Multifunctional Radar Working Mode Recognition Method Based on the Knowledge Graph ROTATE Model

Abstract

This application is applicable to the field of radar signal processing, and provides a multifunctional radar working mode recognition method based on a knowledge graph ROTATE model, including: randomly generating a radar sample data set according to the preset parameter ranges and parameter modulation methods of each working mode of the radar; preprocessing the radar sample data set to form basic unit triples for constructing a knowledge graph; constructing a knowledge graph that describes the relationship between the radar working mode, parameter values, and parameter modulation types based on the basic unit triples; mapping the entities and relationships in the knowledge graph to a low-dimensional continuous vector space for representation, and using the scoring function defined by the ROTATE model algorithm to calculate the confidence of each basic unit triple; training the model by maximizing the confidence of the basic unit triples, and using the trained model to perform working mode recognition on the test set. This application can reduce the recognition time while ensuring a high recognition rate, and has good generalization performance.

Description

Technical Field

[0001] The present invention belongs to the field of radar signal processing, and particularly relates to a multifunctional radar operating mode recognition method based on a knowledge graph ROTATE model. Background Art

[0002] Radar operating mode recognition is an important research direction in modern electronic warfare and signal intelligence. With the continuous progress of radar technology, the complexity and diversity of radar systems have increased significantly, making it increasingly important to effectively recognize the operating modes of radar signals. However, multifunctional radar operating mode recognition currently faces some challenges. The radar operating mode may change dynamically over time, and signal parameters (such as frequency, pulse width, repetition period, etc.) will also be adjusted accordingly, increasing the difficulty of recognition and making algorithm design more complex. When traditional knowledge graph-based signal processing methods face complex signal data sequences, they need to embed high-dimensional vectors, and have a large amount of computation and poor algorithm convergence. Traditional methods mainly include traditional machine learning methods, clustering analysis methods, etc. Zhai Longjun, Dan Bo, Song Weijian, Gao Shan. Multifunctional Radar Operating Mode Recognition Based on Clustering Analysis Method [J]. Ship Electronic Engineering, 2022, 42 (04): 86-88. introduced the K-means clustering algorithm to realize the recognition of four radar operating modes, but its number of iterations reached 100,000 times, with poor algorithm convergence, and all four operating modes in the experiment were fixed repetition frequency signals, without realizing the recognition of complex signal data sequences. Ren Wenbo, Wang Chao, Shi Qingzhan, etc. A Radar Operating Mode Recognition System Based on Neural Network [A] Proceedings of the 2021 National Microwave and Millimeter Wave Conference (Volume II) [C]. Chinese Institute of Electronics, Microwave Society of Chinese Institute of Electronics, 2021: 3. proposed a neural network-based operating mode recognition system, which automatically recognizes different operating modes by extracting multi-level radar feature parameters, with a recognition rate of 90.6%. However, this research did not establish a complete and reliable radar database and radar knowledge base, and could not recognize new and unknown radars. To sum up, the current radar operating mode recognition methods have problems such as low recognition accuracy, poor algorithm convergence, and excessive requirements for the dimensions of knowledge graph entity and relationship embedding vectors. Summary of the Invention

[0003] The present application provides a multifunctional radar operating mode recognition method, device, terminal device, and storage medium based on a knowledge graph ROTATE model, which can solve the problems of low recognition accuracy and excessive requirements for the dimensions of knowledge graph entity and relationship embedding vectors in the existing radar operating mode recognition methods.

[0004] In a first aspect, the present application provides a multifunctional radar working mode recognition method based on a knowledge graph ROTATE model, including the following steps: randomly generate a radar sample data set according to the preset parameter ranges and parameter modulation methods of each radar working mode;

[0005] Preprocess the radar sample data set to form the basic unit triples for constructing the knowledge graph;

[0006] Construct a knowledge graph based on the basic unit triples to describe the relationship between the radar working mode, parameter values, and parameter modulation types;

[0007] Map the entities and relationships in the knowledge graph to a low-dimensional continuous vector space for representation, and use the scoring function defined by the ROTATE algorithm to calculate the confidence of each basic unit triple;

[0008] Maximize the confidence of the basic unit triples to train the model, and use the trained model to recognize the working mode of the test set;

[0009] During the process of training the model, record the scoring function of the negative example triples and use it as the decision threshold for the working mode of unknown models to recognize the working mode of unknown models.

[0010] In a possible implementation manner of the first aspect, the above-mentioned randomly generating a radar sample data set according to the preset parameter ranges and parameter modulation methods of each radar working mode includes:

[0011] Set the number of types of radar working modes, the parameter ranges under each working mode, and the parameter modulation types;

[0012] Randomly generate a sample data set according to the preset content. Each sample data includes a sample number, a parameter modulation type, parameter sequence data, and a radar working mode. Among them, the parameter sequence data represents the parameter change rules in three dimensions of pulse width, signal carrier frequency, and pulse repetition period in the radar pulse descriptor under each working mode as a sequence of a certain length, and this change rule is the parameter modulation type.

