Incremental specific radiation source individual identification method and system based on knowledge calibration

Abstract

The invention discloses an incremental specific radiation source individual identification method and system based on knowledge calibration, belongs to the technical field of specific radiation source identification, and solves the problem that the stability of old category features declines in a long-term incremental learning process, so that the identification performance is reduced. Comprising the following steps: collecting initial radiation source individual signals, and dividing into a training set, a verification set and a test set; constructing a radiation source identification model, and identifying the initial weight of the model; equally dividing the memory bank capacity to each radiation source category, screening representative samples during each incremental learning, and supplementing and updating the memory bank; newly-added radiation source individual signals are collected and divided into a training set, a verification set and a test set; updating the radiation source identification model, expanding the classifier, and updating the model weight; adopting the updated radiation source individual identification model to detect radiation source signals on line; and detecting whether a new radiation source individual exists or not, and if not, completing identification. The method is suitable for low-altitude security and protection, airspace supervision, frequency spectrum game and other scenes.

Description

Technical Field

[0001] This invention belongs to the field of specific radiation source identification technology, and specifically relates to an incremental method for identifying individual specific radiation sources. Background Technology

[0002] Specific radiation source identification technology uses radio frequency fingerprints generated by analyzing inherent differences in the aircraft's hardware circuitry to uniquely identify individual radiation sources. This technology has significant military and civilian value in scenarios such as low-altitude security, airspace surveillance, and spectrum management.

[0003] However, existing research largely focuses on model optimization in closed scenarios, relying on fixed-class datasets and global model update mechanisms, which is insufficient to address the practical needs of the continuously expanding number of radiation sources in open electromagnetic environments. Therefore, designing a specific radiation source identification method with incremental update capabilities, using only newly added radiation source signals to update the model, is of practical significance.

[0004] A literature review of incremental learning techniques in the field of machine learning reveals that related research has made phased progress. For example, SA Rebuffi et al., in their paper "iCaRL: Incremental classifier and representation learning" published at the Conference on Computer Vision and Pattern Recognition (2017, 5533–5542), proposed an incremental learning method combining sample memory and knowledge distillation. This method maintains the recognition ability of old classes by preserving some historical samples and using logit distillation. A. Douillard et al., in their paper "PODNet: Pooled Outputs Distillation for Small-Tasks Incremental Learning" published at the European Conference on Computer Vision (2020, 86–102), proposed an incremental learning method based on feature space distillation. This method alleviates the catastrophic forgetting problem by introducing a pooling output distillation mechanism to align old task features at multiple scales. Although the above methods perform well in incremental learning tasks in computer vision, their application in specific radiation source recognition scenarios still faces many challenges. On the one hand, the time-varying characteristics of electromagnetic signals and the complexity of hardware differences far exceed those of image data, making it difficult for traditional feature space-based distillation methods to directly adapt to the dynamic changes of radio frequency fingerprints. On the other hand, the frequency and scale of incremental updates of radiation sources far exceed those of conventional computer vision tasks, placing higher demands on the real-time performance and memory capacity of the model.

[0005] Current incremental learning schemes generally suffer from the problem of catastrophic forgetting of old categories of knowledge during long-term incremental updates. The root cause is the unsustainability of memory. Continuous parameter updates cause feature representation drift, resulting in catastrophic forgetting of the categories learned in the early stages. Summary of the Invention

[0006] This invention provides an incremental method and system for identifying specific radiation sources based on knowledge calibration. Its purpose is to solve the problem of declining recognition performance caused by the degradation of the stability of old category features during long-term incremental learning.

[0007] The incremental method for identifying specific radiation sources based on knowledge calibration proposed in this invention includes the following steps: S1: Collect initial radiation source individual signals and divide their IQ sequences into training set, validation set and test set; S2: Construct a radiation source identification model based on a deep convolutional network and a linear classifier, and train the initial weights of the radiation source identification model using the training set obtained in S1; S3: Based on the principle of class balance, the memory capacity is evenly distributed among the radiation source categories; during each incremental learning, representative samples are selected based on the cosine similarity with the category prototype to supplement and update the memory. S4: Collect signals from newly added radiation source individuals and divide their IQ sequences into training, validation, and test sets; S5: Update the radiation source identification model, expand the classifier to adapt it to all radiation source identification, and update the model weights through a knowledge calibration mechanism by adding the IQ sequence of individual radiation source signals and the IQ sequence of memory samples. S6: Detect radiation source signals online using an updated individual radiation source identification model; S7: Detect whether there is a new individual radiation source. If yes, repeat S3-S7; otherwise, the identification is complete.

