A radar operating mode recognition algorithm based on multiple adversarial domains and metric learning

Abstract

The present invention discloses a radar working mode recognition algorithm based on multi-adversarial domains and metric learning, comprising the following steps: S1, a source domain radar data sequence with labels and a target domain radar data sequence without labels are input into a feature extractor together; S2, features and inter-class differences of source domain and target domain mixed data are extracted by the feature extractor; S3, the source domain and target domain mixed data features obtained in the feature extractor are input into a feature classifier to obtain a feature classification result, and a classification label of the target domain radar data sequence without labels, a domain discrimination loss, and a source domain classification loss are obtained; S4, the source domain and target domain mixed data features obtained in the feature extractor are input into a domain discriminator to perform multi-feature domain discrimination in the form of data fragments, and a subdomain classification result and a subdomain classification loss are obtained; S5, the model is iteratively solved to finally obtain a high-confidence target domain radar data classification label.

Description

Technical Field

[0001] The present invention relates to the field of radar technology, and in particular to a radar operating mode recognition algorithm based on multiple adversarial domains and metric learning. Background Art

[0002] With the application and development of science and technology in the military, and the increasingly complex battlefield environment, radar jamming countermeasures have gradually become the core of modern military warfare, and radar operating mode recognition is a top priority. Radar operating mode recognition, namely passive detection systems such as RWR (Radar Warning Receiver), ELINTR (Electronic Intelligence Receiver), and ESM (Electronic Support Measurement), identifies the current operating mode of an enemy radar by intercepting and analyzing enemy radar signals.

[0003] In recent years, radar working pattern recognition technology has developed faster and faster, with better recognition effects and higher recognition accuracy. However, almost no one has set foot in the field of radar working pattern recognition without prior knowledge. The difficulties of radar working pattern recognition without prior knowledge are mainly reflected in three aspects: (1) Overall data distribution offset. Overall data distribution offset means that the overall data distribution of the source domain and the target domain is usually not exactly the same. (2) Data distribution offset between categories. Data distribution offset between categories means that the data distribution of the same category between different domains is often different. Regardless of which domain the features come from, features of the same category should be close to each other, and features of different categories should be far apart. (3) The diversity and complexity of time series data. The difficulty brought about by the diversity and complexity of time series data is that the various parameter indicators of the radar signal constitute different channels with different data distributions. The time series change trends of different channels jointly explain the working mode of the radar. Summary of the Invention

[0004] The problem to be solved by the present invention is to provide a radar working mode recognition algorithm based on multi-adversarial domains and metric learning, which can make different categories of target domains move away from classification boundaries and obtain more discriminative features, thereby improving the accuracy of radar working mode recognition of target domain samples.

[0005] The present invention provides a technical solution to solve the above problems: a radar working mode recognition algorithm based on multiple adversarial domains and metric learning, the algorithm includes the following steps:

[0006] S1. The source domain radar data sequence with labels and the target domain radar data sequence without labels are input into the feature extractor together;

[0007] S2, extract the features and inter-class differences of the mixed data of the source domain and the target domain through the feature extractor;

[0008] S3. The mixed data features of the source domain and target domain obtained in the feature extractor are input into the feature classifier to obtain the feature classification result, and the classification label, domain discrimination loss, and source domain classification loss of the target domain radar data sequence without labels are obtained;

[0009] S4. The mixed data features of the source domain and target domain obtained in the feature extractor are input to the domain discriminator, which performs multi-feature domain discrimination in the form of data fragments to obtain the subdomain classification results and subdomain classification loss;

[0010] S5. Based on the above inter-class differences, domain discrimination loss, source domain classification loss, and subdomain classification loss, the model is iteratively solved to finally obtain a highly reliable target domain radar data classification label.

[0011] Preferably, the feature extractor in S1 is G f , is a radar working mode recognition network that combines the BiGRU recurrent neural network trained on the source domain radar data sequence with the attention mechanism. When the input is mixed data of the source domain and target domain radar sequences, it is necessary to add a weight layer θ after the network feature output layer. f Fine-tune the feature extraction network.