[0013] Optionally, in another possible implementation manner of the first aspect, the above-mentioned preprocessing the radar sample data set to form the basic unit triples for constructing the knowledge graph includes:

[0014] Model the generated sample data set. According to the correspondence between the samples, parameter values, sample-to-parameter modulation types, and sample-to-radar operating modes, construct 7 triples for each sample, namely: (sample number, pulse width, sequence value), (sample number, pulse width modulation, type), (sample number, signal carrier frequency, sequence value), (sample number, signal carrier frequency modulation, type), (sample number, pulse repetition period, sequence value), (sample number, pulse repetition period modulation, type), (sample number, operating mode, operating mode type).

[0015] Optionally, in another possible implementation manner of the first aspect, constructing the knowledge graph that describes the relationship between the radar operating mode and parameter values and parameter modulation types based on the basic triples includes:

[0016] Construct the sample number as the head entity;

[0017] Construct the pulse width sequence, signal carrier frequency sequence, and pulse repetition period sequence as the first tail entity, represented as an encoded column vector of length N;

[0018] Construct the pulse width modulation type, signal carrier frequency modulation type, pulse repetition period modulation type, and operating mode type as the second tail entity, represented as text;

[0019] Construct the pulse width, pulse width modulation, signal carrier frequency, signal carrier frequency modulation, pulse repetition period, pulse repetition period modulation, and operating mode as relationships, represented as text;

[0020] Store the radar operating mode knowledge graph constructed by all triples in the neo4j database.

[0021] Optionally, in yet another possible implementation manner of the first aspect, mapping the entities and relationships of the knowledge graph into a low-dimensional continuous vector space for representation, and using the scoring function defined by the ROTATE algorithm to calculate the confidence of each basic unit triple includes:

[0022] Map the entities and relationships of the knowledge graph into the complex vector space, and the dimension of the mapped vector is , and define each relationship as a rotation from the head entity to the tail entity, with representing the head entity vector, relationship vector, and tail entity vector respectively;

[0023] For each triple there is , and the modulus length where ∘ represents the Hadamard product (element-wise product). Specifically, for each element in the embedding, there is , and take the modulus of each element of (i.e., is constrained to , and by this means is in the form of , ( being the imaginary unit), which corresponds to a counterclockwise rotation of radians about the origin of the complex plane and only affects the phase vector space of the entities embedded in the complex number. According to the above definition, for each triple , the distance function is defined as:

[0024]

[0025] Negative sampling is introduced, and the self-adversarial negative sampling technique is adopted to generate negative samples through the current entity and relation embeddings, which follows the following distribution:

[0026]

[0027] where represents the probability, is the j-th negative example triple, is the i-th positive example triple, is the sampling probability, and the probability is used as the weight of the negative samples;

[0028] After introducing self-adversarial negative sampling, the scoring function is obtained, and the expression is:

[0029]

[0030] where is the margin parameter, is the sigmoid function;

[0031] The confidence of the basic unit triple is defined to be inversely proportional to the scoring function. The larger the scoring function, the smaller the confidence, and the smaller the scoring function, the larger the confidence. The value range of the confidence is (0, 1).

[0032] Optionally, in another possible implementation manner of the first aspect, the above-mentioned training of the model by maximizing the confidence of the basic unit triple and the working mode recognition of the trained model for the test set include:

[0033] Parameter setting: Set the dimensions of the head entity, tail entity, and relation embedding vectors in the knowledge graph during training, the training learning rate, the number of negative samples, the self-adversarial negative sampling parameter, the number of training iterations, and the margin parameter; divide multiple basic unit triples into a training set and a test set;

[0034] For each positive example triple in the training set, negative example triples are generated by the current entity and relation embeddings. The specific method is as follows: Take an entity from the entity dataset and randomly select to replace the head entity or the tail entity in the current positive example triple to generate a new negative example triple;

[0035] Train the model using the training set: According to the scoring function formula, use the gradient descent algorithm to optimize the model, update the parameters, and optimize the scoring function to make the positive example triples obtain lower score values;

[0036] Use the trained model to perform working mode recognition on the test set, calculate the scoring function of the test sample and the embedding vectors corresponding to each radar working mode entity, and take the working mode corresponding to the minimum scoring function as the final recognition result.

[0037] Optionally, in another possible implementation manner of the first aspect, recording the scoring function of the negative example triples during the process of training the model and using it as the decision threshold for the working mode of unknown models to identify the working mode of unknown models includes:

[0038] Set the decision threshold for the working mode of unknown models according to the scoring function value of the negative example triples, compare the scoring function value of any appearance sample with the decision threshold. If it is less than the threshold, the sample is identified as a known working mode; otherwise, it is identified as the working mode of an unknown model.

[0039] In a second aspect, an embodiment of the present application provides a multifunctional radar working mode recognition system based on a knowledge graph ROTATE model, including:

[0040] A radar signal data generation module, configured to randomly generate a radar sample dataset according to the preset parameter ranges and parameter modulation methods of each radar working mode;

[0041] A data preprocessing module, configured to preprocess the radar sample data to form the basic unit triples for constructing the knowledge graph;

[0042] A knowledge graph construction module, configured to construct a knowledge graph describing the relationship between the radar working mode, parameter values, and parameter modulation types based on the basic unit triples;

[0043] A model training module, configured to map the entities and relationships of the knowledge graph into a low-dimensional continuous vector space for representation, and use the scoring function defined by the ROTATE model to calculate the confidence of each basic unit triple; maximize the confidence of the basic unit triples to train the model;

[0044] An identification module, configured to perform working mode recognition on the test set using the trained model.