[0008] Furthermore, a preferred solution is provided: In S1, the IQ sequence of the initial radiation source individual signal is divided into a training set, a validation set, and a test set in a ratio of 0.8:0.1:0.1.

[0009] Furthermore, a preferred embodiment is provided: S2 includes: S21: A radiation source identification model is constructed based on a deep convolutional network and a linear classifier, represented as follows: , in Indicates feature extractor, Represents a classifier; S22: Augment the training set of IQ sequence signals from the initial radiation source individuals and input it into a deep convolutional network to extract signal features. ; S23: Input the signal features into a linear classifier for classification and define the loss function; S24: Calculate the importance of each channel parameter in the convolutional layer for radiation source classification using the model's gradient.

[0010] Furthermore, a preferred embodiment is provided: S3 includes: S31: Based on the total number of currently learned categories Fixed-capacity memory Divide equally into A subset; S32: For each category Calculate the average mapping vector of all its samples in the feature space; S33: Based on cosine similarity metric, from category The sample with the highest similarity to the average mapping vector is selected to construct a memory bank sample set that is representative of the category; S34: Aggregate the sample sets in the memory bank into a memory bank for knowledge replay in the subsequent incremental learning phase; S35: During the incremental phase, new task examples are gradually added to the memory.

[0011] Furthermore, a preferred embodiment is provided: S5 includes: S51: Constructing a new radiation source identification model And initialized using the weights of the radiation source identification model; S52: Dynamically expand the weight matrix in the linear classifier to adapt it to the identification of all radiation sources; S53: Perform data augmentation on the IQ sequence training set and memory samples of the newly added radiation source individual signals, and simultaneously input them into the new radiation source identification model. Perform elastic knowledge calibration through the importance weights generated during the training of the radiation source identification model, and define the loss function.

[0012] Furthermore, a preferred embodiment is provided: S5 further includes: Construct positive and negative sample pairs. The positive sample pairs are the same radiation source signals sampled from the memory sample library. The negative sample pairs are different radiation source signals sampled from the memory sample library and the newly added radiation source training set, respectively. This improves the intra-class aggregation degree of old radiation source features and the inter-class separation degree of new and old radiation source features, and performs comparative knowledge calibration.

[0013] Furthermore, a preferred solution is provided: In S4, the IQ sequence of the newly added radiation source individual signal is divided into a training set, a validation set, and a test set in a ratio of 0.8:0.1:0.1.

[0014] This invention also proposes an incremental identification system for specific radiation sources based on knowledge calibration. The system is implemented based on an incremental identification method for specific radiation sources based on knowledge calibration, as described in any one or more of the above-described schemes. The system includes: The first acquisition module is used to acquire the initial radiation source individual signals and divide their IQ sequences into training set, validation set and test set; Model training module: used to build a radiation source identification model based on a deep convolutional network and a linear classifier, and to train the initial weights of the radiation source identification model using the training set obtained in S1; Sample selection module: Based on the principle of class balance, the memory capacity is evenly distributed among the radiation source categories; during each incremental learning, representative samples are selected based on the cosine similarity with the category prototype to supplement and update the memory. The second acquisition module is used to acquire signals from newly added radiation source individuals and divide their IQ sequences into training, validation, and test sets. Model update module: used to update the radiation source identification model, expand the classifier to adapt it to all radiation source identification, and update the model weights through a knowledge calibration mechanism using the IQ sequence of the individual radiation source signal and the IQ sequence of the memory sample. Signal recognition module: used to detect radiation source signals online using an updated individual radiation source recognition model; Signal detection module: Used to detect the presence of new radiation source individuals. If so, the steps from sample screening module to signal detection module are repeated; otherwise, the identification is completed.

[0015] The present invention also proposes a computer device, the computer device including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the incremental specific radiation source individual identification method based on knowledge calibration described in any one or more of the above-described schemes.

[0016] The present invention also proposes a computer-readable storage medium for storing a computer program that executes a knowledge-calibrated incremental method for identifying specific radiation sources, as described in any one or more of the above-described schemes.

