Radar signal pulse description word data generalization method based on generative pre-training Transform
Through the method based on the generative pre-training Transformer, the radar signal pulse description word data is preprocessed and word-partitioned, and new data is generated using the GPT-2 model, which solves the problems of insufficient data generalization capabilities and high computational complexity in the existing methods, and achieves high-quality and low-cost data generalization enhancement.
Patent Information
- Application Number
- CN202411872193.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-18
- Publication Date
- 2025-04-22
AI Technical Summary
When the existing generative network method realizes the generalization enhancement of radar signal pulse description word data, there are problems such as high requirements for the number of learning samples, unstable training process, high computational complexity and high training time cost.
Using a method based on generative pre-training Transformer, the radar signal pulse description word data is preprocessed and improved word segmentation processing is performed. The data is input into the fine-tuned GPT-2 model in text form, and its powerful text generation ability is used to generate new and diverse pulse description word data.
Through this method, the training and simulation environment of the radar signal sorting system can be effectively enhanced, and the generated data sets are highly representative and diverse, solving the problem of insufficient data generalization capabilities in the existing methods.
Smart Images
Figure CN119862369A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of data processing, and in particular relates to a radar signal pulse description word data generalization method based on a generative pre-trained Transformer. Background Art
[0002] The increasingly complex electromagnetic environment has seriously threatened the accuracy of radar emitter identification. The radar emitter sorting method based on deep learning is heavily dependent on the quantity and accuracy of pre-training sample data. Enhancing the generalization of training data from the data generation quality level is a prerequisite for accurately training classifiers and improving the classifier's ability to accurately distinguish the type and style of emitters. Therefore, studying high-quality generalization methods for radar signal PDW datasets can provide guarantees for the quality of classification models and has important application value.
[0003] At present, the mainstream methods for generating generalized discrete data similar to PDW include synthetic minority oversampling technology (SMOTE), generative adversarial networks (GAN), variational autoencoders (VAEs), and diffusion models (DDPM). SMOTE is a method commonly used to deal with unbalanced data sets, which increases the number of minority class samples by interpolating and generating new samples. GAN generates high-quality samples through adversarial training of generators and discriminators. VAEs use encoders to compress input data into latent space, and then use decoders to reconstruct input data from latent space to generate high-quality samples. DDPM destroys data by gradually adding noise, and then learns the inverse process to reconstruct data from noise to generate high-quality samples. However, in the prior art, the generalization enhancement of data set samples using generative networks overcomes the shortcomings of artificial dependence and poor robustness of traditional enhancement methods. However, the existing generative network data generation methods have problems such as high requirements on the number of learning samples, unstable training process, high computational complexity, and high training time cost. For example, the new samples generated by SMOTE are interpolations between existing samples, but these samples are still very close to the original samples, and cannot capture the true distribution of the data well, and even have overfitting problems. The training process of GAN is unstable, and the learning of network structure and parameter estimation on high-dimensional data have high computational complexity, and training usually requires a lot of computing resources and time. The training of VAEs involves complex mathematical derivations such as variational inference, and the training complexity is high. The DDPM training process needs to simulate the forward and reverse processes of multiple time steps and perform multiple back propagations. The reverse process of the model is difficult to learn, the computational complexity is high, and the training time is usually long. Therefore, it is of great significance to study fast, high-quality, and low-cost PDW dataset generalization methods. Summary of the invention
[0004] In order to solve the above problems existing in the prior art, the present invention provides a radar signal pulse description word data generalization method based on generative pre-trained Transformer. The technical problem to be solved by the present invention is achieved through the following technical solutions:
[0005] In a first aspect, the present invention provides a radar signal pulse description word data generalization method based on a generative pre-trained Transformer, comprising:
[0006] Obtain a PDW data training set, where the PDW data training set includes multiple PDW data samples, and each PDW data sample includes multiple features;
[0007] Preprocess the PDW data training set and create a dictionary;
[0008] Perform word segmentation on the preprocessed PDW data training set to obtain a data set that meets the language model processing requirements;
[0009] Inputting a data set that meets the language model processing requirements into an initial language model for training to obtain a trained language model, and the trained language model generates a text format training set based on the data set that meets the language model processing requirements;
[0010] According to the dictionary, the text format training set is decoded to obtain the generalized and expanded PDW data training set.