[0045] In a third aspect, an embodiment of the present application provides a computer program product. When the computer program product runs on a terminal device, it causes the terminal device to execute the multi-functional radar working mode recognition method based on the knowledge graph ROTATE model according to any one of the above first aspects.

[0046] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. The computer program, when executed by a processor, implements the multi-functional radar working mode recognition method based on the knowledge graph ROTATE model as described above.

[0047] The beneficial effects of the present application are as follows: The present application proposes a multi-functional radar working mode recognition method based on the knowledge graph ROTATE model. First, preset the preset parameter ranges of each working mode of the radar and the parameter modulation method to randomly generate a radar sample data set, and preprocess the data to form the basic unit triples for constructing the knowledge graph; secondly, construct a knowledge graph describing the relationship between the radar working mode, parameter values, and parameter modulation types; map the entities and relationships of the radar working mode knowledge graph into a low-dimensional continuous vector space for representation, and define a scoring function to calculate the confidence of a triple; finally, maximize the confidence to train the model, use the trained model to recognize the working mode of the test set, and record the scoring function of the negative example triples during the process of training the model and use it as a decision threshold to recognize the working mode of unknown models. The recognition method uses the ROTATE algorithm in knowledge graph embedding learning, and obtains a recognition model through training. The recognition model can achieve good recognition results and improve the recognition accuracy. The present application innovatively combines knowledge graph learning with radar working mode recognition, can recognize signals in multiple frequency bands and multiple working modes, can process radar signal data sequences with a relatively simple algorithm and achieve a high correct recognition rate of radar working modes, has a fast recognition speed, good algorithm convergence, and can discriminate the working mode of unknown models. Description of the Drawings

[0048] Figure 1 is a schematic flowchart of a multi-functional radar working mode recognition method based on the knowledge graph ROTATE model provided by an embodiment of the present application;

[0049] Figure 2 is a graph of the recognition accuracy of the radar working mode under different embedding vectors provided by an embodiment of the present application and a comparison graph with the model TransE;

[0050] Figure 3 is a schematic diagram of the recognition accuracy of the ROTATE model provided by an embodiment of the present application for the working mode of unknown models under different iteration times;

[0051] Figure 4It is a schematic structural diagram of a multi-functional radar working mode recognition system based on the knowledge graph ROTATE model provided by an embodiment of the present application. Detailed implementation manners

[0052] In the following description, specific details such as specific system structures and technologies are proposed for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present application. However, those skilled in the art should clearly understand that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from interfering with the description of the present application.

[0053] It should be understood that when used in the specification and appended claims of the present application, the term "comprising" indicates the presence of the described features, wholes, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0054] It should also be understood that the term "and / or" used in the specification and appended claims of the present application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0055] As used in the specification and appended claims of the present application, the term "if" can be interpreted as "when", "once", "in response to determining", or "in response to detecting" depending on the context. Similarly, the phrase "if determined" or "if detecting [the described condition or event]" can be interpreted as meaning "once determined", "in response to determining", "once detecting [the described condition or event]", or "in response to detecting [the described condition or event]" depending on the context.

[0056] In addition, in the description of the specification and appended claims of the present application, the terms "first", "second", "third", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0057] The reference to "one embodiment" or "some embodiments" etc. described in the specification of the present application means that a specific feature, structure, or characteristic described in connection with the embodiment is included in one or more embodiments of the present application. Thus, the statements "in one embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments", etc. that appear in different places in this specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "comprising", "including", "having", and their variants all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0058] Figure 1 Shows the schematic flow chart of a multi-functional radar working mode recognition method based on the knowledge graph ROTATE model provided by an embodiment of the present application.

[0059] Step 101, randomly generate a radar sample data set according to the preset parameter ranges and parameter modulation methods of each radar working mode.

[0060] Further, in the embodiment of the present application, the above step 101 includes:

[0061] Step 1011, preset the number of types of radar working modes, the parameter ranges under each working mode, and the parameter modulation types;

[0062] Step 1012, randomly generate a sample data set according to the preset content. Each sample data includes a sample number, a parameter modulation type, parameter sequence data, and a radar working mode. Among them, the parameter sequence data characterizes the parameter change rules in three dimensions of pulse width, signal carrier frequency, and pulse repetition period in the radar pulse descriptor under each working mode as a sequence of a certain length, and this sequence change rule is the parameter modulation type.

