Radar signal detection method and system based on multi-task learning of time-frequency domain feature fusion network

Abstract

The present invention discloses a radar signal detection method and system based on multi-task learning of a time-frequency domain feature fusion network. The method includes the following steps: establishing a radar signal mathematical model and using the mathematical model to obtain various types of radar signal sequences; processing the radar signal sequences to obtain radar signal I sequences and radar signal Q sequences; labeling the radar signal I sequences and radar signal Q sequences and constructing a training data set; constructing a time-frequency domain feature fusion network and training the time-frequency domain feature fusion network using the training data set to obtain a signal detection and recognition model; obtaining a radar signal to be detected and completing radar signal detection using the signal detection and recognition model. The present invention performs combined modeling and analysis on multiple radar signals, and uses a time-frequency domain feature fusion network to extract local and global signal features from the perspective of signal I / Q, and fuses frequency domain features to reduce the impact of noise on signal recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal detection and recognition, and specifically relates to a radar signal detection method and system based on multi-task learning of time-frequency domain feature fusion network. Background Art

[0002] Radar signal detection and recognition are fundamental tasks in radar reconnaissance and serve as the foundation for subsequent radar system processing. Efficiently and accurately detecting radar signals and identifying the correct intrapulse type is crucial. Existing deep learning-based signal detection and recognition methods can be divided into two categories, each utilizing different signal input formats. Methods based on time-frequency images typically employ convolutional neural network (CNN) architectures, using established frameworks such as VGG, YOLO, GoogleNet, and ResNet. However, the high computational complexity of generating time-frequency images limits the real-time processing capabilities of these methods.

[0003] The signal detection and recognition method based on signal sequence does not require complex time-frequency image calculation at the input end of the signal. Compared with the method using time-frequency image, it has higher detection and recognition efficiency. This method usually uses recurrent neural network (RNN) and its variants. Mustafa Atahan "Deep learning for radar signal detection in electronic warfare systems" published in "IEEE Radar Conference.IEEE" (2020, 21(9): 1-6.) introduced a bidirectional long short-term memory network to realize signal detection. Mengtao Zhu et al. published "JMRPE-Net: Joint modulation recognition and parameter estimation of cognitive radar signals with a deep multitask network" in "IET Radar, Sonar Navigation" (2021, 15(11): 1508-1524) and proposed a time series-based network to complete signal recognition. These methods mainly consider the isolated detection and recognition of each individual pulse signal, without studying the relationship between continuous signals to improve the detection and recognition performance of the network.

[0004] Ziwei Zhang published "JDMR-Net: Joint Detection and Modulation Recognition Networks for Radar Signals" in "IEEE Transactions on Aerospace and Electronic Systems" (2023, 59(6): 7575-7589). The paper introduced a multi-task learning method to detect and recognize signals, and studied the relationship between continuous signals to improve the network's detection and recognition performance. However, the paper did not analyze the signal frequency domain, resulting in unsatisfactory detection and recognition results in low signal-to-noise ratio environments. Summary of the Invention

[0005] The present invention aims to solve the deficiencies of the prior art and provides the following solutions:

[0006] The radar signal detection method based on multi-task learning of time-frequency domain feature fusion network includes the following steps:

[0007] Establishing a radar signal mathematical model and using the mathematical model to obtain various types of radar signal sequences;

[0008] Processing the radar signal sequence to obtain a radar signal I sequence and a radar signal Q sequence;

[0009] Annotating the radar signal I sequence and the radar signal Q sequence, and constructing a training data set;

[0010] Constructing a time-frequency domain feature fusion network, and training the time-frequency domain feature fusion network using the training data set to obtain a signal detection and recognition model;

[0011] Acquire the radar signal to be detected, and use the signal detection and recognition model to complete radar signal detection.

[0012] Preferably, the radar signal mathematical model includes:

[0013] s[k]=x[k]+n[k],0≤k≤K

[0014] Where S[k] represents the radar signal, x[k] represents the noise-free radar signal, n[k] represents the receiver noise, k represents the sampling index, and K represents the number of samples.

[0015] Preferably, the method for obtaining the radar signal sequence includes: obtaining a radar signal using the radar signal mathematical model, and performing a Hilbert transform on the radar signal to obtain the radar signal sequence:

[0016] s′[k]=s[k]+jH(s[k])

[0017] Where S'[k] represents the radar signal sequence, j represents the imaginary unit, and H() represents the Hilbert transform.