[0011] In a second aspect, the present invention further provides a radar signal pulse description word data generalization device based on a generative pre-trained Transformer, comprising:
[0012] A data acquisition module is used to acquire a PDW data training set, where the PDW data training set includes multiple PDW data samples, and each PDW data sample includes multiple features;
[0013] Data preprocessing module, used to preprocess the PDW data training set and create a dictionary;
[0014] The data segmentation processing module is used to perform segmentation processing on the pre-processed PDW data training set to obtain a data set that meets the language model processing requirements;
[0015] A data processing module is used to input a data set that meets the language model processing requirements into an initial language model for training to obtain a trained language model. The trained language model generates a text format training set based on the data set that meets the language model processing requirements.
[0016] The data decoding module is used to decode the text format training set according to the dictionary to obtain the generalized and expanded PDW data training set.
[0017] Beneficial effects of the present invention:
[0018] The present invention provides a radar signal pulse description word data generalization method based on generative pre-trained Transformer. By preprocessing PDW data and improving word segmentation processing, PDW discrete data is input into a fine-tuned GPT-2 model in the form of text. It not only utilizes the powerful text generation capability of GPT-2, but also adapts to the characteristics of radar PDW signal data through specific preprocessing and word segmentation techniques, so that it can process and understand the unique format and semantics of PDW data. Furthermore, by training the model to learn the intrinsic patterns and variations of PDW data, new and diverse pulse description word data can be generated, which is of high value for enhancing the training and simulation environment of radar signal sorting systems.
[0019] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 It is a flow chart of a radar signal pulse description word data generalization method based on a generative pre-trained Transformer provided in an embodiment of the present invention;
[0021] Figure 2 It is a schematic diagram of a radar signal pulse description word data generalization method based on a generative pre-trained Transformer provided in an embodiment of the present invention;
[0022] Figure 3 It is a schematic diagram of the cumulative probability density distribution of the radar radiation source PDW generated data provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0023] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.
[0024] In view of the defects of the prior art mentioned above, the present invention plans to solve the problem of insufficient representativeness and diversity of PDW sample data sets due to system noise, measurement error, insufficient sample collection, etc. in the training set preparation stage of training radar emitter classification models based on radar emitter pulse descriptive words (PDW) using neural network related methods. The present invention combines the idea of natural language generation to propose a radar signal pulse descriptive word data generalization enhancement method based on the generative pre-training model (Generative Pre-Training-2, GPT-2), and uses a text-based word segmentation method to avoid the dimensionality explosion problem encountered when processing high-dimensional data, thereby achieving rapid convergence of the model; by conducting targeted training on the initialized large language model, the model's adaptability to specific categories of PDW data is enhanced, and an efficient, accurate and stable radar signal PDW data set generalization enhancement method is proposed.
[0025] See also Figure 1 , Figure 1 is a flow chart of a radar signal pulse description word data generalization method based on a generative pre-trained Transformer provided in an embodiment of the present invention. Figure 2 is a schematic diagram of a radar signal pulse description word data generalization method based on a generative pre-trained Transformer provided by an embodiment of the present invention. The radar signal pulse description word data generalization method based on a generative pre-trained Transformer provided by the present invention includes:
[0026] S101. Obtain a PDW data training set, where the PDW data training set includes multiple PDW data samples, and each of the PDW data samples includes multiple features.
