A zero sample generation method and device based on expert knowledge and Gaussian SMOTE

Abstract

The application discloses a zero sample generation method and device based on expert knowledge and Gaussian SMOTE, generates initial sample data by using a Gaussian noise addition method according to radar rules and expert knowledge accumulated through public data inquiry or other channels, further discriminates and screens the initial sample by using a K nearest neighbor algorithm to form seed samples, and finally expands the seed sample data by using an oversampling technology to form sample data suitable for intelligent algorithm training, so as to improve the generalization ability and applicability of the intelligent algorithm, improve the identification accuracy of the zero sample radar model, and solve the problem that the artificial intelligence identification method based on sample training cannot adapt to the situation that part of radar signals are less collected, and sample is missing in electronic reconnaissance and electromagnetic spectrum processing.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic spectrum processing technology, and in particular to a zero-sample generation method and apparatus based on expert knowledge and Gaussian SMOTE. Background Technology

[0002] The electromagnetic spectrum is characterized by a wide variety of information types, complex and variable signals, and high randomness. Timely, effective, and comprehensive acquisition of electromagnetic information plays a fundamental role in accurately perceiving the electromagnetic situation. Radar radiation source identification is an important part of this process: based on signal sorting, the detected radar parameters are analyzed to accurately identify the radar model or even individual radars, providing important support for a comprehensive understanding of electromagnetic spectrum information.

[0003] In recent years, with the continuous upgrading and updating of radar technology, radar signals have become characterized by their large quantity, high density of overlap, dynamic overlap, and rapid pattern changes. The development of big data and artificial intelligence technologies has brought new approaches to radar radiation source identification: based on a large amount of received data, machine learning training methods can be used to automatically extract radar features and construct an identification space, thereby solving the problem of accurate identification of new and multifunctional radar systems. Common intelligent radar radiation source identification methods include machine learning-based identification methods and neural network-based identification methods.

[0004] Public research indicates that, compared to traditional library matching methods, intelligent radar source identification methods, due to their nonlinear capabilities based on data-driven autonomous learning, can finely distinguish between models with overlapping parameters, significantly improving the accuracy of identifying new radar systems. However, due to the non-cooperative nature of electromagnetic space, there is insufficient detection data for a large number of new radar systems, with some radar signals even exhibiting a "zero" state. This fails to meet the prerequisite of intelligent identification methods requiring a large amount of sample data, thus limiting their application in the field of electromagnetic spectrum processing. Summary of the Invention

[0005] The main objective of this invention is to provide a zero-sample generation method and apparatus based on expert knowledge and Gaussian SMOTE, aiming to solve the problem that in current electronic reconnaissance and electromagnetic spectrum processing, some radar signals are poorly received and samples are missing, making artificial intelligence recognition methods based on sample training unsuitable.

[0006] To achieve the above objectives, this invention provides a zero-shot generation method based on expert knowledge and Gaussian SMOTE, the method comprising the following steps:

[0007] S1: Construction of radar signal pattern rules;

[0008] S2: Sample generation based on Gaussian function;

[0009] S3: Sample discrimination and screening based on KNN;

[0010] S4: Sample augmentation based on SMOTE.

[0011] Optionally, step S1 specifically includes: acquiring radar radiation source rules accumulated by experts, and converting the radar radiation source rules into a set of multiple signal patterns of the target radar model.

[0012] Optionally, step S2 specifically includes:

[0013] S21: For each combination of the signal patterns in the set, generate the parameters corresponding to each combination using the Gaussian function noise addition method;

[0014] S22: Obtain initial samples based on the relationship between the combinations determined for each signal pattern and the parameters corresponding to each combination;

[0015] S23: Repeat steps S21-S22 until all signal patterns corresponding to the radar radiation source rules have been traversed to obtain the initial sample set.

[0016] Optionally, step S3 specifically includes:

[0017] S31: For each combination, calculate the Euclidean distance between the initial sample and every sample in the sample set of all models to obtain the distance set;

[0018] S32: Sort the distance values ​​in the distance set in ascending order and take the sample corresponding to the minimum distance of the first preset number of samples;

[0019] S33: Obtain the category labels corresponding to a preset number of neighboring samples from the label set, calculate the overlap of the initial samples, and remove the sample if the overlap is greater than the preset value.