[0017] Compared with the prior art, the advantages of the present invention are: 1. This invention dynamically evaluates the importance of network parameters for the identification of old types of radiation sources based on gradient sensitivity analysis, regulates the parameter evolution during the interaction between new and old knowledge, and effectively alleviates feature representation drift through elastic knowledge calibration; 2. This invention constructs positive and negative sample pairs and combines a dual-branch contrastive loss function to simultaneously constrain intra-class aggregation and inter-class separation, avoiding confusion and conflict between the feature spaces of the new and old classes. Through contrastive knowledge calibration, it effectively suppresses topological distortion of the feature spaces of the new and old classes. 3. The core of this invention lies in the design of a specific radiation source individual identification framework for long incremental scenarios, which supports full-cycle application from initial deployment to multi-stage expansion, improves identification accuracy in incremental scenarios, and has strong engineering feasibility.

[0018] This invention is applicable to scenarios such as low-altitude security, airspace surveillance, and spectrum gaming. Attached Figure Description

[0019] Figure 1 This is a flowchart of the incremental method for identifying specific radiation sources based on knowledge calibration, as described in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the technology described in a specific embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the incremental recognition accuracy according to a specific embodiment two of the present invention; Figure 4 This is a schematic diagram of the incremental recognition forgetting rate as described in Embodiment 2 of the present invention. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0021] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0023] Implementation Method 1: Reference Figure 1 , Figure 2 This implementation method is described below.

[0024] An incremental method for identifying specific radiation sources based on knowledge calibration includes the following steps: S1: Acquire the initial radiation source individual signal and generate the IQ sequence of the initial radiation source individual signal. The dataset was divided into training, validation, and test sets in a ratio of 0.8:0.1:0.1. S2: Construct a radiation source identification model based on a deep convolutional network and a linear classifier, and train the initial weights of the radiation source identification model using the training set obtained in S1; S21: A radiation source identification model is constructed based on a deep convolutional network and a linear classifier, represented as follows: (1) in Indicates feature extractor, Represents a classifier; S22: Augment the training set of IQ sequence signals from the initial radiation source individuals and input it into a deep convolutional network to extract signal features. ; S23: Input the signal features into a linear classifier for classification, and define the loss function as follows: (2) in Represents the logarithmic function. Represents an exponential function. This represents the truncation function. This indicates the number of samples in the training batch. This indicates the first training batch. The true label of each sample This indicates the first training batch. One sample, This is the weight matrix. Indicates the first Class weight vector, This represents the weight vector corresponding to the true class. For learnable scaling parameters, Adjustable constants to facilitate inter-class separation Indicates transpose S24: Calculate the importance of each channel parameter in the convolutional layer for radiation source classification using the model's gradient, expressed as: (3) in For mathematical expectation, Describing the L2 norm, Indicates the first The first layer of convolution Feature mapping of each channel This indicates the number of convolutional layers in a deep convolutional network. Indicates the first The number of channels in each convolutional layer This represents the training dataset and stores the importance weights. , This represents the classification loss.

[0025] S3: Based on the principle of class balance, the memory capacity is evenly distributed among the radiation source categories; during each incremental learning, representative samples are selected based on the cosine similarity with the category prototype to supplement and update the memory. S31: Based on the total number of currently learned categories Fixed-capacity memory Divide equally into A subset, represented as: (4) in Indicates the number of samples in the set. Indicates rounding down; S32: For each category Calculate the sum of all its samples in the feature space The average mapping vector in the matrix is ​​represented as: (5) in Indicates the first Number of class samples Indicates the first Samples of the class; S33: Based on cosine similarity metric: (6) From category Selecting the average mapping vector The most similar A sample set of samples was used to construct a class-representative memory bank sample set. ; S34: Aggregate the sample sets of the memory into a memory, represented as: (7) Knowledge replay for subsequent incremental learning phases, among which Represents the union; S35: During the incremental phase, new task examples are gradually added to the memory.

[0026] S4: Collect signals from newly added radiation source individuals and process their IQ sequences. The dataset was divided into training, validation, and test sets in a ratio of 0.8:0.1:0.1. S5: Update the radiation source identification model, expand the classifier to adapt it to all radiation source identification, and update the model weights through a knowledge calibration mechanism by adding the IQ sequence of individual radiation source signals and the IQ sequence of memory samples. S51: Constructing a new radiation source identification model And adopt the radiation source identification model described in S2. The weights are initialized; S52: Weight Matrix in Dynamically Expanded Linear Classifiers This makes it adaptable to the identification of all radiation sources; S53: Add the IQ sequence training set and memory samples of the newly added individual radiation source signals. Perform data augmentation and simultaneously input it into the new radiation source identification model. In the process of training the radiation source identification model, importance weights are generated. Perform elastic knowledge calibration and define a loss function, expressed as: (8) in Denotes the Frobenius norm. and They represent the first and second digits in the old and new models, respectively. The first layer of convolution Feature mapping of each channel; S54: Construct positive and negative sample pairs, wherein the positive sample pairs are signals from the same radiation source sampled from the memory sample library; Construct the set of positive sample pairs: (9) in For reference sample, Positive samples of the same type This represents the total number of categories already learned. The negative sample pairs are different radiation source signals sampled from the memory sample library and the newly added radiation source training set, respectively, and are represented as follows: (10) in For negative samples of different classes, This indicates the total number of newly added radiation sources.