[0027] Specifically, in this embodiment, the radar emitter signal modulation mechanism is used to simulate and establish a radar emitter signal PDW data set. In terms of characterizing the radar emitter pulse, each pulse is usually characterized by a PDW containing 5-dimensional parameters, including: arrival time (TOA), arrival direction (DOA), radiation source frequency (RF), pulse width (PW), and pulse amplitude (PA). In order to learn the distribution of a specific type of radar emitter signal pulse in a high-dimensional parameter space (DOA, RF, PW, PA), the ability of the network to generate new data to expand the training set of each type of PDW data is tested.
[0028] In this embodiment, the PDW data sample includes at least the characteristics of arrival direction, radiation source frequency, pulse width and pulse amplitude, that is, the simulated PDW data training set is set with four characteristics: {DOA, RF, PW, PA}.
[0029] S102: Preprocess the PDW data training set and create a dictionary.
[0030] Specifically, in this embodiment, the four features {DOA, RF, PW, PA} are used to describe the numerical expression of certain aspects of radar signal information, which is different from general natural language information. Therefore, it is necessary to adjust the word segmentation strategy according to the needs of the application scenario of radar signal data generalization enhancement. Specifically:
[0031] like Figure 2 As shown in the figure, first, all feature columns and corresponding feature data of the PDW data sample are converted to string type; secondly, a unique integer label is assigned to each feature data, and a dictionary is created to save the category feature data and integer labels and their mapping relationship for subsequent decoding; finally, the label encoder is used to convert the feature data into numerical data, and the dictionary containing the column information and the label encoder is added to the label encoder list, and the encoded data frame is returned.
[0032] S103: Perform word segmentation processing on the preprocessed PDW data training set to obtain a data set that meets the language model processing requirements.
[0033] For more details, please see Figure 2 In this embodiment, for the input table type training data, first, each row of feature data is converted and regarded as a sentence; secondly, the input data is converted into a data frame containing input identifier and attention mask information based on the word segmentation processing method; then, the processed data frame is converted into a standardized open source platform (HuggingFace) dataset object, so that the diverse data processing model training functions provided by the open source library of the open source platform can be utilized.
[0034] In this embodiment, the word segmentation process is completely based on text, which effectively avoids the problem of dimensional explosion and reliance on large sample training data when the generation network processes high-dimensional data. At the same time, after the feature data sentences are word segmented, the length of the training batch sequence can be guaranteed to be consistent, which significantly reduces the training time and improves the overall effectiveness and efficiency of the generated data.
[0035] S104: input the data set that meets the language model processing requirements into the initial language model for training to obtain a trained language model. The trained language model generates a text format training set based on the data set that meets the language model processing requirements.
[0036] For more details, please see Figure 2 In this embodiment, a randomly initialized multi-layer Transformer-based GPT2 language model (distilgpt2) is used to specifically train the PDW data table to optimize the model. In this way, the use of targeted optimization of the pre-trained language model emphasizes the adaptability of the generative model to the PDW discrete table data, and improves the generation rate, accuracy and stability of the model. The actual results show that compared with a well-trained language model, training PDW data based on a randomly initialized language model will make the model converge faster.
[0037] In this embodiment, a data set that meets the language model processing requirements is input into the initial language model for training to obtain a trained language model, including:
[0038] The loss function used in training is minimized within the preset number of iterations, and the parameters of the language model are determined;
[0039] Among them, the loss function used in training is the cross entropy loss function.
[0040] In this embodiment, the trained language model generates a text format training set based on a data set that meets the language model processing requirements, including:
[0041] like Figure 2 As shown, the trained language model (LM1) is loaded, the data set that meets the language model processing requirements input during the training process is used as input, the discrete column distribution learned during the training process is used as the target distribution, text format data is generated, and then it is converted back to table format.
[0042] It should be noted that using GPT-2 as a basic language model and fine-tuning it to handle generative tasks has significant advantages, because the pre-training foundation of GPT-2 is extremely powerful and can understand and generate complex text structures. This model performs well in generating fluent and coherent text. By fine-tuning the model, that is, performing word segmentation processing, its behavior can be adjusted according to the needs of specific tasks, making the model better adapted to the language usage in specific fields. These features together make GPT-2 a common and effective technical choice for technicians in the field of generative networks, especially in the field of language text-based generative networks.