[0020] S34: Repeat steps S31-S34 to obtain the filtered seed sample set.

[0021] Optionally, step S4 specifically includes:

[0022] S41: For all models that have been accumulated, calculate the average number of samples.

[0023] S42: Determine whether the number of seed samples in the seed sample set is less than the average number of samples. If so, use the SMOTE method to amplify the samples of this type.

[0024] S43: Perform step S42 for each type of seed sample to complete the amplification of all seed samples.

[0025] Optionally, step S42 specifically includes:

[0026] S421: Determine the sample amplification ratio based on the number of samples of all models and the number of seed samples. If the sample amplification ratio is greater than 1, proceed to step S422.

[0027] S422: For each sample in the seed sample set, amplify the sample according to the sample amplification ratio and the preset number of nearest neighbors.

[0028] Optionally, if the sample amplification rate is not greater than 1, return to step S34.

[0029] Optionally, step S422 specifically includes:

[0030] S4221: For samples in the seed sample set Calculate its M nearest neighbor samples Where 1≤h≤M, forming Nearest neighbor sample set

[0031] S4222: From the nearest neighbor sample set Randomly selected samples And generate a random number τ between [0,1];

[0032] S4223: Using linear interpolation, synthesize... and New samples ;

[0033] S4224: Add new sample Add to the seed sample set.

[0034] Furthermore, to achieve the above objectives, the present invention also provides a zero-shot generation device based on expert knowledge and Gaussian SMOTE, the device comprising:

[0035] A building module for constructing radar signal pattern rules;

[0036] The generation module is used for sample generation based on the Gaussian function;

[0037] The filtering module is used for sample discrimination and filtering based on KNN;

[0038] An expansion module for SMOTE-based sample expansion.

[0039] This invention proposes a zero-shot generation method and apparatus based on expert knowledge and Gaussian SMOTE. Based on radar rules and expert knowledge accumulated through publicly available information or other channels, initial sample data is generated using a Gaussian noise addition method. Further, the K-nearest neighbor algorithm is used to discriminate and filter the initial samples, forming seed samples. Finally, oversampling technology is used to expand the seed sample data, forming sample data suitable for training intelligent algorithms. This improves the generalization ability and applicability of the intelligent algorithm, enhancing the accuracy of zero-shot radar model identification. By combining accumulated rule knowledge, sample generation for zero-shot models is achieved, solving the problem of difficulty in identifying radars with known rules but missing samples. This expands the adaptability of intelligent algorithms and improves the accuracy of equipment and systems in identifying radar models. It also addresses the problem in electronic reconnaissance and electromagnetic spectrum processing where some radar signals are poorly received and samples are missing, making sample-based artificial intelligence identification methods unsuitable. Attached Figure Description

[0040] Figure 1 This is a schematic diagram illustrating the principle of the SMOTE algorithm.

[0041] Figure 2 This is a schematic diagram of the overall framework of the zero-shot generation method based on expert knowledge and Gaussian SMOTE in an embodiment of the present invention.

[0042] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0043] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0044] Currently, in the relevant technical fields, some radar signals are poorly detected and lack samples in electronic reconnaissance and electromagnetic spectrum processing, making it impossible for artificial intelligence recognition methods based on sample training to adapt.

[0045] To address this problem, various embodiments of the zero-shot generation method based on expert knowledge and Gaussian SMOTE are proposed in this invention. The zero-shot generation method based on expert knowledge and Gaussian SMOTE provided by this invention combines accumulated rule knowledge to generate samples for zero-shot radar models, thereby solving the problem of identification difficulty for radar models where some rules are known but samples are missing. This expands the adaptability of intelligent algorithms and improves the accuracy of equipment and systems in identifying radar models. It also solves the problem that in electronic reconnaissance and electromagnetic spectrum processing, some radar signals are poorly received and samples are missing, making sample-based artificial intelligence identification methods unsuitable.

[0046] This embodiment utilizes parameter rules accumulated by experts to add Gaussian noise to the radio frequency, repetition interval, and pulse width, extracting initial samples from the signal pattern rules. Furthermore, the KNN algorithm is used to discriminate and filter the initial samples, eliminating samples that overlap with other models to form seed samples. Finally, the SMOTE method is used to generate training samples, thereby forming the sample data required for intelligent recognition for zero-sample models, meeting the training requirements of the intelligent recognition algorithm, and improving recognition accuracy.