[0027] To improve the intra-class aggregation of old radiation source features and the inter-class separation of old and new radiation source features, comparative knowledge calibration is performed, and a loss function is defined as follows: (11) in This is the cosine similarity calculation function. This is the temperature coefficient.

[0028] By using formulas (2), (8), and (11) for joint optimization, the total loss function is obtained, which is expressed as: (12) in It is an adjustable hyperparameter.

[0029] The importance weights are recalculated using formula (3). This will be used as the weight for the next incremental update.

[0030] S6: Detect radiation source signals online using an updated specific radiation source individual identification model; S7: Detect whether there is a new individual radiation source. If yes, repeat S3-S7; otherwise, the identification is complete.

[0031] Implementation Method Two: Reference Figure 3 , Figure 4 This implementation method is described below.

[0032] This embodiment is a further illustrative example of the incremental method for identifying specific radiation sources based on knowledge calibration described in Embodiment 1.

[0033] This implementation uses a real radio signal dataset from the ADS-B special aviation monitoring system, with a sampling frequency of 50 MHz.

[0034] This implementation selects 100 different types of long signals from the dataset, each containing 4,800 sampling points.

[0035] In this implementation, the initial number of radiation sources is set to 20 categories, and 5 new categories are added each time until the total number of radiation source categories reaches 100, with a total of 16 incremental updates.

[0036] The experimental parameters for this implementation method are set as follows: constants Set to 4, temperature coefficient Setting it to 0.1 allows for adjustment of the hyperparameter. Set to 1.0. If set to 0.05, the memory sample library can store a maximum of 2000 signal samples.

[0037] like Figure 3 Figure 4 As shown, comparing the recognition accuracy curve and the forgetting rate curve demonstrates the advantages of the proposed method over traditional methods in long-term incremental update scenarios, and breaks through some application limitations of traditional incremental learning methods.

[0038] Implementation Method 3: A knowledge-calibrated incremental individual identification system for specific radiation sources is provided. The system is implemented based on a knowledge-calibrated incremental individual identification method for specific radiation sources as described in Embodiment 1. The system includes: The first acquisition module is used to acquire the initial radiation source individual signals and divide their IQ sequences into training set, validation set and test set; Model training module: used to build a radiation source identification model based on a deep convolutional network and a linear classifier, and to train the initial weights of the radiation source identification model using the training set obtained in S1; Sample selection module: Based on the principle of class balance, the memory capacity is evenly distributed among the radiation source categories; during each incremental learning, representative samples are selected based on the cosine similarity with the category prototype to supplement and update the memory. The second acquisition module is used to acquire signals from newly added radiation source individuals and divide their IQ sequences into training, validation, and test sets. Model update module: used to update the radiation source identification model, expand the classifier to adapt it to all radiation source identification, and update the model weights through a knowledge calibration mechanism using the IQ sequence of the individual radiation source signal and the IQ sequence of the memory sample. Signal recognition module: used to detect radiation source signals online using an updated individual radiation source recognition model; Signal detection module: Used to detect the presence of new radiation source individuals. If so, the steps from sample screening module to signal detection module are repeated; otherwise, the identification is completed.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A knowledge-based calibration incremental specific radiation source individual identification method, characterized in that, The method comprises the following steps: S1: collecting initial radiation source individual signals, and dividing IQ sequences of the initial radiation source individual signals into a training set, a validation set and a test set; S2: constructing a radiation source identification model based on a deep convolutional network and a linear classifier, and training initial weights of the radiation source identification model using the training set obtained in S1; S3: based on a class balance principle, dividing a memory bank capacity among the radiation source classes; in each incremental learning, screening representative samples according to cosine similarity with a class prototype, and supplementing and updating the memory bank; S4: collecting new radiation source individual signals, and dividing IQ sequences of the new radiation source individual signals into a training set, a validation set and a test set; S5: updating the radiation source identification model, and expanding the classifier to adapt to identification of all radiation sources, and updating model weights through a knowledge calibration mechanism using IQ sequences of the new radiation source individual signals and memory sample IQ sequences; S6: using the updated radiation source individual identification model to detect radiation source signals online; S7: detecting whether there is a new radiation source individual, and if yes, repeating S3-S7, and if no, completing identification.