[0043] S105: Decode the text format training set according to the dictionary to obtain the generalized and expanded PDW data training set.
[0044] Specifically, in this embodiment, the text format training set is converted into a generalized and expanded PDW data training set using feature data corresponding to the features stored in the dictionary, integer labels, and mapping relationships converted to strings, and the quality of the generated data is evaluated.
[0045] In summary, the present invention provides a radar signal pulse description word data generalization method based on generative pre-trained Transformer. By preprocessing PDW data and improving the word segmentation processing, the PDW discrete data is input into the fine-tuned GPT-2 model in the form of text. It not only utilizes the powerful text generation ability of GPT-2, but also adapts to the characteristics of radar PDW signal data through specific preprocessing and word segmentation technology, so that it can process and understand the unique format and semantics of PDW data. Furthermore, by training the model to learn the intrinsic patterns and variations of PDW data, new and diverse pulse description word data can be generated, which is of high value for enhancing the training and simulation environment of radar signal sorting systems.
[0046] In an optional embodiment of the present invention, the effectiveness of the method provided in the above embodiment is verified by simulation experiments, specifically:
[0047] 1. Dataset
[0048] In order to verify the reliability of the method proposed in the present invention, the experimental data used is a simulated PDW data set of radar radiation source signals with specific feature categories, the number of pulses is 100, the DOA range is 62~128°, the RF agility range is 5208~5213GHz, the PW agility range is 12~14μs, and the PA range is -40~-30.
[0049] Implementation Details
[0050] The hardware platform of the simulation experiment of the present invention is: 13th Gen Intel(R)Core(TM)i9-13905H, fourteen cores, main frequency of 2.60GHz, memory size of 32GB, graphics card of 4090, and video memory size of 24GB. The present invention uses Adam optimizer and adaptive linear learning rate scheduler, uses cross entropy loss function, and sets the initial learning rate to 5e-5. The training samples are input into the network for training, and 100 epochs are trained for all models, and global parameter adjustment is performed and the parameters of the optimal model are saved.
[0051] Load the trained language model (LM1), use the data set that meets the language model processing requirements input during the training process as input, and the discrete column distribution learned during the training process as the target distribution, generate text format data, and then convert it back to table format, and finally restore the generated data to the original PDW sample data format according to the dictionary, and compare it with the evaluation results of other table data synthesis methods
[0052] See also Figure 3 , Figure 3It is a schematic diagram of the cumulative probability density distribution of the radar radiation source PDW generated data provided by an embodiment of the present invention. The results show that the radar signal pulse descriptor data generalization enhancement method based on generative pre-trained Transformer proposed in the present invention is closest to the real sample data in terms of the cumulative probability density distribution and the change with the index of the data set data generated by it compared with SMOTE, GAN, VAEs, and DDPM, showing the reliability of generalizing PDW data based on this algorithm.
[0053] Based on the same inventive concept, the present invention also provides a radar signal pulse description word data generalization device based on a generative pre-trained Transformer, which is used to implement the method provided by the above embodiment of the present invention. The embodiment of the method is referred to above and will not be repeated here. The device includes:
[0054] A data acquisition module, used to acquire a PDW data training set, wherein the PDW data training set includes a plurality of PDW data samples, and each of the PDW data samples includes a plurality of features;
[0055] Data preprocessing module, used to preprocess the PDW data training set and create a dictionary;
[0056] A data segmentation processing module is used to perform segmentation processing on the pre-processed PDW data training set to obtain a data set that meets the language model processing requirements;
[0057] A data processing module is used to input a data set that meets the language model processing requirements into an initial language model for training to obtain a trained language model, and the trained language model generates a text format training set based on the data set that meets the language model processing requirements;
[0058] The data decoding module is used to decode the text format training set according to the dictionary to obtain a generalized and expanded PDW data training set.