[0012] Preferably, the feature of the mixed data in S2 is the output feature vector of the feature classifier recurrent network BiGRU, and the inter-class difference metric function L lj The formula is shown in the following formula (1):

[0013]

[0014] In formula (1), I N×N Indicates that the current mixed output feature of the feature extractor is N features, and each feature vector is N-dimensional; f i s represents the features extracted from the i-th sample data in the source domain after passing through the feature extractor, f i sT represents the transpose of the extracted feature matrix, ||·||2 represents the 2-norm of the matrix, and m is the number of samples in the radar data sequence.

[0015] Preferably, the feature classifier in S3 is G y , whose control weight is θ y , the classification label is obtained by the Softmax classification layer with weights to obtain the radar working mode classification label vector; the domain discrimination loss metric function L y The formula is shown in formula (2):

[0016]

[0017] In formula (2), m is the number of radar data sequence samples, represents the i-th data in the source domain, represents the label of the i-th data in the source domain, G f (·) indicates feature extraction of data, G y (·) indicates feature classification of data, L c is the cross entropy function; the source domain classification loss is the improved additive cosine margin loss metric function L gams , and its formula is shown in the following formula (3):

[0018]

[0019] In formula (3), m represents the number of radar data sequence samples, C represents the number of radar sequence classifications, S represents the cosine coefficient, k represents the cosine interval, and θ yi represents the angle between the source domain radar sample and its own label weight, θ j Represents the angle between the source domain radar sample and the weight of the non-self label.

[0020] Preferably, the domain discriminator in S4 is G di ,i=1,2,...,I, its control weight is θ di ,i=1,2,...,I, the domain classification result is to judge whether the current feature belongs to the source domain or the target domain, and its subdomain classification loss is L ll , and its formula is shown in the following formula (4):

[0021]

[0022] In formula (4), G di is the identification network function for the i-th domain, For sample x j The nth component of the pseudo label vector.

[0023] Preferably, the iterative solution process in S5 is as follows:

[0024] In the early stage of training, the grey correlation function is used to calculate the source domain sample S i and the target domain sample T j The gray correlation degree between the target domain sample and the source domain sample is used as the initial pseudo label of the target domain sample: the label generated in each iteration is The calculation formula is shown in the following formula (5):

[0025]

[0026] Among them, N epochis the number of iterations required for model training, q is the number of current iterations, is the pseudo label of the previous iteration, is the initial label;

[0027] Iterative process ①: When the feature extractor inputs the source domain sample, it adjusts the parameter θ f and θ y , minimize the loss as much as possible, improved additive cosine margin loss function L gams value, making it increase;

[0028] Iterative process②: After inputting the mixed data of source domain and target domain, in order to improve the accuracy of identifying the sample domains of source domain and target domain, the parameter θ is adjusted f , minimize the domain classification loss L y ;

[0029] Iterative process ③: In order to increase the interval between the features of samples of different categories, that is, to increase the separability, by adjusting the parameter θ f , minimize the domain classification loss L lj ;

[0030] Iterative process ④: In order to reduce the feature difference between the source domain and the target domain samples of the same type, that is, to reduce the intra-class distance, by adjusting the parameter θ di ,i=1,2,...,I, maximize the domain classification loss L ll ;

[0031] Adjustment parameter θ di Make the domain classification loss L ll Increasing will make it easier for the subdomain discriminator to determine whether the data comes from the source domain or the target domain, and adjusting the parameter θ f To reduce the gap between domains, adjust θ f Value and θ di The value is constantly in conflict, when adjusting θ di Make the subdomain classification loss L ll Reaching the maximum value, at this time if the subdomain discriminator G di When it is impossible to determine whether the radar data sequence comes from the source domain or the target domain, the features extracted by the feature extractor are common features of the target domain and source domain radars, and the feature classifier can make more accurate classification predictions for the unlabeled target domain radar data.

