A feature coding method for individual identification of radiation sources

Abstract

The present invention discloses a feature coding method for individual identification of radiation sources. Based on the Gram angle and / or difference field transformation, the electromagnetic signals of the radiation source sampled in two channels, IQ, are encoded and processed. The real part, imaginary part and amplitude feature sequence of the signal are subjected to Gram angle and / or difference field transformation respectively to obtain a matrix containing time domain and frequency domain correlation features between sampling points. The obtained coding features can be used as inputs of a deep neural network, and the individual identification of radiation sources can be realized by training the obtained model. The present invention provides ideas for the extraction of time domain and frequency domain features of various electromagnetic signals such as radar and communication, and provides a data preprocessing solution for individual identification tasks for various types of radiation sources. Compared with traditional coding methods, the present invention has better feature extraction effects on the IQ two-channel collected signals of the radiation source, and has the characteristics of high recognition accuracy and a wide range of signal applicability, which is of great significance for individual identification applications of radiation sources such as electronic countermeasures.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and in particular relates to a feature coding method for individual identification of radiation sources. Background Art

[0002] Electromagnetic signal recognition aims to obtain useful information from the emitted electromagnetic signals to provide support for applications such as situation assessment. It requires feature analysis of the captured signals to obtain various characteristics of the signals, and then obtain the signal fingerprint characteristics of the radiation source to perform individual radiation source target identification tasks.

[0003] Conventional electronic signal feature extraction, such as simple features like modulation parameters like carrier frequency, bandwidth, and symbol rate, is no longer sufficient to identify individual radiation sources and cannot meet the demands of modern applications. Advances in computer technology, fusion algorithms, and electronic information science have led to rapid developments in signal feature analysis and extraction, enabling the full mining and utilization of information. Fingerprint feature extraction and application of subtle individual features of radiation source targets, which can reveal their unique properties, hold significant research and application value.

[0004] Existing methods for extracting and identifying radiation source signals have limited systematic feature extraction methods, and most fail to perform effective encoding preprocessing on the signal before feature extraction. "A Quasi-Incremental Radiation Source Individual Identification Method Based on Knowledge Distillation Mechanism" (CN114492745A) and "A Radiation Source Individual Identification Method Based on Deep Residual Shrinkage Network" (CN114091545A) propose methods for identifying radiation source IQ signals using deep residual neural networks. However, these methods directly utilize waveforms of both IQ and Q signals, requiring high waveform resolution for feature extraction and lacking preliminary feature extraction of the radiation source signal. "A Radiation Source Individual Identification Method in a Small Sample Scenarios" (CN114492604A) proposes inputting the time-frequency features of signals extracted using different time-frequency analysis methods into a neural network for training. After performing a preliminary analysis of the radiation source signal, the neural network is used to extract features. However, the applicable scenarios are limited by conditions such as the sample size and the length of the radiation source signal, resulting in weak adaptability to different scenarios. Summary of the Invention

[0005] To address these issues, the present invention discloses a feature encoding method for identifying individual radiation sources. This method provides a method for extracting time-domain and frequency-domain features from various electromagnetic signals, such as radar and communications signals, and offers a data preprocessing solution for identifying individual radiation sources. Compared to traditional encoding methods, this method achieves superior feature extraction for both I / Q signals collected from radiation sources, boasting high recognition accuracy and a wide range of signal applicability. This method is of great significance for applications in identifying individual radiation sources, such as electronic countermeasures.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] (1) Transform the radiation source signal of the IQ two-way sampling. Obtain the IQ sampling one-dimensional signal of the I-way and record it as Re, the Q-way one-dimensional signal as Im, and the amplitude (square sum root) of the IQ two-way signal as Am; for Re, Im and Am, common time-frequency analysis methods such as short-time Fourier transform, wavelet transform, and Hilbert-Huang transform can be used. When performing short-time Fourier transform and wavelet transform, calculate the signal power spectrum density at different moments, take the maximum value of the frequency distribution of the two-dimensional transform result in each time window, and obtain a one-dimensional time-frequency feature sequence. When performing Hilbert-Huang transform, the intrinsic modal component decomposed by the empirical mode decomposition method is then subjected to Hilbert transform, and finally the instantaneous frequency and instantaneous amplitude of the signal are obtained, which are the time-frequency features after decomposition. The number of one-dimensional feature sequences obtained is equal to the number of empirical mode decompositions.