[0063] Preferably, the specific implementation manner of the above step 101 can be described with reference to the following embodiments:

[0064] S1, preset the parameter ranges and parameter modulation types of each radar working mode, set the number of known working mode types to 10, numbered 1 - 10, and the number of unknown model working mode types to 5, numbered 11 - 15;

[0065] For mode 1, the parameter RF range is 4000 - 4500 MHz, fixed carrier frequency, the parameter PW range is 1 - 4 us, fixed pulse width, the parameter PRI range is 30 - 35 us, fixed pulse repetition frequency;

[0066] For mode 2, the parameter RF range is 4000 - 4500 MHz, carrier frequency hopping, the parameter PW range is 1 - 4 us, pulse width jitter, the parameter PRI range is 30 - 35 us, pulse repetition frequency stagger;

[0067] For mode 3, the parameter RF range is 4000 - 4500 MHz, carrier frequency hopping, the parameter PW range is 1 - 4 us, pulse width jitter, the parameter PRI range is 30 - 35 us, pulse repetition frequency sliding;

[0068] For mode 4, the parameter RF range is 4000 - 4500 MHz, carrier frequency inter-pulse agility, the parameter PW range is 1 - 4 us, pulse width jitter, the parameter PRI range is 30 - 35 us, pulse repetition frequency sliding;

[0069] Mode 5 parameters: RF range 4500 - 5000 MHz, carrier frequency pulse group agility, parameter PW range 1 - 4 us, pulse width jitter, parameter PRI range 35 - 40 us, pulse repetition frequency (PRF) slide;

[0070] Mode 6 parameters: RF range 4500 - 5000 MHz, carrier frequency inter-pulse agility, parameter PW range 1 - 4 us, pulse width jitter, parameter PRI range 35 - 40 us, pulse repetition frequency stagger;

[0071] Mode 7 parameters: RF range 5000 - 5500 MHz, carrier frequency inter-pulse agility, parameter PW range 5 - 8 us, pulse width jitter, parameter PRI range 40 - 45 us, pulse repetition frequency stagger;

[0072] Mode 8 parameters: RF range 5000 - 5500 MHz, carrier frequency hopping, parameter PW range 5 - 8 us, pulse width jitter, parameter PRI range 40 - 45 us, pulse repetition frequency jitter;

[0073] Mode 9 parameters: RF range 5000 - 5500 MHz, carrier frequency pulse group agility, parameter PW range 5 - 8 us, pulse width jitter, parameter PRI range 40 - 45 us, pulse repetition frequency slide;

[0074] Mode 10 parameters: RF range 5000 - 5500 MHz, carrier frequency pulse group agility, parameter PW range 5 - 8 us, pulse width fixed, parameter PRI range 40 - 45 us, pulse repetition frequency stagger;

[0075] Mode 11 parameters: RF range 4000 - 4500 MHz, carrier frequency inter-pulse agility, parameter PW range 1 - 4 us, pulse width fixed, parameter PRI range 30 - 35 us, pulse repetition frequency stagger;

[0076] Mode 12 parameters: RF range 4500 - 5000 MHz, carrier frequency fixed, parameter PW range 1 - 4 us, pulse width fixed, parameter PRI range 35 - 40 us, pulse repetition frequency jitter;

[0077] Mode 13 parameters: RF range 4500 - 5000 MHz, carrier frequency pulse group agility, parameter PW range 1 - 4 us, pulse width fixed, parameter PRI range 35 - 40 us, pulse repetition frequency stagger;

[0078] Mode 14 parameters: RF range 4500 - 5000 MHz, carrier frequency hopping, parameter PW range 5 - 8 us, pulse width fixed, parameter PRI range 35 - 40 us, pulse repetition frequency slide;

[0079] Mode 15 parameters: RF range 5000 - 5500 MHz, carrier frequency inter-pulse agility, parameter PW range 5 - 8 us, pulse width fixed, parameter PRI range 40 - 45 us, pulse repetition frequency stagger;

[0080] S2, randomly generate sample data according to the above settings, where each mode 1 to 10 generates 2000 samples, and each mode 11 to 15 generates 100 samples. Each sample randomly generates a radar characteristic parameter sequence, where the parameter sequence is a sequence of a certain length that characterizes the parameter change law in the three dimensions of pulse width, signal carrier frequency, and pulse repetition period in the radar pulse description word under each working mode. The change law is the parameter modulation type. Generate radar characteristic parameter sequences according to the law, where each parameter sequence is 24 in length, and each parameter sequence is generated as follows:

[0081] S21, randomly generates a pulse width sequence. There are two types of pulse width modulation:

[0082] Fixed pulse width: ;

[0083] Pulse Width Jitter: .

[0084] in express The pulse width value at the moment, represents a fixed pulse width value, k represents the number of pulse width types, Indicates the kth type pulse width value.

[0085] S22, randomly generates a signal carrier frequency sequence. There are four types of modulation of the signal carrier frequency:

[0086] Fixed carrier frequency: ;

[0087] Carrier frequency hopping: ;

[0088] Carrier frequency pulse-to-pulse agility: ;

[0089] Carrier frequency pulse group agility: ;

[0090] in, express The signal carrier frequency value at time represents a fixed carrier frequency value, k represents the number of carrier frequency types, represents the carrier frequency value of the k-th signal, is the frequency agility range, is in the interval A uniformly distributed random number, is the number of pulses for each frequency.

[0091] S23, randomly generates a pulse repetition period sequence, and there are four modulation types of pulse repetition period:

[0092] PRF fixed: ;

[0093] PRF jitter: ;

[0094] PRF stagger: m-staggered radar: , ;

[0095] PRF sliding: ;

[0096] Among them, represents the pulse repetition period value at time represents the fixed pulse repetition period value is PRI the initial value of is a uniformly distributed random number within, controlled by the maximum jitter amount, and the maximum jitter amount , generally taking the value of , is the increment of the value, changing according to the sliding rule is PRI the sliding range of represents the modulo operation.