[0032] Preferably, the grey relational function is as shown in the following formula (6):

[0033]

[0034] In formula (6), n represents the number of radar sequence division units, The physical feature displacement difference that represents the similarity between the source domain and the target domain is expressed as follows:

[0035]

[0036] In formula (6) The physical characteristic velocity difference (first-order slope difference) that represents the similarity between the source domain and the target domain is expressed as follows:

[0037]

[0038] In formula (6) The physical feature acceleration difference (second-order slope difference) that represents the similarity between the source domain and the target domain is expressed as follows:

[0039]

[0040] Compared with existing technologies, the present invention offers the following advantages: Compared to traditional radar operating mode recognition methods, the present invention proposes a domain adaptation algorithm for radar operating mode recognition based on multiple adversarial domains and metric learning. During training, the extracted features are constrained by inter-class differences. Combined with multiple adversarial domains and metric learning, this algorithm reduces the distance between samples of the same class and increases the distance between samples of different classes. This allows different classes in the target domain to move away from classification boundaries, resulting in more discriminative features and thus improving the accuracy of radar operating mode recognition for target domain samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings described herein are used to provide further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0042] Figure 1 Schematic diagram of the radar working mode recognition model based on multiple adversarial domains and metric learning of the present invention;

[0043] Figure 2 This is a structural diagram of the improved additive cosine interval loss metric function of the present invention. DETAILED DESCRIPTION

[0044] The following will describe the embodiments of the present invention in detail with reference to the accompanying drawings and examples, so that the implementation process of how the present invention applies technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0045] In the description of the present invention, it should be noted that, for directional words, such as the terms "center", "horizontal", "longitudinal", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and so on, indicating directions and positional relationships, are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and cannot be understood as limiting the specific scope of protection of the present invention.

[0046] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features. Therefore, the terms "first" and "second" may explicitly or implicitly refer to one or more of these features. Throughout the description of the present invention, "several" means two or more, unless otherwise specifically defined.

[0047] In the present invention, unless otherwise specified or limited, the terms "assemble," "connect," and "connect" should be understood in a broad sense. For example, they may refer to fixed connection, detachable connection, or integral connection; mechanical connection; direct connection, connection through an intermediate medium, or internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0048] It will be understood that when used in this specification and the appended claims, the terms “comprises” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0049] It should also be understood that the terms used in this description of the embodiments of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the embodiments of the present invention. As used in the description of the embodiments of the present invention and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0050] This embodiment discloses a radar working mode recognition algorithm based on multiple adversarial domains and metric learning, the algorithm comprising the following steps:

[0051] S1. The source domain radar data sequence with labels and the target domain radar data sequence without labels are input into the feature extractor together;

[0052] S2, extract the features and inter-class differences of the mixed data of the source domain and the target domain through the feature extractor;

[0053] S3. The mixed data features of the source domain and target domain obtained in the feature extractor are input into the feature classifier to obtain the feature classification result, and the classification label, domain discrimination loss, and source domain classification loss of the target domain radar data sequence without labels are obtained;

[0054] S4. The mixed data features of the source domain and target domain obtained in the feature extractor are input to the domain discriminator, which performs multi-feature domain discrimination in the form of data fragments to obtain the subdomain classification results and subdomain classification loss;

[0055] S5. Based on the above inter-class differences, domain discrimination loss, source domain classification loss, and subdomain classification loss, the model is iteratively solved to finally obtain a highly reliable target domain radar data classification label.

[0056] Among them, the feature extractor in S1 is G f , is a radar working mode recognition network that combines the BiGRU recurrent neural network trained on the source domain radar data sequence with the attention mechanism. When the input is mixed data of the source domain and target domain radar sequences, it is necessary to add a weight layer θ after the network feature output layer. f Fine-tune the feature extraction network.