[0008] In this step, it is necessary to screen the time-frequency analysis method in a targeted manner according to the analysis requirements of the signal and the characteristics of the signal itself.

[0009] (2) For the time domain and time-frequency domain feature results obtained in step (1), perform Gram angular field transform (GASF / GADF transform) to transform the original time series and time-frequency domain feature sequence of length n into an n×n feature coding matrix.

[0010] (3) The I-path, Q-path, amplitude and time-frequency domain feature coding matrix of the radiation source IQ signal in the cascade step (2) are selected to form a matrix array of size k×n×n, which is applied to the individual recognition task based on deep neural network.

[0011] For different received signals, the feature matrices to be concatenated are selected in a targeted manner. Signals with a large number of sampling points are more likely to select time-frequency domain features for concatenation, while signals with a large number of sampling points are more likely to select time-domain features for concatenation. When it is necessary to simultaneously concatenate time-domain and time-frequency domain features with different feature sequence lengths, the encoding matrix size with the largest feature dimension is selected as the benchmark, and the smaller encoding matrices are padded with zeros to make them the same size as the benchmark. When the number of samples is sufficient, by transforming the combination of feature concatenations, the features with the highest discrimination and the best recognition effect are selected as the encoding method for the final signal, which serves as the feature required for individual recognition.

[0012] As a specification and improvement of the present invention, step (1) proposes a solution for limiting the range and different signal conditions at the sampling point, thereby increasing the electromagnetic signal range applicable to the present invention.

[0013] The applied radiation source IQ signal is a digital signal data with discrete amplitude and time. The data is divided into two channels, each modulated by a carrier wave, and the two carrier waves are orthogonal to each other. The two channels have the same frequency and a 90-degree phase difference. After being modulated, the I and Q signals are transmitted simultaneously. The applied radiation source IQ signal has certain requirements regarding the number of sampling points. Because this method uses Gram angle sum / difference field transformation, a time series signal with n sampling points, after encoding preprocessing, produces an n×n matrix. Therefore, the required number of radiation source IQ signal sampling points is less than 1E3. For radiation source IQ signals with long sampling times and more than 1E3 sampling points, envelope detection can be used to further obtain the peak characteristics of the signal, and this characteristic can be applied in conjunction with the processing method for subsequent processing. Alternatively, direct time-frequency domain analysis can be performed by controlling the window size to match the output length to the number of sampling points in this step. For sequence data of length n, the converted n×n matrix can be approximated using segmented aggregation to reduce the sequence length before performing the conversion. This involves segmenting the sequence and then compressing the subsequences within each segment into a single value through averaging. The applied radiation source IQ signal has no requirements for absolute time parameters such as sampling rate; this method only considers the relative time relationship between different sampling points of the radiation source IQ signal during processing.

[0014] As an improvement of the present invention, the time-frequency domain analysis method in step (1) needs to process the analysis results in the time-frequency-amplitude direction to match the Gram angle field transform format requirements and visualization requirements in step (2). The specific processing method is as follows:

[0015] In step (1), after performing time-frequency analysis using short-time Fourier transform, wavelet transform and its optimization method, a three-dimensional matrix result in the format of TFA is obtained. The complex result obtained is modulo to obtain the frequency-amplitude characteristics at different sampling points; for each different sampling point, the frequency with the highest amplitude at the point is taken as the frequency characteristic of this point, or the average frequency characteristic of this sampling point is obtained by weighting.