[0097] Step 102: Preprocess the radar sample data set to form the basic unit triples for constructing the knowledge graph.

[0098] Furthermore, in the embodiment of the present application, the above step 102 includes:

[0099] Model the generated sample data set, and construct 7 triples for each sample according to the corresponding relationships between the sample and the parameter value, the sample and the parameter modulation type, and the sample and the radar operating mode, namely {sample number, pulse width, sequence value}, {sample number, pulse width modulation, type}, {sample number, signal carrier frequency, sequence value}, {sample number, signal carrier frequency modulation, type}, {sample number, pulse repetition period, sequence value}, {sample number, pulse repetition period modulation, type}, {sample number, operating mode, operating mode type}.

[0100] Step 103: Construct a knowledge graph describing the relationship between the radar operating mode and the parameter value and the parameter modulation type based on the basic unit triples.

[0101] Furthermore, in the embodiment of the present application, the above step 103 includes:

[0102] Step 1031: Construct the sample number as the head entity;

[0103] Step 1032: Construct the pulse width sequence, signal carrier frequency sequence, and pulse repetition period sequence as the first tail entity, represented as a coded column vector of length N;

[0104] Step 1033: Construct the pulse width modulation type, signal carrier frequency modulation type, pulse repetition period modulation type, and working mode type as the second tail entity, represented as text;

[0105] Construct the relationships among the pulse width, pulse width modulation, signal carrier frequency, signal carrier frequency modulation, pulse repetition period, pulse repetition period modulation, and working mode, represented as text;

[0106] Step 1034: Integrate all triples to construct a radar working mode knowledge graph and store it in the neo4j database.

[0107] Step 104: Map the entities and relationships of the knowledge graph into a low-dimensional continuous vector space for representation, and use the scoring function defined by the ROTATE model algorithm to calculate the confidence of each basic unit triple.

[0108] Further, in the embodiment of the present application, the above Step 104 includes:

[0109] Step 1041: Map the entities and relationships of the knowledge graph into a complex vector space, and define each relationship as a rotation from the head entity to the tail entity, so as to represent the head entity vector, relationship vector, and tail entity vector respectively;

[0110] Step 1042: Introduce negative sampling, adopt the self-adversarial negative sampling technique, generate negative samples through the current entity and relationship embeddings, and follow the following distribution:

[0111]

[0112] where represents probability, is the j-th negative example triple, is the i-th positive example triple, is the sampling probability, and take the probability as the weight of the negative sample, represents the distance function, ;

[0113] Step 1043: After introducing self-adversarial negative sampling, obtain the scoring function, and the expression is:

[0114]

[0115] where, is a boundary parameter, is the sigmoid function;

[0116] Step 1044, define that the confidence of the basic unit triple is inversely proportional to the scoring function.

[0117] Preferably, in one embodiment, map the entities and relationships of the radar working mode knowledge graph into a low-dimensional continuous vector space for representation, map the entities and relationships into a complex vector space, and the lengths of the mapped vectors are 20, 40, 60, 80, 100, 200, 400, 600, 800 respectively, and define each relationship as a rotation from the head entity to the tail entity. Let represent the head entity vector, the relationship vector, and the tail entity vector respectively.

[0118] Use the ROTATE algorithm to define a scoring function to calculate the confidence of a triple.

[0119] For each triple there is , the modulus where ∘ represents the Hadamard product (element-wise product). Specifically, for each element in the embedding, there is . Let the modulus of each element of (i.e., ) be constrained to , and in this way, is in the form of ( is the imaginary unit), which corresponds to a counterclockwise rotation of radians around the origin of the complex plane and only affects the phase vector space of the entity embeddings in the complex numbers. According to the above definition, for each triple , define the distance function as:

[0120]

[0121] Introduce negative sampling and adopt the self-adversarial negative sampling technique to generate negative samples through the current entity and relationship embeddings. The expression is:

[0122]

[0123] where, is the sampling probability, and take the probability as the weight of the negative samples. Therefore, use the self-adversarial negative sampling scoring function to calculate the confidence of a triple. The expression is:

[0124]

[0125] where, is the boundary parameter, is the sigmoid function, ( ) is the i-th negative example triple.

[0126] It is defined that the confidence is inversely proportional to the score function. The larger the score function, the smaller the confidence; the smaller the score function, the larger the confidence, and the value range of the confidence is (0, 1).

[0127] Step 105: Maximize the confidence of the basic unit triples to train the model, and use the trained model to perform working mode recognition on the test set.

[0128] Furthermore, in the embodiment of the present application, the above step 105 includes:

[0129] Step 1051: Parameter setting: Set the dimensions of the head entity, tail entity, and relationship embedding vectors in the knowledge graph during training, the training learning rate, the number of negative samplings, the adversarial negative sampling parameter, the number of training iterations, and the boundary parameter; divide multiple basic unit triples into a training set and a test set;

[0130] Step 1052: For each positive example triple in the training set, generate a negative example triple by generating the current entity and relationship embedding. The specific method is as follows: Randomly select an entity from the entity dataset to replace the head entity or tail entity in the current positive example triple to generate a new negative example triple;

[0131] Step 1053: Use the training set to train the model: Optimize the score function of the model using the gradient descent algorithm according to the score function formula;

[0132] Step 1054: Use the trained model to perform working mode recognition on the test set, calculate the score function of the test sample and the embedding vectors corresponding to each radar working mode entity, and take the working mode corresponding to the minimum score function as the final recognition result.

