent feature extraction capabilities. Each sub-network consists of three residual connection blocks. Residual blocks are a structure that effectively alleviates the gradient vanishing problem and promotes deep network training, enabling the network to learn deeper abstract features without losing too much information during back propagation.

[0062] The distance calculation layer is located after the output of the two twin networks. Its function is to receive the output feature vectors (Embedding) of the two sub-networks and calculate the distance between the two feature vectors in the metric space. This distance can be Euclidean distance, cosine similarity, etc., depending on the task requirements and the choice of metric space.

[0063] In order to make the output of the network better match the supervision label (that is, the distance label between 0 and 1), the sigmoid function is used to normalize the output of the distance calculation layer to ensure that the prediction result falls within the probability range, which is convenient for subsequent loss function calculation and optimization.

[0064] Step S300 includes:

[0065] The two signal features to be compared in the processed training data set are input into two twin input layers respectively, and then input into the feature extraction network for mapping to obtain two feature vectors Embedding;

[0066] Input the two feature vectors Embedding into the distance calculation layer to obtain the semantic distance in the metric space;

[0067] The semantic distance in the metric space is normalized by sigmoid to obtain the contrast loss function and binary cross entropy loss;

[0068] The dual loss function is calculated based on the contrastive loss function and the binary cross entropy loss.

[0069] The network accepts a pair of signal samples as input, one sample enters the upper sub-network and the other enters the lower sub-network. Even if the samples come from the same source, they can be transformed or collected under different conditions and mapped to the feature space through the twin network. Each sub-network automatically extracts and enhances the key features of the input data through residual blocks. These feature vectors capture the essential properties of the signal.

[0070] By calculating the distance between two feature vectors and using a dual loss function (such as Contrastive Loss or Triplet Loss) to minimize the distance between similar sample pairs and maximize the distance between dissimilar sample pairs, the network gradually learns how to more accurately distinguish the signal characteristics of different radiation sources during training. The upper and lower sub-networks share the weight parameter W, which means that the network has a consistent processing of sample transformations, so that it can focus more on learning to distinguish the subtle features of different signals rather than the differences in specific instances, enhancing the generalization ability of the model.

[0071] The contrastive loss function pushes the network to learn a feature space where similar samples are close together and dissimilar samples are far apart. for: ,in is the number of samples, is the label of whether two samples match. is the Euclidean distance between two sample features, To set the threshold.

[0072] , represents the Euclidean distance between two sample features, is the label of whether two samples match. It means that the two samples are similar. represents dissimilarity, m is a set threshold (in the embodiment of the present invention, m=1), is the number of samples.

[0073] When the samples are similar, , the loss function is That is, the greater the distance between features, the greater the loss function. If the Euclidean distance D of originally similar samples in the feature space is large, it means that the current model feature extraction is not good, and it is necessary to increase the loss and further optimize the neural network parameters in the direction of smaller feature space distance.

[0074] When the samples are not similar, , the loss function is , that is, in Within the range, the smaller the distance between features, the larger the loss function. When the distance is the threshold m, the loss can be considered to be 0. When the samples are dissimilar, the Euclidean distance in the feature space should be larger. If the distance is small, the network will have a large loss value and the supervision model will be optimized in the direction of increasing the distance.

[0075] The contrast loss function essentially determines the feature extraction direction of the neural network. The goal is to make the distance between the extracted features of similar pairs of samples as small as possible, and the distance between the extracted features of dissimilar pairs of samples as large as possible. It is equivalent to redefining the distance between the two samples in the high-dimensional space. Compared with the method of directly calculating the distance in the original data domain, this mapping to the high-dimensional feature space is supervised and trained. Therefore, the mentioned features will be more discriminative, which can make originally similar samples closer in the feature space and originally dissimilar samples farther apart in the feature space.

[0076] Binary Cross Entropy Loss for: ,in is the sample size, is the true label, The label value predicted by the neural network.

[0077] is the real label 1 or 0, representing a binary classification problem. In order to unify with the distance metric from a physical concept, the present invention defines similar labels as 0 and dissimilar labels as 1. The labels represent the binary semantic distance between samples in a physical sense. = 1, the samples are not similar. , The closer it is to 1, the smaller the loss function value. The closer it is to 0, the larger the loss function value. = 0, the samples are similar, , The closer it is to 0, the smaller the loss function value is. The closer it is to 1, the larger the loss function value.

[0078] The above two losses are combined into a total loss function according to a certain weight (half each in this case). This is done to take into account both the feature similarity measure and classification accuracy. The dual loss function consists of two parts: the contrastive loss with margin and the binary cross entropy loss: .