[0057] More specifically, the feature of the mixed data in S2 is the output feature vector of the feature classifier recurrent network BiGRU, and the inter-class difference metric function L lj The formula is shown in the following formula (1):

[0058]

[0059] In formula (1), I N×N Indicates that the current mixed output feature of the feature extractor is N features, and each feature vector is N-dimensional; f i s represents the features extracted from the i-th sample data in the source domain after passing through the feature extractor, f i sT represents the transpose of the extracted feature matrix, ||·||2 represents the 2-norm of the matrix, and m is the number of samples in the radar data sequence.

[0060] Among them, the feature classifier in S3 is G y , whose control weight is θ y , the classification label is obtained by the Softmax classification layer with weights to obtain the radar working mode classification label vector; the domain discrimination loss metric function L y The formula is shown in formula (2):

[0061]

[0062] In formula (2), m is the number of radar data sequence samples, represents the i-th data in the source domain, represents the label of the i-th data in the source domain, G f (·) indicates feature extraction of data, G y (·) indicates feature classification of data, L c is the cross entropy function; the source domain classification loss is the improved additive cosine margin loss metric function L gams , and its formula is shown in the following formula (3):

[0063]

[0064] In formula (3), m represents the number of radar data sequence samples, C represents the number of radar sequence classifications, S represents the cosine coefficient, k represents the cosine interval, and θ yi represents the angle between the source domain radar sample and its own label weight, θ j Represents the angle between the source domain radar sample and the weight of the non-self label.

[0065] Among them, the domain discriminator in S4 is G di ,i=1,2,...,I, its control weight is θ di ,i=1,2,...,I, the domain classification result is to judge whether the current feature belongs to the source domain or the target domain, and its subdomain classification loss is L ll , and its formula is shown in the following formula (4):

[0066]

[0067] In formula (4), G di is the identification network function for the i-th domain, For sample x j The nth component of the pseudo label vector.

[0068] Preferably, the iterative solution process in S5 is as follows:

[0069] In the early stage of training, the grey correlation function is used to calculate the source domain sample S i and the target domain sample T j The grey correlation between

[0070] The grey relational function is shown in the following formula (5):

[0071]

[0072] In formula (5), n represents the number of radar sequence division units, The physical feature displacement difference that represents the similarity between the source domain and the target domain is expressed as follows:

[0073]

[0074] In formula (5) The physical characteristic velocity difference (first-order slope difference) that represents the similarity between the source domain and the target domain is expressed as follows:

[0075]

[0076] In formula (5) The physical feature acceleration difference (second-order slope difference) that represents the similarity between the source domain and the target domain is expressed as follows:

[0077]

[0078] The maximum grey correlation between the target domain sample and the source domain sample is used as the initial pseudo label of the target domain sample: the label generated in each iteration is The calculation formula is shown in the following formula (9):

[0079]

[0080] Among them, N epoch is the number of iterations required for model training, q is the number of current iterations, is the pseudo label of the previous iteration, is the initial label;

[0081] Iterative process ①: When the feature extractor inputs the source domain sample, it adjusts the parameter θ f and θ y , minimize the loss as much as possible, improved additive cosine margin loss function L gams value, making it increase;

[0082] Iterative process②: After inputting the mixed data of source domain and target domain, in order to improve the accuracy of identifying the sample domains of source domain and target domain, the parameter θ is adjusted f , minimize the domain classification loss L y ;

[0083] Iterative process ③: In order to increase the interval between the features of samples of different categories, that is, to increase the separability, by adjusting the parameter θ f , minimize the domain classification loss L lj ;

[0084] Iterative process ④: In order to reduce the feature difference between the source domain and the target domain samples of the same type, that is, to reduce the intra-class distance, by adjusting the parameter θ di ,i=1,2,...,I, maximize the domain classification loss Lll ;

[0085] Adjustment parameter θ di Make the domain classification loss L ll Increasing will make it easier for the subdomain discriminator to determine whether the data comes from the source domain or the target domain, and adjusting the parameter θ f To reduce the gap between domains, adjust θ f Value and θ di The value is constantly in conflict, when adjusting θ di Make the subdomain classification loss L ll Reaching the maximum value, at this time if the subdomain discriminator G di When it is impossible to determine whether the radar data sequence comes from the source domain or the target domain, the features extracted by the feature extractor are common features of the target domain and source domain radars, and the feature classifier can make more accurate classification predictions for the unlabeled target domain radar data.