[0133] It should be noted that the entity dataset is derived from the sample data generation process before constructing the basic unit triples. Specifically, by presetting the parameter range and modulation type of the radar working mode, radar sample data is randomly generated. These data include the numbers of each sample, the sequence values of various parameters, and the corresponding modulation types. During the process of constructing the knowledge graph, this information is used as different entities respectively: the sample number constitutes the head entity; the parameter sequence (such as the numerical sequence of pulse width, signal carrier frequency, and pulse repetition period) constitutes part of the tail entity; the modulation type (such as text information like "fixed", "jitter", etc.): also serves as the tail entity. The set of entities extracted from these sample data constitutes the entity dataset, and subsequent entities are randomly selected from this entity dataset for replacement when generating negative examples.

[0134] Preferably, in one embodiment, the parameter settings are as follows: in the training, the dimensions of the head entity, tail entity, and relationship embedding vectors are set to 20, 40, 60, 80, 100, 200, 400, 600, 800 respectively; the training learning rate is 0.0008; the number of negative samples is 512; the adversarial negative sampling parameter is 2.0; the number of training steps is 20000; the margin parameter is 10.0; the training set and the test set are divided, and the ratio is 0.95:0.05.

[0135] For the training set, for each positive example triple, through the current entity and relationship, an entity is randomly selected from the entity dataset to replace the head entity or tail entity in the current triple to generate a new negative example triple.

[0136] Use the training set to train the model. According to the scoring function formula, use the gradient descent algorithm to optimize the model, update the parameters, and optimize the scoring function to make the positive example triple obtain a lower score value.

[0137] Calculate the scoring function of the test sample and the embedding vector corresponding to each radar operating mode entity, and take the operating mode corresponding to the minimum scoring function as the final recognition result. Change the vector dimension and repeat the training and recognition process. The recognition accuracy rates under different embedding vector dimensions are as Figure 2 shown. It can be seen that in the low-dimensional case where the embedding vector is 20 - 60 dimensions, the correct recognition rate of the present invention is more than 10% higher than that of the TransE model; in the high-dimensional case where the embedding vector is 200 - 1000 dimensions, the correct recognition rate of the present invention is more than 1% higher than that of the TransE model.

[0138] Step 106, record the scoring function of the negative example triple during the process of training the model and use it as the decision threshold for the operating mode of the unknown model to identify the operating mode of the unknown model.

[0139] Further, in the embodiment of the present application, the above step 106 includes:

[0140] Set the decision threshold for the operating mode of the unknown model according to the scoring function value of the negative example triple, compare the scoring function value of any presented sample with the decision threshold. If it is less than the threshold, the sample is identified as a known operating mode; otherwise, it is identified as the operating mode of the unknown model.

[0141] Preferably, in one embodiment, during the process of training the model, the scoring function of negative example triples is recorded and used as the decision threshold for unknown model working mode recognition. During the training process, the scoring function value of positive example triples is relatively small, while that of negative example triples is relatively large. Triples constructed from the working modes of unknown model working mode samples and other known samples in the knowledge graph should all be incorrect and belong to negative example triples. According to the scoring function values of negative example triples, the corresponding scoring function decision thresholds are set. A total of 1000 scoring function values of negative example triples are counted, and the 1st, 10th, 100th, and 1000th values sorted from largest to smallest in the scoring function values are used as decision thresholds, that is, the decision thresholds are respectively: 7.674 (false alarm rate PF = 10 -3 ), 7.635 (false alarm rate PF = 10 -2 ), 7.544 (false alarm rate PF = 10 -1 ), 6.515 (false alarm rate PF = 1). Compare the smallest scoring function of each sample with the decision threshold. If it is less than the threshold, the sample model is a known working mode; otherwise, it is an unknown model working mode. Thus, the recognition of unknown model working modes can be realized. And the recognition rates of unknown model working modes under each threshold are obtained. The recognition accuracies under different thresholds when the model iteration times are 300, 400, and 500 are as Figure 3 shown.

[0142] The multi-functional radar working mode recognition method based on the knowledge graph ROTATE model provided by this application first presets the parameter range of each working mode of the radar and randomly generates a radar sample data set with parameter modulation methods, and preprocesses the data to form the basic unit triples for constructing the knowledge graph; secondly, constructs a knowledge graph describing the relationship between radar working modes, parameter values, and parameter modulation types; maps the entities and relationships of the radar working mode knowledge graph into a low-dimensional continuous vector space for representation, and defines a scoring function to calculate the confidence of a triple; finally, maximizes the confidence to train the model, uses the trained model to recognize the working mode of the test set, and records the scoring function of negative example triples during the process of training the model and uses it as the decision threshold to recognize unknown model working modes. The recognition method uses the ROTATE algorithm in knowledge graph embedding learning, and through training, an identifier model can be obtained. The identifier can achieve good recognition effects and improve the recognition accuracy. This application innovatively combines knowledge graph learning with radar working mode recognition, can recognize signals in multiple frequency bands and multiple working modes, can process radar signal data sequences with relatively simple algorithms and achieve a relatively high correct recognition rate of radar working modes, has a relatively fast recognition speed, good algorithm convergence, and can discriminate unknown model working modes.