[0079] By constructing a dual loss function, the characteristics of the two loss functions are fully utilized to comprehensively optimize the twin network. The contrast loss function ensures the similarities and differences of samples in the feature space from a micro level, while the binary cross entropy loss further constrains the final output of the network to be consistent with the label from a macro level.

[0080] In a possible embodiment, the present application verifies the convergence of the network model using a single loss function and a dual loss function through a set of ablation experiments. Figure 3a A comparison chart of the accuracy of the three loss functions proposed in the embodiments of the present application is shown. Figure 3b The loss comparison diagram of the three loss functions proposed in the embodiment of the present application is shown. The solid line in the figure represents the dual loss proposed by the present invention, and the two dotted lines represent the accuracy and loss changes of the network when a certain loss function is used alone.

[0081] Using contrast loss alone: ​​It mainly learns by optimizing the feature distance between samples, emphasizing the aggregation of similar samples and pushing away dissimilar samples. This helps the network learn a feature representation space with better discrimination, but may lack in the precise classification decision boundary.

[0082] Using binary cross entropy loss alone emphasizes the classification accuracy of samples, and works by maximizing the log-likelihood probability of correctly classified samples. Although this method can optimize classification performance, its ability to distinguish between similarities and dissimilarities in feature space may not be as good as a dedicated distance metric loss.

[0083] The dual loss function (a combination of contrastive loss and binary cross entropy loss) is used in combination. This configuration not only promotes the effective organization of sample distances in the feature space (through contrastive loss), but also strengthens the classification ability of the model through binary cross entropy loss, ensuring the optimization of the predicted probability distribution. The results show that this combination can better utilize the label information in the data, allowing the model to distinguish between similar and dissimilar samples while also improving the accuracy of sample classification.

[0084] Step S400: propagate the dual loss function through the neural network reverse gradient and minimize the loss function, and obtain a converged similarity comparison model after multiple iterative training.

[0085] Using gradient descent or other optimization algorithms, the gradient of the loss relative to the network parameters is calculated according to the combined dual loss function, and the network parameters are updated. This process is repeated multiple times on the entire training data set until the loss function reaches a minimum or the stopping condition is met. As the training progresses, the model gradually learns how to extract features from the input signal that can effectively distinguish the similarity and dissimilarity of samples, thereby reaching a better performance state after multiple iterations, that is, the model converges.

[0086] Step S500: input two signal features to be compared in the training data set into a similarity comparison model, and obtain a normalized semantic space distance through calculation.

[0087] After the model training is completed, for any two new signals that need to be compared, their feature vectors are first passed as input to the trained similarity comparison model. After processing these feature vectors, a normalized semantic space distance is output. , this distance is a scalar value that reflects the relative position or degree of difference between two feature vectors in the high-dimensional feature space. Normalization means The value range of is usually constrained between 0 and 1 for easy understanding and application.

[0088] according to The definition of When it is close to 1, it indicates that the two signal features are far apart in the semantic space, and they can be considered dissimilar; on the contrary, if If it is close to 0, it means that the characteristics of the two signals are very similar and can be considered similar. , we can directly obtain a value reflecting the similarity of signal features. The closer the value is to 1, the more similar the signals are, and the closer it is to 0, the less similar the signals are.

[0089] In a possible embodiment, in order to prove the effectiveness of the method, the proposed method is verified by measured data: first, a data set is created using historically accumulated measured data of various radiation sources to construct a set of similar / dissimilar sample pairs in the entire sample space, and the training data and test data are randomly split in proportion, with the ratio of training data set to test data being 3:1.

[0090] Then, a residual-connected deep twin network with shared weights is constructed. The number of residual blocks is set to 3 for feature extraction of the input signal. The extracted features are 32-dimensional Embedding vectors. The contrast loss and binary cross entropy loss are calculated as dual losses to jointly optimize the network. Then, a large number of signal features are used to train the network model for 100 epochs until the network converges. Figure 4a The figure shows the accuracy change during the dual twin network training process proposed in the embodiment of the present application. Figure 4b The loss function change diagram during the dual twin network training process proposed in the embodiment of the present application is shown. After the training is completed, the test set is processed and the results of the similarity evaluation are statistically analyzed. Table 1 shows a pulse feature similarity comparison table:

[0091] Table 1

[0092]

[0093] It represents the two signal pattern pairs to be compared. The values ​​in the upper row are the true values ​​of the similarity of the signal pattern pairs, and the values ​​below are the similarity judgment results of the dual twin network ( ), the higher the similarity value is, the more similar the two signals are. It can be seen that the method proposed in the present invention can effectively distinguish similar / dissimilar situations and shows good judgment performance.

[0094] In summary, the method proposed in this application can fully extract the feature differences of signal features in high-dimensional space, and based on this, achieve fast and accurate similarity judgment, proving its effectiveness.