[0016] When using empirical mode decomposition, Hilbert transform, and its optimization method for time-frequency analysis in step (1), it is necessary to consider the different characteristics of similar signals, which lead to differences in the number of decompositions. Therefore, the number of decompositions of the signal with the least number of decompositions among all signals that require coding preprocessing is taken as the unified number of decompositions. After obtaining the transformed results, the multi-level decomposition results of a single signal are all used as a time-frequency feature of the signal to proceed to the next step.

[0017] As an improvement of the present invention, the radiation source individual identification method based on feature coding preprocessing in step (2) uses Gram angle sum and difference field transformation, and the specific processing method is as follows:

[0018] Scaling is performed to scale the data range obtained from step (1) to [-1, 1] or [0, 1]; the scaled sequence data is converted to a polar coordinate system, that is, the numerical value is regarded as the cosine value of the angle and the timestamp is regarded as the radius. If the data scaling range is [-1, 1], the converted angle range is [0, π]; if the scaling range is [0, 1], the converted angle range is

[0019] Subsequently, an inner product-like operation is performed in the rectangular coordinate system, and the scaled one-dimensional sequence data is converted from the rectangular coordinate system to the polar coordinate system. Then, the temporal correlation of different time points is identified by considering the angles (sum and difference) between different points.

[0020] The resulting feature-encoded image has the following characteristics:

[0021] (1) The encoded image is a square matrix, distributed in chronological order from the upper left corner to the lower right corner

[0022] (2) The image is a heat image, and the grayscale size represents the eigenvalue size of the point

[0023] (3) Each point in the image contains the amplitude-time related information and frequency-time related information contained in the sequence at that time point after time domain and time-frequency domain analysis

[0024] (4) The matrix slices of the image matrix also have features that are distributed in chronological order from the upper left corner to the lower right corner, and the image is not sparse. In applications such as individual recognition tasks using deep residual neural networks, the proposed method has strong information fusion and high robustness, which has been verified in experiments.

[0025] As an improvement of the present invention, the transformations performed in step (2) are all transformations of the single result obtained in step (1) for a single signal. Therefore, in order to obtain the complete preprocessing coding result of a single signal, the above transformation results need to be cascaded. When it is necessary to simultaneously cascade the time domain and time-frequency domain features with different feature sequence lengths, the encoding matrix size with the largest feature dimension is selected as the benchmark, and the smaller encoding matrix is ​​padded with zeros to make its size the same as the benchmark, providing a standard format for direct use for individual recognition using a deep residual neural network in step (3).

[0026] The beneficial effects of the present invention are:

[0027] The proposed coding preprocessing method for radiation source IQ signals provides a method for extracting features from radiation source IQ signals in both the time and time-frequency domains, offering a data preprocessing solution for artificial intelligence projects targeting various signals. Compared to traditional coding methods, this method achieves superior feature extraction for both IQ and Q signals, and boasts excellent visualization, strong interpretability, and a wide range of signal applicability. This method is of great significance for applications in individual radiation source identification and modern electronic countermeasures. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of the present invention;

[0029] Figure 2 This is a flow chart of time-frequency analysis after preliminary decomposition according to the present invention;

[0030] Figure 3 This is the feature coding flow chart of the present invention;

[0031] Figure 4 The Im sequence image after preliminary decomposition in the example of the present invention;

[0032] Figure 5 The Re sequence image after preliminary decomposition in the example of the present invention;

[0033] Figure 6 The Am sequence image after preliminary decomposition in the example of the present invention;

[0034] Figure 7 The image obtained by encoding the preprocessed Im sequence after the initial decomposition in the example of the present invention;

[0035] Figure 8 The image is the result of Re sequence encoding preprocessing after preliminary decomposition in the example of the present invention;

[0036] Figure 9 The image is the result of the Am sequence coding preprocessing after the preliminary decomposition in the example of the present invention;

[0037] Figure 10 This is the result diagram of the example where the encoding method of the present invention is not adopted;

[0038] Figure 11 This is a diagram showing the result of using the encoding method of the present invention in an example. DETAILED DESCRIPTION

[0039] The present invention will be further explained below with reference to the accompanying drawings and a specific implementation of a preamble signal of a civil aviation ADS-B signal. It should be understood that the following specific implementation is only used to illustrate the present invention and is not intended to limit the scope of the present invention.