[0059] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants are intended to cover non-exclusive inclusion, so that the article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the article or device including the elements. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The orientation or position relationship indicated by "up", "down", "left", "right", etc. is based on the orientation or position relationship shown in the accompanying drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention.
[0060] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.
[0061] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.
Claims
1. A radar signal pulse descriptor data generalization method based on generative pre-trained Transformer, characterized in that: include: Acquire a PDW data training set, wherein the PDW data training set includes a plurality of PDW data samples, and each of the PDW data samples includes a plurality of features; Preprocessing the PDW data training set and creating a dictionary; Performing word segmentation processing on the preprocessed PDW data training set to obtain a data set that meets the language model processing requirements; Inputting the data set that meets the language model processing requirements into the initial language model for training to obtain a trained language model, wherein the trained language model generates a text format training set according to the data set that meets the language model processing requirements; According to the dictionary, the text format training set is decoded to obtain a generalized and expanded PDW data training set.
2. The radar signal pulse descriptor data generalization method based on generative pre-trained Transformer according to claim 1 is characterized in that: The PDW data samples include at least arrival direction, radiation source frequency, pulse width and pulse amplitude characteristics.
3. The radar signal pulse descriptor data generalization method based on generative pre-trained Transformer according to claim 1 is characterized in that: The preprocessing of the PDW data training set and creation of a dictionary include: Convert the features of all PDW data samples in the PDW data training set and the feature data corresponding to the features into string types, assign integer labels to the feature data corresponding to the features, and create the dictionary; Using a label encoder, the feature data is converted into numerical data.
4. The radar signal pulse descriptor data generalization method based on generative pre-trained Transformer according to claim 3 is characterized in that: The dictionary includes feature data corresponding to the feature, integer labels, and mapping relationships converted into strings.
5. The radar signal pulse descriptor data generalization method based on generative pre-trained Transformer according to claim 1 is characterized in that: The word segmentation processing of the pre-processed PDW data training set includes: Converting the feature data of the PDW data samples in the preprocessed PDW data training set into feature data in sentence form; Based on a word segmentation method, the feature data in the sentence form is converted into a data frame including input identifier and attention mask information; Convert the data frame into a standardized open source platform dataset object.
6. The radar signal pulse descriptor data generalization method based on generative pre-trained Transformer according to claim 1 is characterized in that: The language model is the GPT-2 language model.
7. The radar signal pulse descriptor data generalization method based on generative pre-trained Transformer according to claim 1 is characterized in that: Inputting the data set that meets the language model processing requirements into the initial language model for training to obtain a trained language model, including: The data set that meets the language model processing requirements is input into the initial language model training PWD data table to optimize the language model and generate language model parameters to construct the trained language model.
8. The radar signal pulse descriptor data generalization method based on generative pre-trained Transformer according to claim 1 is characterized in that: Inputting the data set that meets the language model processing requirements into the initial language model for training to obtain a trained language model, including: The loss function used in training is minimized within the preset number of iterations, and the parameters of the language model are determined; Among them, the loss function used in training is the cross entropy loss function.
9. A radar signal pulse description word data generalization device based on generative pre-trained Transformer, characterized in that: include: A data acquisition module, used to acquire a PDW data training set, wherein the PDW data training set includes a plurality of PDW data samples, and each of the PDW data samples includes a plurality of features; A data preprocessing module, used for preprocessing the PDW data training set and creating a dictionary; A data segmentation processing module is used to perform segmentation processing on the pre-processed PDW data training set to obtain a data set that meets the language model processing requirements; A data processing module, used for inputting the data set that meets the language model processing requirements into the initial language model for training to obtain a trained language model, wherein the trained language model generates a text format training set according to the data set that meets the language model processing requirements; The data decoding module is used to decode the text format training set according to the dictionary to obtain a generalized and expanded PDW data training set.