[0047] It's important to note that the K-Nearest Neighbors (KNN) algorithm is a common classification method in statistical analysis. Essentially, it uses statistical probability principles to automatically learn from known samples, build a feature system, and predict the category of unknown samples. Specifically, it finds the K nearest neighbors of the test sample from the known samples and determines the category of the test sample based on the categories of these K nearest neighbors. KNN has been widely used in text classification, pattern recognition, image classification, and other fields.

[0048] Synthetic Minority Oversampling Technique (SMOTE) is a widely used data preprocessing method for imbalanced classification problems. The basic idea of ​​SMOTE is to linearly interpolate between each minority class sample and its K nearest neighbors within each class, randomly generating a new sample. Because the synthesized sample is a random value between two samples, it effectively increases the diversity of the minority class samples, solving the overfitting problem caused by random oversampling. A schematic diagram of the SMOTE principle is shown below. Figure 1 As shown, taking K=5 as an example, in the minority class sample X i and its 5 nearest neighbor samples X i1 ,...,X i5 Perform linear interpolation to generate new samples.

[0049] like Figure 2 The diagram shown is a schematic representation of the overall framework of the zero-shot generation method based on expert knowledge and Gaussian SMOTE in this embodiment.

[0050] Specifically, this invention mainly includes three parts: (1) sample generation based on Gaussian function, (2) sample discrimination and screening based on KNN, and (3) sample expansion based on SMOTE.

[0051] (1) Sample generation based on Gaussian function.

[0052] The first part of this technical solution involves sample generation based on Gaussian functions. The radar radiation source rules accumulated by experts can be expressed as a set Π={sig} of multiple signal patterns for that radar model. i Each signal pattern can be represented as:

[0053] sig i =[rule RF rule PRI rule Pw ] T

[0054] Among them, rule RF rule PRI rule PW This can be represented as a combination of typical value and tolerance:

[0055] rule = {(ρ k ,σ k )}

[0056] Where, ρ k For typical values, σ k This corresponds to the tolerance. The Gaussian function is defined as:

[0057]

[0058] The specific steps for generating samples based on Gaussian functions are as follows:

[0059] Step 1: For each combination R of RF, PRI, and PW, generate parameters using the Gaussian function noise addition method. Taking RF as an example, the generated parameter para... rf for:

[0060] para rf =ρ k,rf +G s (σ k,rf ,x)

[0061] Where σ k,rf For ρ k,rf The corresponding variance, where x is a random number. para pri and para pw The generation method is the same.

[0062] Step 2, according to signal pattern S i Determined rule RF rule PRI rule PW Combining relationships to form the initial generated samples:

[0063] para init =[para rf ,para pw ,para pri ] T

[0064] Step 3: Repeat steps 1-2 until all rules for model R have been traversed. The initial sample generated using the Gaussian function is as follows:

[0065] PARA init,R ={para init,k}

[0066] (2) Sample discrimination and screening based on KNN.

[0067] The second part of this technical solution is sample discrimination and screening based on KNN. For the Gaussian simulation samples generated in the first part, the KNN method is used to remove overlapping samples with other models. The specific steps are as follows:

[0068] Step 4, for an initial sample of model R In all sample sets of all models In j = 1, ..., sampleNum, calculate With each sample The Euclidean distance is as follows:

[0069]

[0070] Step 5, for the distance set j = 1, ..., sampleNum The smaller the value, the more similar the two samples are to the set. Sort in ascending order and take the samples corresponding to the first K minimum distances. k = 1, ..., K.

[0071] Step 6, from the label set Label = {L j Obtain the class labels of the K nearest neighbor samples from the sampleNum, j=1,...,sampleNum, and calculate the sample... degree of overlap:

[0072]

[0073] overlap level i To indicate the degree of overlap, num other Let K be the number of samples belonging to other categories among the K nearest neighbors.

[0074] Step 7: Filter samples based on the degree of overlap. If the overlap level... i If the value is greater than δ, it indicates that the sample overlaps significantly with other models and needs to be removed.