[0143] It should be understood that the sequence numbers of the steps in the above embodiments do not indicate the order of execution, and the order of execution of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0144] A multi-functional radar working mode recognition method based on the knowledge graph ROTATE model corresponding to the above embodiment. Figure 4 The structural block diagram of the multi-functional radar working mode recognition system provided by the embodiment of the present application is shown. For the convenience of description, only the parts related to the embodiment of the present application are shown.

[0145] Refer to Figure 4 , the system 400 includes:

[0146] A radar signal data generation module 401, configured to randomly generate a radar sample data set according to the preset parameter range and parameter modulation method of each working mode of the radar.

[0147] A data preprocessing module 402, configured to preprocess the radar sample data to form a basic unit triple for constructing a knowledge graph.

[0148] A knowledge graph construction module 403, configured to construct a knowledge graph describing the relationship between the radar working mode, parameter values, and parameter modulation types based on the basic unit triples.

[0149] A model training module 404, configured to map the entities and relationships of the knowledge graph into a low-dimensional continuous vector space for characterization, and use the scoring function defined by the ROTATE model to calculate the confidence of each basic unit triple, and maximize the confidence of the basic unit triple to train the model.

[0150] An identification module 405, configured to perform working mode identification on the test set using the trained model.

[0151] In actual use, the multi-functional radar working mode recognition system based on the knowledge graph ROTATE model provided by the embodiment of the present application can be configured in any terminal device to execute the foregoing multi-functional radar working mode recognition method based on the knowledge graph ROTATE model.

[0152] The multi-functional radar operating mode recognition system based on the knowledge graph ROTATE model provided by this application first presets the preset parameter ranges and parameter modulation methods of each radar operating mode to randomly generate a radar sample data set, and preprocesses the data to form the basic unit triples for constructing the knowledge graph. Secondly, a knowledge graph describing the relationship between the radar operating mode, parameter values, and parameter modulation types is constructed. The entities and relationships of the radar operating mode knowledge graph are mapped into a low-dimensional continuous vector space for representation, and a scoring function is defined to calculate the confidence of a triple. Finally, the confidence is maximized to train the model. The trained model is used to recognize the operating mode of the test set, and the scoring function of the negative example triples is recorded during the process of training the model and used as a decision threshold to recognize the operating modes of unknown models. The recognition method uses the ROTATE algorithm in knowledge graph embedding learning to obtain an identifier model through training. The identifier can achieve good recognition results and improve the recognition accuracy. This application innovatively combines knowledge graph learning with radar operating mode recognition, can recognize signals in multiple frequency bands and various operating modes, can process radar signal data sequences with relatively simple algorithms and achieve a high correct recognition rate of radar operating modes, has a fast recognition speed, good algorithm convergence, and can discriminate the operating modes of unknown models.

[0153] It should be noted that, regarding the information interaction, execution process, etc. between the above systems / units, since they are based on the same concept as the method embodiments of this application, for their specific functions and the technical effects brought about, reference can be specifically made to the method embodiment part, and details will not be elaborated here.

[0154] Those skilled in the art can clearly understand that, for the convenience and conciseness of description, only the above division of each functional unit and module is used for illustration. In actual applications, the above functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for the convenience of mutual distinction and do not limit the protection scope of this application. The specific working processes of the units and modules in the above system can refer to the corresponding processes in the foregoing method embodiments, and details will not be elaborated here.

[0155] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A multi-functional radar operating mode recognition method based on the knowledge graph ROTATE model, characterized in that Including the following steps: Randomly generate a radar sample data set according to the preset parameter ranges and parameter modulation methods for each radar operating mode; Preprocess the radar sample data set to form the basic unit triples for constructing the knowledge graph; Construct a knowledge graph based on the basic unit triples to describe the relationship between the radar operating mode, parameter values, and parameter modulation types; Map the entities and relationships in the knowledge graph into a low-dimensional continuous vector space for representation, and use the scoring function defined by the ROTATE model algorithm to calculate the confidence of each basic unit triple. Define that the confidence of the basic unit triple is inversely proportional to the scoring function; Maximize the confidence of the basic unit triples to train the model, and use the trained model to identify the operating mode of the test set; During the process of training the model, record the scoring function of the negative example triples and use it as the decision threshold for the operating mode of unknown models to identify the operating mode of unknown models.

2. The multi-functional radar working mode recognition method based on the knowledge graph ROTATE model according to claim 1, characterized in that, The randomly generating a radar sample data set according to the preset parameter ranges and parameter modulation methods for each radar operating mode includes: Preset the number of types of radar operating modes, the parameter ranges under each operating mode, and the parameter modulation types in advance; Randomly generate a sample data set according to the preset content. Each sample data includes a sample number, a parameter modulation type, parameter sequence data, and a radar operating mode. Among them, the parameter sequence data represents the parameter change rules in three dimensions of pulse width, signal carrier frequency, and pulse repetition period in the radar pulse descriptor under each operating mode as a sequence of a certain length, and this sequence change rule is the parameter modulation type.