[0040] like Figure 1As shown, the coding preprocessing method of the radiation source IQ signal described in the present invention can be roughly divided into the following steps: preliminary decomposition, time-frequency analysis (optional), feature coding, and application in artificial intelligence engineering:

[0041] Initial decomposition: Decompose the two signal paths. Take the radiation source IQ signal (I: in-phase) as the real part Re, take the Q: quadrature (orthogonal) as the imaginary part Im, and take the geometric mean of the two radiation source IQ signals as Am. Obtain the Re, Im, and Am time series.

[0042] Time-frequency analysis: To further obtain the time-frequency domain characteristics of the IQ signal of the radiation source, the acquired signal time series is subjected to time-frequency analysis. Short-time Fourier transform can be applied to calculate the signal power spectrum at different moments, and the maximum value of the frequency distribution of the two-dimensional analysis result in each time window is taken to obtain a one-dimensional time-frequency feature sequence; empirical mode decomposition and Hilbert transform are applied to obtain the signal decomposition results based on the time domain characteristics of the data itself, the time-frequency characteristics after decomposition, and the results obtained with the appropriate number of empirical mode decompositions. Other common time-frequency joint analysis methods that can be applied in this step include wavelet transform and other common methods.

[0043] After performing time-frequency analysis using short-time Fourier transform, wavelet transform and its optimization method, a three-dimensional matrix result in the TFA format is obtained. The complex number result is modulo the obtained frequency-amplitude characteristics at different sampling points; for each different sampling point, the frequency with the highest amplitude at the point is taken as the frequency characteristic of this point, or the average frequency characteristic of this sampling point is obtained by weighted method.

[0044] When using empirical mode decomposition, Hilbert transform, and its optimization methods for time-frequency analysis, it's important to consider the varying characteristics of similar signals, which can lead to differences in the number of decompositions possible. Therefore, the uniform number of decompositions is determined by taking the signal with the lowest number of decompositions among all signals requiring coding preprocessing. After obtaining the transformed results, the multi-level decomposition results for each signal are used as a time-frequency feature for the next step.

[0045] The specific process is as follows Figure 2 shown.

[0046] Feature coding preprocessing: For the different results obtained through time domain and frequency domain analysis in step (1), Gram angular field transform (GASF / GADF transform) is performed to obtain the feature coding matrix.

[0047] Scaling is performed to scale the data range obtained from step (1) to [-1, 1] or [0, 1]; the scaled sequence data is converted to a polar coordinate system, that is, the numerical value is regarded as the cosine value of the angle and the timestamp is regarded as the radius. If the data scaling range is [-1, 1], the converted angle range is [0, π]; if the scaling range is [0, 1], the converted angle range is Subsequently, an inner product-like operation is performed in the rectangular coordinate system, and the scaled one-dimensional sequence data is converted from the rectangular coordinate system to the polar coordinate system. Then, the temporal correlation of different time points is identified by considering the angles (sum and difference) between different points.

[0048] The encoding preprocessing process is as follows Figure 3 shown.

[0049] The preliminary decomposition method of the present invention decomposes the two signal paths, taking the radiation source IQ signal (I: in-phase) as the real part Re, the Q (quadrature) as the imaginary part Im, and the geometric mean of the two radiation source IQ signals as Am. The time series of Im, Re, and Am is obtained. Figure 4 、 Figure 5 、 Figure 6 The Im, Re, and Am results after preliminary decomposition of the IQ signal of the example radiation source are shown respectively.