2. The knowledge-based calibration incremental specific radiation source individual identification method according to claim 1, characterized in that, In S1, the IQ sequences of the initial radiation source individual signals are divided into the training set, the validation set and the test set according to a ratio of 0.8:0.1:0.

1.

3. The knowledge-based calibration incremental specific radiation source individual identification method according to claim 1, characterized in that, S2 comprises: S21: constructing a radiation source identification model based on a deep convolutional network and a linear classifier, and representing the radiation source identification model as: ， wherein denotes a feature extractor, denotes a classifier; S22: Data augmentation is performed on the IQ sequence training set of the initial radiation source individual signal, and the IQ sequence training set is input into a deep convolutional network to extract signal features ; S23: inputting signal features into the linear classifier for classification, and defining a loss function; S24: calculating, through a gradient of the model, an importance degree of each channel parameter in a convolutional layer to radiation source classification.

4. The knowledge-based calibration incremental specific radiation source individual identification method according to claim 1, characterized in that, S3 comprises: S31: According to the current total number of learned classes a memory bank of fixed capacity is equally divided into subsets; S32: For each class , compute the average mapping vector of all samples in the feature space; S33: Based on cosine similarity metric, from category The sample with the highest similarity to the average mapping vector is selected to construct a memory bank sample set that is representative of the category; S34: aggregating memory bank sample sets into a memory bank for knowledge playback in a subsequent incremental learning stage; S35: gradually supplementing new task samples to the memory bank in an incremental stage.

5. The knowledge-based calibration incremental specific radiation source individual identification method according to claim 1, characterized in that, S5 comprises: S51: Construct a new radiation source identification model and the weights of the radiation source identification model are initialized; S52: dynamically expanding a weight matrix in the linear classifier to adapt to identification of all radiation sources; S53: performing data enhancement on the IQ sequence training set of the new radiation source individual signals and memory samples, and simultaneously inputting the IQ sequence training set and the memory samples into a new radiation source identification model, and performing elastic knowledge calibration through importance weights generated in radiation source identification model training, and defining a loss function.

6. The knowledge-based calibration incremental specific radiation source individual identification method according to claim 5, characterized in that, S5 further comprises: constructing positive and negative sample pairs, the positive sample pairs being same radiation source signals sampled from the memory sample bank, and the negative sample pairs being different radiation source signals respectively sampled from the memory sample bank and the new radiation source training set, improving in-class aggregation degree of old class radiation source features and inter-class separation degree of new and old radiation source features, and performing comparative knowledge calibration.

7. The knowledge-based calibration incremental specific radiation source individual identification method according to claim 1, characterized in that, In S4, the IQ sequences of the new radiation source individual signals are divided into the training set, the validation set and the test set according to a ratio of 0.8:0.1:0.

1.

8. A knowledge-based calibration incremental specific radiation source individual identification system, characterized in that, The system is implemented based on the incremental specific radiation source individual identification method based on knowledge calibration according to any one of claims 1-7, and the system comprises: a first acquisition module configured to collect initial radiation source individual signals, and divide IQ sequences of the initial radiation source individual signals into a training set, a validation set and a test set; The model training module is configured to construct a radiation source identification model based on a deep convolutional network and a linear classifier, and train initial weights of the radiation source identification model by using the training set obtained by the S1; The sample screening module is configured to divide the memory capacity among the radiation source categories based on a class balance principle, and screen representative samples according to the cosine similarity with the category prototype to supplement and update the memory at each incremental learning. The second acquisition module is configured to acquire the signal of the new radiation source individual, and divide the IQ sequence of the signal into a training set, a validation set and a test set. The model updating module is configured to update the radiation source identification model, and expand the classifier to adapt to identification of all radiation sources, and update the model weights by using the IQ sequence of the new radiation source individual signal and the memory sample IQ sequence through a knowledge calibration mechanism. The signal recognition module is configured to detect the radiation source signal online by using the updated radiation source individual identification model. The signal detection module is configured to detect whether there is a new radiation source individual, and if yes, repeat the steps of the sample screening module to the signal detection module, and if no, complete the identification.

9. A computer device, comprising: The computer device comprises a memory and a processor, and the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the incremental specific radiation source individual identification method based on knowledge calibration according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium is configured to store a computer program, and the computer program executes the incremental specific radiation source individual identification method based on knowledge calibration according to any one of claims 1-7.

Images