[0086] Among them, it should be noted that Figure 1 The method shown is a radar working mode recognition method based on multi-adversarial domains and metric learning. The source domain and target domain radar data sequences are input into the feature extractor to extract features, and a layer of weights is added after the feature extractor output layer to measure the difference between classes. The extracted features are input into the feature classifier for classification, and the weight parameters in the feature classifier are trained by the gradient descent method to finally train a feature classifier suitable for the target domain. The extracted features are input into the multi-adversarial domain discriminator for domain discrimination, and the domain discriminator is enhanced by the gradient descent method so that the domain discriminator can prepare to discriminate whether the features extracted by the feature extractor belong to the source domain or the target domain. The feature extractor weights are re-modulated so that the newly extracted features can confuse the multi-adversarial domain discriminator. The feature extractor domain and the multi-adversarial domain discriminator continuously confront the optimization parameters, and finally the domain discriminator can no longer judge whether the features extracted in the feature extractor belong to the source domain or the target domain. At this time, the features extracted by the feature extractor are the features shared by the source domain and the target domain.

[0087] Figure 2 The figure shows a schematic diagram of the improved additive cosine metric function space, where x1 represents the input data space vector, W1 represents the clustering vector family of features of the same type as x1, and W2 represents the clustering vector family of features of different types from x1. The improved additive cosine metric function reduces the clustering angle of features of the same type as x1 while increasing the angle with features of non-same type, that is, it reduces the intra-class distance while increasing the inter-class distance.

[0088] Compared to traditional radar mode recognition methods, this paper proposes a domain adaptation algorithm for radar mode recognition based on multiple adversarial domains and metric learning. During training, the extracted features are constrained by inter-class differences. Combined with multiple adversarial domains and metric learning, this algorithm reduces the distance between samples of the same class and increases the distance between samples of different classes. This allows different classes in the target domain to move away from classification boundaries, resulting in more discriminative features and improving the accuracy of radar mode recognition for target domain samples.

[0089] The above description is merely a description of the preferred embodiment of the present invention and is not to be construed as limiting the scope of the claims. The present invention is not limited to the above embodiment, and variations in the specific structure are permitted. Any variations within the scope of the independent claims of the present invention are also within the scope of protection of the present invention.

Claims

1. A radar operating mode recognition algorithm based on multiple adversarial domains and metric learning, characterized by: The algorithm comprises the following steps, S1. The source domain radar data sequence with labels and the target domain radar data sequence without labels are input into the feature extractor together; S2, extract the features and inter-class differences of the mixed data of the source domain and the target domain through the feature extractor; S3. The mixed data features of the source domain and target domain obtained in the feature extractor are input into the feature classifier to obtain the feature classification result, and the classification label, domain discrimination loss, and source domain classification loss of the target domain radar data sequence without labels are obtained; S4. The mixed data features of the source domain and target domain obtained in the feature extractor are input to the domain discriminator, which performs multi-feature domain discrimination in the form of data fragments to obtain the subdomain classification results and subdomain classification loss; S5. Based on the above inter-class differences, domain discrimination loss, source domain classification loss, and subdomain classification loss, the model is iteratively solved to ultimately obtain a highly reliable target domain radar data classification label. The feature of the mixed data in S2 is the output feature vector of the feature classifier recurrent network BiGRU, and the inter-class difference metric function The formula is shown in the following formula (1): (1) In formula (1) Indicates that the current mixed output feature of the feature extractor is features, each feature vector is N-dimensional; Indicates the source domain The features extracted by the feature extractor after the sample data passes through the feature extractor, represents the transpose of the extracted feature matrix, represents the 2-norm of the matrix, is the number of radar data sequence samples; The feature classifier in S3 is , and its control weight is The classification label is obtained by the weighted Softmax classification layer to obtain the radar working mode classification label vector; the domain discrimination loss metric function The formula is shown in formula (2): (2) In formula (2) is the number of radar data sequence samples, Indicates the source domain data, Indicates the source domain The labels of the data, Indicates feature extraction of data. Indicates feature classification processing of data. is the cross entropy function; the source domain classification loss is the improved additive cosine margin loss metric function , and its formula is shown in the following formula (3): (3) In formula (3) represents the number of radar data sequence samples, represents the number of radar sequence classifications, represents the cosine coefficient, represents the cosine interval, Represents the angle between the source domain radar sample and the label weight it carries. Represents the angle between the source domain radar sample and the weight of the non-self label; The domain discriminator in S4 is , and its control weight is , the domain classification result is to judge whether the current feature belongs to the source domain or the target domain, and its subdomain classification loss is , and its formula is shown in the following formula (4): (4) In formula (4), For the A domain identification network function, For samples The pseudo label vector A portion.