3. According to the multi-functional radar operating mode recognition method based on the knowledge graph ROTATE model described in claim 2, the preprocessing of the radar sample data set to form the basic unit triples for constructing the knowledge graph includes: Model the generated sample data set. According to the corresponding relationships from the sample to the parameter value, from the sample to the parameter modulation type, and from the sample to the radar operating mode, construct 7 triples for each sample, namely: (sample number, pulse width, sequence value), (sample number, pulse width modulation, type), (sample number, signal carrier frequency, sequence value), (sample number, signal carrier frequency modulation, type), (sample number, pulse repetition period, sequence value), (sample number, pulse repetition period modulation, type), (sample number, operating mode, operating mode type).

4. The multifunctional radar working mode recognition method based on the knowledge graph ROTATE model according to claim 3, wherein The constructing a knowledge graph based on the basic triples to describe the relationship between the radar operating mode, parameter values, and parameter modulation types includes: Construct the sample number as the head entity; Construct the pulse width sequence, signal carrier frequency sequence, and pulse repetition period sequence as the first tail entity, represented as an encoded column vector of length N; Construct the pulse width modulation type, signal carrier frequency modulation type, pulse repetition period modulation type, and operating mode type as the second tail entity, represented as text; Construct pulse width, pulse width modulation, signal carrier frequency, signal carrier frequency modulation, pulse repetition period, pulse repetition period modulation, and operating mode as relationships, represented as text; Store the radar operating mode knowledge graph constructed by all triples in the neo4j database.

5. The multifunctional radar operating mode recognition method based on the knowledge graph ROTATE model according to claim 4, wherein Mapping the entities and relationships of the knowledge graph into a low-dimensional continuous vector space for representation, and using the scoring function defined by the algorithm of the ROTATE model to calculate the confidence of each basic unit triple, the steps include: Mapping the entities and relationships of the knowledge graph into a complex vector space, and defining each relationship as a rotation from the head entity to the tail entity. Let h, r, and t represent the head entity vector, relationship vector, and tail entity vector respectively; Introducing negative sampling, adopting the self-adversarial negative sampling technique, generating negative samples through the current entity and relationship embeddings, following the following distribution: where p represents probability, (h' j , r, t' j ) is the j-th negative example triple, (h i , r i , t i ) is the i-th positive example triple, α is the sampling probability, taking the probability α as the weight of the negative sample, f r represents the distance function, After introducing self-adversarial negative sampling, the scoring function is obtained, and the expression is: where γ is the margin parameter and σ is the sigmoid function; Defining that the confidence of the basic unit triple is inversely proportional to the scoring function.

6. The multi-functional radar operating mode recognition method based on the knowledge graph ROTATE model according to claim 5, characterized in that, Maximizing the confidence of the basic unit triple to train the model, and using the trained model to perform working mode recognition on the test set, including: Parameter setting: Setting the dimensions of the head entity, tail entity, and relationship embedding vectors of the knowledge graph in training, the training learning rate, the number of negative samples, the adversarial negative sampling parameter, the number of training iterations, and the margin parameter; Dividing multiple basic unit triples into a training set and a test set; For each positive example triple in the training set, generating a negative example triple through the current entity and relationship embeddings. The specific method is: randomly select an entity from the entity dataset to replace the head entity or tail entity in the current positive example triple to generate a new negative example triple; Training the model using the training set: Optimizing the model using the gradient descent algorithm according to the scoring function formula; Using the trained model to perform working mode recognition on the test set, calculating the scoring function of the test sample and the embedding vectors corresponding to each radar working mode entity, and taking the working mode corresponding to the minimum scoring function as the final recognition result.

7. A multifunctional radar working mode recognition method based on the knowledge graph ROTATE model according to claim 6, characterized in that During the process of training the model, recording the scoring function of the negative example triple and using it as the decision threshold for the working mode of unknown models, and recognizing the working mode of unknown models, including: Setting the decision threshold for the working mode of unknown models according to the scoring function value of the negative example triple, comparing the scoring function value of any appearing sample with the decision threshold. If it is less than the threshold, the sample is recognized as a known working mode; otherwise, it is recognized as the working mode of an unknown model.

8. A multifunctional radar working mode recognition system based on the knowledge graph ROTATE model, characterized in that Including: A radar signal data generation module, used to randomly generate a radar sample dataset according to the preset parameter range and parameter modulation method of each radar working mode; A data preprocessing module, used to preprocess the radar sample data to form basic unit triples for constructing the knowledge graph; A knowledge graph construction module, used to construct a knowledge graph describing the relationship between radar working modes and parameter values and parameter modulation types based on the basic unit triples; A model training module, used to map the entities and relationships of the knowledge graph into a low-dimensional continuous vector space for representation, and using the scoring function defined by the ROTATE model to calculate the confidence of each basic unit triple, defining that the confidence of the basic unit triple is inversely proportional to the scoring function, and maximizing the confidence of the basic unit triple to train the model; An identification module, configured to perform working mode identification on a test set using a trained model.

9. A computer program product, which, when running on a terminal device, causes the terminal device to perform operations corresponding to the multi-functional radar working mode identification method based on the knowledge graph ROTATE model according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, The computer program, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images