[0050] The time-frequency analysis technique described in the present invention employs a short-time Fourier transform to calculate the signal power spectrum at each moment. The maximum value of the frequency distribution of the two-dimensional analysis result within each time window is taken to obtain a one-dimensional time-frequency feature sequence. Empirical mode decomposition and the Hilbert transform are then applied to obtain the signal decomposition results based on the time-domain characteristics of the data itself, the decomposed time-frequency features, and the appropriate number of empirical mode decompositions. Other common time-frequency joint analysis methods applicable to this step include wavelet transforms and other common methods.

[0051] In step (1), after performing time-frequency analysis using short-time Fourier transform, wavelet transform and its optimization method, a three-dimensional matrix result in the format of TFA is obtained. The complex result obtained is modulo to obtain the frequency-amplitude characteristics at different sampling points; for each different sampling point, the frequency with the highest amplitude at the point is taken as the frequency characteristic of this point, or the average frequency characteristic of this sampling point is obtained by weighting.

[0052] When using empirical mode decomposition, Hilbert transform, and its optimization method for time-frequency analysis in step (1), it is necessary to consider the different characteristics of similar signals, which lead to differences in the number of decompositions. Therefore, the number of decompositions of the signal with the least number of decompositions among all signals that require coding preprocessing is taken as the unified number of decompositions. After obtaining the transformed results, the multi-level decomposition results of a single signal are all used as a time-frequency feature of the signal for the next step.

[0053] The feature coding preprocessing of the present invention: for the different results obtained by time domain and frequency domain analysis in step (1), perform Gram angle sum / difference field transformation to obtain a feature coding matrix. Scaling is performed to scale the data range obtained by step (1) analysis to [-1, 1] or [0, 1]; the scaled sequence data is converted to a polar coordinate system, that is, the numerical value is regarded as the cosine value of the angle and the timestamp is regarded as the radius. If the data scaling range is [-1, 1], the converted angle range is [0, π]; if the scaling range is [0, 1], the converted angle range is Subsequently, an inner product-like operation is performed in the rectangular coordinate system, and the scaled one-dimensional sequence data is converted from the rectangular coordinate system to the polar coordinate system. Then, the temporal correlation of different time points is identified by considering the angles (sum and difference) between different points.

[0054] In this example, the Im, Re, Am sequence obtained by the preliminary decomposition is subjected to the coding preprocessing adopted by the present invention to obtain an image such as Figure 7 、 Figure 8 、 Figure 9 shown.

[0055] The resulting feature-encoded image has the following characteristics:

[0056] (1) The encoded image is a square matrix, distributed in chronological order from the upper left corner to the lower right corner

[0057] (2) The image is a thermal image, and the grayscale depth represents the value of the point

[0058] (3) Each point in the image contains the amplitude-time related information and frequency-time related information contained in the sequence at that time point after time domain and time-frequency domain analysis

[0059] (4) The matrix slices of the image matrix also have features that are distributed in chronological order from the upper left corner to the lower right corner, and the image is not sparse. In applications such as individual recognition tasks using deep residual neural networks, the proposed method has strong information fusion and high robustness, which has been verified in experiments.

[0060] The individual identification task described in the present invention is to cascade the results of the above encoding preprocessing, train them through a convolutional neural network, and obtain the radiation source individual identification results. The feature encoding preprocessing results are used as training samples and sent to the deep neural network for further feature extraction to finally obtain the individual identification results. After training without the encoding method of the present invention, the signal classification result confusion matrix is ​​obtained as follows Figure 10 As shown, the accuracy is 89.7%. After using the feature encoding method adopted by the present invention, the signal classification result confusion matrix is ​​obtained as follows Figure 11 As shown in the figure, the accuracy rate is 93.2%, which significantly improves the recognition accuracy under the condition of feature encoding preprocessing.