[0075] Step 8, repeat steps 4 to 7. R Next, the filtered seed sample set S is obtained. sel,R ={parasel,k}

[0076] (3) Sample augmentation based on SMOTE.

[0077] The third part of this technical solution is a sample amplification method based on SMOTE. The seed samples formed in the second part are amplified using the SMOTE method. The specific steps are as follows:

[0078] Step 9, Calculate the average sample size. For the accumulated sample S of all models... 1 ,S 2 ,…,S L Calculate its average sample size in The number of samples for model i.

[0079] Step 10, Sample amplification processing. Regarding the number of seed samples, if... The SMOTE method is then used to amplify this type of sample, specifically as follows:

[0080] Step 10.1, Calculate the sample amplification rate. Calculate the sample amplification rate n = |N / N sel,R If n≤1, then no sample processing is required, proceed to step 8.

[0081] Step 10.2, Minority Class Sample Synthesis. For seed sample S sel,R For each sample in the dataset, amplification is performed based on the sample amplification fold n and the preset number of nearest neighbors L, specifically as follows:

[0082] Step 10.2.1, for the sample Calculate its M nearest neighbor samples Where 1≤h≤M, forming Nearest neighbor sample set .

[0083] Step 10.2.2, from the nearest neighbor sample set Randomly selected samples And generate a random number τ between [0,1].

[0084] Step 10.2.3: Use linear interpolation to synthesize... and New samples The linear interpolation formula is:

[0085]

[0086] Step 10.2.4, take the new sample Add to sample set S sel,R middle.

[0087] Step 10.3: Repeat steps 10.2.2 to 10.2.4 n times until n new samples are added.

[0088] Step 10.4, for the minority class sample set S i Each sample in Repeat steps 10.2 to 10.3.

[0089] Step 11: For each type of seed sample, perform step 10 to complete the amplification of all seed samples.

[0090] Step 12, the sample set formed after the SMOTE method is S * The expanded sample has a balanced number of classes, making it suitable for training various machine learning algorithms.

[0091] This embodiment provides a zero-shot generation method based on expert knowledge and Gaussian SMOTE. The specific effect of this invention in practical engineering projects is that, for situations where radar detection data is scarce and samples are missing in electronic reconnaissance and electromagnetic spectrum processing, a zero-shot generation method based on expert knowledge and Gaussian SMOTE is proposed. This method combines rule-based knowledge accumulated by experts and gradually applies Gaussian noise addition, KNN, and SMOTE methods to generate samples. This satisfies the prerequisite of intelligent recognition methods requiring a large number of samples, expands the applicability of intelligent recognition, and ultimately improves the accuracy of equipment and systems in identifying radar models.

[0092] To explain this application more clearly, a specific example of a zero-shot generation method based on expert knowledge and Gaussian SMOTE is provided below, based on the steps described above.

[0093] In this example, the zero-shot generation method based on expert knowledge and Gaussian SMOTE is applied to the identification of radar radiation sources in the fields of electronic reconnaissance and electromagnetic processing. Taking the generation of radar model samples with missing four types of samples but regular radar signal patterns as an example, the specific implementation scheme of this technology is as follows:

[0094] First, steps 1 to 3 are performed to generate samples using a Gaussian function. Taking radar model A as an example, the rules for signal pattern 1 are shown below, and an example of samples generated using the Gaussian function is shown below. A schematic diagram of the generated initial samples is shown in Table 2.

[0095] Table 1: Signal Pattern Rules of Radar Model A

[0096]

[0097] Table 2: Initial Samples Generated According to Signal Pattern 1 Rules of Model A

[0098]

[0099] The initial samples are filtered using KNN, with K set to 5 and δ set to 0.5 in step 7. Steps 4 to 8 are then executed to generate seed samples for four types of radar.

[0100] Finally, the SMOTE-based method is used to augment the samples, following steps 9 to 12. This forms the training sample set S. * .

[0101] In summary, the number of rules for the four types of radar and the final number of training samples are shown in Table 3.

[0102] Table 1: Number of Rules and Samples for Four Types of Radar

[0103]

[0104] To demonstrate the effectiveness of this technical solution, the generated samples were used to train the convolutional neural network (CNN) model, and the test data was used for testing. The recognition accuracy is shown in Table 4.