[0095] In this application, by introducing the dual loss function, this method can simultaneously consider the contrast loss (strengthening the distinction between similar and dissimilar samples in the feature space) and the binary cross entropy loss (optimizing the classification decision boundary) during the training process. This design can more comprehensively explore and utilize the complex correlations between signal features, and improve the accuracy and robustness of the model in similarity evaluation.

[0096] Twin network architecture: Using twin networks, one network processes the features of one signal and the other processes the features of another signal. They share weights, which not only simplifies the model structure and reduces computational complexity, but also allows parallel processing, significantly improving computational efficiency. In addition, shared weights ensure consistency in the processing of input signals, which is conducive to capturing purer signal differences rather than deviations caused by differences in network structure.

[0097] Feature extraction and high-dimensional space differences: In high-dimensional feature space, the model can express the differences in signal features more finely. This is due to the powerful nonlinear expression ability of deep neural networks. Even without strict feature alignment, the similarity between signals can be effectively evaluated, solving the problem of dependence on precise feature alignment in traditional methods.

[0098] Data-driven similarity calculation: Unlike the traditional method of manually setting thresholds, this invention uses a large amount of labeled data for training, allowing the model to automatically learn how to calculate the similarity between signals. This not only avoids the subjectivity and limitations of manually setting thresholds, but also makes full use of the patterns and relationships in historical data, making the model's similarity assessment of new samples more accurate.

[0099] Reducing the need for feature alignment: Since the model can learn similarities from higher-level feature representations, the need for strict alignment of original signal features is reduced, which is particularly beneficial for complex signal processing in practical applications and improves the practicality and versatility of the method.

[0100] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0101] First, compared with traditional similarity evaluation methods, this application can make full use of historical annotation information and expert experience, complete end-to-end signal feature similarity calculation in a data-driven manner, and can automatically learn high-dimensional signal features, avoiding the defects of manual feature processing.

[0102] Second, the twin network structure adopted can process multiple groups of signal features to be compared in parallel at the same time, which greatly improves the efficiency compared with the traditional serial calculation method.

[0103] Third, the designed dual loss can extract the features of the data more comprehensively and improve the accuracy of similarity calculation.

[0104] Fourth, it has good generalization and transfer learning capabilities, and can be effectively applied to various other types of data such as images and speech.

[0105] A possible implementation of a radiation source signal similarity evaluation device is provided below, which is used to execute each execution step and corresponding technical effect of the radiation source signal similarity evaluation method shown in the above embodiment and possible implementation, and the device includes:

[0106] A sample pair division module is used to divide the acquired radiation source signals into similar sample pairs and dissimilar sample pairs according to the radiation sources, which together form a training data set;

[0107] A cleaning and labeling module is used to clean and label abnormal samples in the training data set to obtain the processed training data set;

[0108] A first computing module is used to input the processed training data set into the constructed dual Siamese network to calculate the dual loss function;

[0109] The training module is used to propagate the dual loss function through the reverse gradient of the neural network and minimize the loss function. After multiple iterations of training, a converged similarity comparison model is obtained;

[0110] The second calculation module is used to input two signal features to be compared in the training data set into the similarity comparison model, and obtain the normalized semantic space distance through calculation.

[0111] In a possible embodiment, the first computing module includes:

[0112] The input module is used to input the two signal features to be compared in the processed training data set into two twin input layers respectively, and then input them into the feature extraction network for mapping to obtain two feature vectors Embedding;

[0113] The semantic distance generation module is used to input two feature vector embeddings into the distance calculation layer to obtain the semantic distance in the metric space;

[0114] Normalization module, which is used to obtain the contrast loss function and binary cross entropy loss by normalizing the semantic distance in the metric space through sigmoid operation;

[0115] The calculation submodule is used to calculate the dual loss function based on the contrast loss function and the binary cross entropy loss.

[0116] In a possible embodiment, the feature extraction network includes two sub-networks, each sub-network includes three residual connection blocks and shares the same parameters.

[0117] This preferred embodiment provides a computer device, which can implement the steps in any embodiment of the radiation source signal similarity evaluation method provided in the embodiments of the present application. Therefore, the beneficial effects of the radiation source signal similarity evaluation method provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0118] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by controlling related hardware through instructions, and the instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. To this end, an embodiment of the present application provides a storage medium, in which a plurality of instructions are stored, and the instructions can be loaded by a processor to execute the steps of any embodiment of the radiation source signal similarity evaluation method provided in the embodiment of the present application.

[0119] The storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0120] Since the instructions stored in the storage medium can execute the steps in any radiation source signal similarity evaluation method embodiment provided in the embodiments of the present application, the beneficial effects that can be achieved by any radiation source signal similarity evaluation method provided in the embodiments of the present application can be achieved. For details, please refer to the previous embodiments and will not be repeated here.