Claims

1. A feature coding method for individual identification of radiation sources, characterized in that: The following steps are involved: (1) Transform the radiation source signals sampled in both I and Q channels; Obtain IQ sampling, record the I-channel one-dimensional signal as Re, the Q-channel one-dimensional signal as Im, and the amplitude of the two-channel IQ signal as Am; for Re, Im and Am, use common time-frequency analysis methods such as short-time Fourier transform (STFT), wavelet transform (WT), and Hilbert-Huang transform (HHT); when performing short-time Fourier transform and wavelet transform, calculate the signal power spectral density at different times, take the maximum value of the frequency distribution of the two-dimensional transform result in each time window, and obtain a one-dimensional time-frequency feature sequence; When performing the Hilbert-Huang transform, the intrinsic modal components (IMFs) decomposed by the empirical mode decomposition (EMD) are then Hilbert transformed to obtain the instantaneous frequency and instantaneous amplitude of the signal, which are the decomposed time-frequency characteristics. The number of one-dimensional feature sequences obtained is equal to the number of empirical mode decompositions. In this step, it is necessary to select the time-frequency analysis method in a targeted manner based on the analysis requirements and characteristics of the signal itself. (2) For the time domain and time-frequency domain feature results obtained in step (1), perform Gram angle sum / difference field GASF / GADF transformation; convert the original time series and time-frequency domain feature sequence with a length of n into an n×n feature coding matrix; (3) Cascading the I-path, Q-path, amplitude and time-frequency domain feature encoding matrix of the radiation source IQ signal in step (2), selecting k features with high discrimination, forming a matrix array of size k×n×n, and applying it to the individual recognition task based on deep neural network; For different signals received, the cascaded feature matrix is ​​selected in a targeted manner; For signals with more sampling points, more time-frequency domain features are selected for cascading, and vice versa, time-domain features are cascaded; When it is necessary to simultaneously cascade time domain and time-frequency domain features with different feature sequence lengths, the encoding matrix size with the largest feature dimension is selected as the benchmark, and the smaller encoding matrix is ​​padded with zeros to make its size the same as the benchmark; when the number of samples is sufficient, by transforming the combination of feature cascades, the features with the highest discrimination and the best recognition effect are selected as the encoding method of the final signal, which is the feature required for individual identification.

2. The feature encoding method for radiation source individual identification according to claim 1, characterized in that: Based on the Gram angular field GASF / GADF transform, the time series signal with n sampling points is recorded as vector V, and V×V T Then, a matrix of size n×n is obtained, and the required number of sampling points of the radiation source IQ signal is below the order of 1E3; for the radiation source IQ signal with a long sampling time and a number of sampling points higher than 1E3, the peak characteristics of the signal are further obtained by the envelope detection method, the length of the obtained sequence is reduced, and the characteristics and processing method are applied for subsequent processing; in addition, the time-frequency domain analysis method in step (1) can be directly used to control the output length to match the order of magnitude of the number of sampling points in this step by analyzing the size of the window and the window displacement length; for the sequence data of length n, the converted n×n matrix uses segmented aggregation approximation to first reduce the sequence length and then convert it; that is, the sequence is segmented, and then the subsequences in each segment are compressed into a value by averaging; the applied radiation source IQ signal has no requirements on the absolute time parameters of the sampling rate, and only the relative time relationship between different sampling points of the radiation source IQ signal is considered during processing.

3. The feature encoding method for radiation source individual identification according to claim 1, characterized in that: When the Gram angular field transform is used in step (2), the following transformation requirements are met: Scaling is performed to scale the data range obtained by step (1) to [-1, 1] or [0, 1]; the scaled sequence data is converted to a polar coordinate system, that is, the numerical value is regarded as the cosine value of the angle and the timestamp is regarded as the radius; if the data scaling range is [-1, 1], the converted angle range is [0, π]; if the scaling range is [0, 1], the converted angle range is After scaling, an inner product-like operation is performed in the rectangular coordinate system. The scaled one-dimensional series data is converted from the rectangular coordinate system to the polar coordinate system. Then, the angle sum GASF / angle difference GADF between different points is considered to identify the time correlation of different time points, and decide whether to perform the angle sum or angle difference.

Images