2. The radar operating mode recognition method based on multiple adversarial domains and metric learning according to claim 1 is characterized in that: The feature extractor in S1 is It is a radar working mode recognition network that combines the BiGRU recurrent neural network trained on the source domain radar data sequence with the attention mechanism. When the input is mixed data of the source domain and target domain radar sequences, it is necessary to add a weight layer after the network feature output layer. Fine-tune the feature extraction network.

3. The radar operating mode recognition method based on multiple adversarial domains and metric learning according to claim 1 is characterized in that: The iterative solution process in S5 is as follows: In the early stage of training, the grey correlation function is used to calculate the source domain samples. and target domain samples The gray correlation degree between the target domain sample and the source domain sample is used as the initial pseudo label of the target domain sample: the label generated in each iteration is , and its calculation formula is shown in the following formula (5): (5) in, is the number of iterations required for model training, is the number of current iterations, is the pseudo label of the previous iteration, is the initial label; Iterative process ①: When the feature extractor inputs the source domain sample, it adjusts the parameters and Minimize loss as much as possible, improved additive cosine margin loss function value, making it increase; Iterative process②: After inputting the mixed data of source domain and target domain, in order to improve the accuracy of identifying the sample domains of source domain and target domain, the parameters are adjusted , minimize the domain classification loss as much as possible ; Iterative process ③: To increase the interval between features of samples of different categories, that is, to increase separability, by adjusting the parameters , minimize the domain classification loss as much as possible ; Iterative process ④: To reduce the feature differences between similar samples in the source domain and the target domain, that is, to reduce the intra-class distance, the parameters , maximize the domain classification loss ; Adjustment parameters Make the domain classification loss Increasing will make it easier for the subdomain discriminator to determine whether the data comes from the source domain or the target domain, and adjusting the parameters To reduce the gap between domains, adjust Value and Values ​​continue to fight, when adjusting Make the subdomain classification loss Reaching the maximum value, if the subdomain discriminator When it is impossible to determine whether the radar data sequence comes from the source domain or the target domain, the features extracted by the feature extractor are common features of the target domain and source domain radars, and the feature classifier can make more accurate classification predictions for the unlabeled target domain radar data.

4. The radar operating mode recognition method based on multiple adversarial domains and metric learning according to claim 3 is characterized by: The grey relational function is shown in the following formula (6): (6) In formula (6) Indicates the number of radar sequence division units, The physical feature displacement difference that represents the similarity between the source domain and the target domain is expressed as follows: (7) In formula (6) The physical feature velocity difference that represents the similarity between the source domain and the target domain is expressed as follows: (8) In formula (6) The physical feature acceleration difference that represents the similarity between the source domain and the target domain is expressed as follows: (9)。

Images