[0105] Table 4: Identification Results for Different Models

[0106]

[0107] As shown in Table 4, the samples generated using this technical solution, combined with the intelligent recognition algorithm, can accurately identify zero-sample radar models (with an accuracy of no less than 90%).

[0108] In a preferred embodiment, the present invention proposes a zero-shot generation device based on expert knowledge and Gaussian SMOTE, comprising:

[0109] A building module for constructing radar signal pattern rules;

[0110] The generation module is used for sample generation based on the Gaussian function;

[0111] The filtering module is used for sample discrimination and filtering based on KNN;

[0112] An expansion module for SMOTE-based sample expansion.

[0113] Other embodiments or specific implementations of the zero-shot generation device based on expert knowledge and Gaussian SMOTE of the present invention can be referred to the above-described method embodiments, and will not be repeated here.

[0114] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0115] Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware, and of course, it can also be implemented by special hardware including application-specific integrated circuits, special CPUs, special memory, special components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for the present invention, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, portable hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

Claims

1. A zero-shot generation method based on expert knowledge and Gaussian SMOTE, characterized in that, The method comprises the following steps: S1: construction of a radar signal pattern rule; S2: sample generation based on a Gaussian function; S3: sample discrimination screening based on KNN; S4: sample expansion based on SMOTE; The step S1 specifically comprises: acquiring radar emitter rules accumulated by experts, and converting the radar emitter rules into a set of multiple signal patterns of a target radar model; The step S2 specifically comprises: S21: for each combination of the signal patterns in the set, generating parameters corresponding to each combination by using a Gaussian function noise method; S22: obtaining initial samples according to the relationship between each combination of the signal patterns and the parameters corresponding to each combination; S23: repeating steps S21-S22 until all signal patterns corresponding to the radar emitter rules are traversed, to obtain an initial sample set.

2. The method for zero-shot generation based on expert knowledge and Gaussian SMOTE as claimed in claim 1, wherein, The step S3 specifically comprises: S31: for each initial sample corresponding to a combination, calculating the Euclidean distance between the initial sample and each sample in the sample set of all models to obtain a distance set; S32: arranging distance values in the distance set in ascending order, and taking samples corresponding to the smallest distances in a preset number of samples; S33: obtaining category labels corresponding to the preset number of neighbor samples from a label set, calculating the overlap degree of the initial sample, and if the overlap degree is greater than a preset value, eliminating the sample; S34: repeating steps S31-S34 to obtain a screened seed sample set.

3. The method of claim 2, wherein the method is based on expert knowledge and Gaussian SMOTE. The step S4 specifically comprises: S41: calculating the average sample number of all models of samples accumulated; S42: determining whether the number of seed samples in the seed sample set is less than the average sample number, and if so, performing sample expansion on the seed samples by using the SMOTE method; S43: performing step S42 on seed samples of each model to complete expansion of all seed samples.

4. The method of claim 3, wherein the method is based on expert knowledge and Gaussian SMOTE. The step S42 specifically comprises: S421: determining a sample expansion rate according to the number of samples of all models and the number of seed samples, and if the sample expansion rate is greater than 1, performing step S422; S422: for each sample in the seed sample set, performing sample expansion according to the sample expansion rate and a preset number of neighbors.

5. The method of claim 4, wherein the method is based on expert knowledge and Gaussian SMOTE. If the sample expansion rate is not greater than 1, return to perform step S34.

6. The method of generating zero-shot based on expert knowledge and Gaussian SMOTE as claimed in claim 4, wherein, The step S422 specifically comprises: S4221: for a sample in the set of seed samples , compute its M nearest neighbor samples , where , form the set of nearest neighbor samples of the sample ; S4222: randomly select a sample from the set of neighboring samples from the set of neighboring samples and generate a random number τ in [0, 1] S4223: Synthesis of new samples between S4222 and S4224 using linear interpolation method and ;​ S4224: Add new sample to the seed sample set. is added to the seed sample set.

7. A zero-shot generation apparatus based on expert knowledge and Gaussian SMOTE, for implementing the method of any one of claims 1-6, characterized in that, The device comprises: A construction module for construction of a radar signal pattern rule; A generation module for sample generation based on a Gaussian function; A screening module for sample discrimination screening based on KNN; An expansion module for sample expansion based on SMOTE.

Images