[0121] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for evaluating the similarity of radiation source signals, characterized in that: The method comprises: The acquired radiation source signals are paired and marked according to the radiation source. If two samples come from the same radiation source, they are marked as similar sample pairs. If the samples come from different radiation sources, they are marked as dissimilar sample pairs. Similar sample pairs and dissimilar sample pairs are divided to form a training data set. The radiation source signals reflect the characteristics of different radiation sources. The training data set is used to improve the ability of the model to generalize to unseen radiation source signals. According to system requirements and the characteristics of different radiation sources, signal samples with different qualities generated by the same radiation source are screened, and abnormal samples in the training data set are cleaned and labeled to obtain the processed training data set; Input the processed training data set into the constructed dual Siamese network to calculate the dual loss function; The dual twin network includes two twin input layers, a feature extraction network, and a distance calculation layer. The steps of inputting the processed training data set into the constructed dual twin network to calculate the dual loss function include: The two signal features to be compared in the processed training data set are input into two twin input layers respectively, and then input into the feature extraction network for mapping to obtain two feature vectors Embedding; Input the two feature vectors Embedding into the distance calculation layer to obtain the semantic distance in the metric space; The semantic distance in the metric space is normalized by sigmoid to obtain the contrast loss function and binary cross entropy loss; The dual loss function is calculated based on the contrast loss function and the binary cross entropy loss. The dual loss function is used to minimize the distance between similar sample pairs and maximize the distance between dissimilar sample pairs. The dual twin network gradually learns to distinguish the signal characteristics of different radiation sources during the training process. The dual loss function is propagated through the reverse gradient of the neural network and the loss function is minimized. After multiple iterative training, a converged similarity comparison model is obtained. The two signal features to be compared in the training data set are input into the similarity comparison model, and the normalized semantic space distance is calculated to calculate and measure the similarity between the radiation source signal features.

2. The method for evaluating the similarity of radiation source signals according to claim 1, wherein: The feature extraction network consists of two sub-networks, each of which consists of three residual connection blocks and shares the same parameters.

3. The radiation source signal similarity evaluation method according to claim 1, characterized in that: Contrastive loss function for: ,in is the number of samples, Is the label of whether two samples match, is the Euclidean distance between two sample features, To set the threshold.

4. The radiation source signal similarity evaluation method according to claim 1, characterized in that: Binary Cross Entropy Loss for: ,in is the sample size, is the true label, The label value predicted by the neural network.

5. A radiation source signal similarity evaluation device, characterized in that: The device comprises: A sample pair division module is used to pair and mark the acquired radiation source signal samples according to the radiation source. If two samples come from the same radiation source, they are marked as similar sample pairs. If the samples come from different radiation sources, they are marked as dissimilar sample pairs. Similar sample pairs and dissimilar sample pairs are divided to form a training data set. The radiation source signal reflects the characteristics of different radiation sources. The training data set is used to improve the ability of the model to generalize to unseen radiation source signals. The cleaning and labeling module is used to screen signal samples with different qualities generated by the same radiation source according to system requirements and the characteristics of different radiation sources, and to clean and label abnormal samples in the training data set to obtain the processed training data set; The input module is used to input the two signal features to be compared in the processed training data set into two twin input layers respectively, and then input them into the feature extraction network for mapping to obtain two feature vectors Embedding; The semantic distance generation module is used to input two feature vector embeddings into the distance calculation layer to obtain the semantic distance in the metric space; Normalization module, which is used to obtain the contrast loss function and binary cross entropy loss by normalizing the semantic distance in the metric space through sigmoid operation; The calculation submodule is used to calculate the dual loss function based on the contrast loss function and the binary cross entropy loss. The dual loss function is used to minimize the distance between similar sample pairs and maximize the distance between dissimilar sample pairs. The dual twin network gradually learns to distinguish the signal characteristics of different radiation sources during the training process. The training module is used to propagate the dual loss function through the reverse gradient of the neural network and minimize the loss function. After multiple iterative training, a converged similarity comparison model is obtained. The second calculation module is used to input two signal features to be compared in the training data set into the similarity comparison model, and obtain the normalized semantic space distance through calculation, so as to calculate and measure the similarity between the radiation source signal features.

6. The radiation source signal similarity evaluation device according to claim 5, characterized in that: The feature extraction network consists of two sub-networks, each of which consists of three residual connection blocks and shares the same parameters.

7. A computer device, characterized in that: The computer device includes a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the radiation source signal similarity evaluation method according to any one of claims 1 to 4.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, which is loaded and executed by a processor to implement the radiation source signal similarity evaluation method according to any one of claims 1 to 4.

Images