[0018] Preferably, the radar signal I sequence is:

[0019] I[k]=Re(S′[k])

[0020] Where I[k] represents the radar signal I sequence, and Re() represents the real part;

[0021] The radar signal Q sequence is:

[0022] Q[k]=Im(S′[k])

[0023] Wherein, Q[k] represents the radar signal Q sequence, and Im() represents the imaginary part.

[0024] Preferably, the time-frequency domain feature fusion network includes: a time-domain feature extraction module, a frequency-domain feature extraction module, a cross-attention module and a detection task head;

[0025] The time domain feature extraction module is used to extract the time domain feature matrix in the training data set;

[0026] The frequency domain feature extraction module is used to extract the frequency domain feature matrix in the training data set;

[0027] The cross attention module is used to perform attention mechanism calculation and data splicing on the time domain feature matrix and the frequency domain feature matrix to obtain an output result;

[0028] The detection task head is used to perform radar signal detection based on the output result.

[0029] Preferably, the time domain feature extraction module includes: a signal sliding window cutting module and a first local feature extraction module;

[0030] The signal sliding window cutting module is used to segment the radar initial signal in the training data set to obtain n ordered sequences:

[0031]

[0032] Among them, χ represents the sequence set, X n represents the nth sequence, represents the real number field, n represents the number of ordered sequences, and W represents the sliding window length;

[0033] The first local feature extraction module adopts an improved DLSTM structure, and is used to extract the features of the sequence set χ to obtain the time domain feature matrix

[0034] Preferably, the frequency domain feature extraction module includes: a data segmentation unit, a time-frequency conversion unit and a second local feature extraction module;

[0035] The data segmentation unit is used to segment the original time series in the training data set to obtain segmented data;

[0036] The time-frequency conversion unit performs time-frequency conversion processing on the segmented data using EDFT to obtain frequency domain data:

[0037]

[0038] Where F[k] represents the spectrum, L represents the input length, T represents the output length, and j represents the imaginary unit;

[0039] The second local feature extraction module adopts an improved DLSTM structure, and the second local feature extraction module is used to extract the features of the frequency domain data to obtain the frequency domain feature matrix

[0040] Preferably, the cross attention module is composed of two layers of cross multi-head attention, each layer including two of the multi-head attentions.

[0041] The present invention also provides a radar signal detection system based on time-frequency domain feature fusion network multi-task learning, wherein the detection system applies any of the above-mentioned detection methods and comprises: a signal sequence acquisition module, a sequence processing module, a labeling module, a network construction module, and a detection module;

[0042] The signal sequence acquisition module is used to establish a radar signal mathematical model and use the mathematical model to acquire various types of radar signal sequences;

[0043] The sequence processing module is used to process the radar signal sequence to obtain a radar signal I sequence and a radar signal Q sequence;

[0044] The labeling module is used to label the radar signal I sequence and the radar signal Q sequence and construct a training data set;

[0045] The network construction module is used to construct a time-frequency domain feature fusion network, and train the time-frequency domain feature fusion network using the training data set to obtain a signal detection and recognition model;

[0046] The detection module is used to obtain the radar signal to be detected and complete the radar signal detection using the signal detection and recognition model.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] This paper addresses the low accuracy of existing radar signal detection and recognition methods under low signal-to-noise ratio conditions by designing a more intelligent radar signal detection and recognition method. By performing combined modeling and analysis on multiple radar signals, a time-frequency domain feature fusion network is employed to extract local and global signal features from the perspective of signal I / Q. Frequency domain features are then integrated to reduce the impact of noise on signal recognition. Simulation experiments demonstrate the effectiveness of the proposed time-frequency domain feature fusion network in signal detection and identification of radar intrapulse modulation types. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention;

[0051] Figure 2 Schematic diagram of the time-frequency domain feature fusion network structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] Example 1

[0055] In this embodiment, if Figure 1 As shown in FIG, the radar signal detection method based on multi-task learning of time-frequency domain feature fusion network includes the following steps:

[0056] S1. Establish a mathematical model of radar signals and use the mathematical model to obtain various types of radar signal sequences.

[0057] The radar signal mathematical model includes:

[0058] s[k]=x[k]+n[k],0≤k≤K

[0059] Where S[k] represents the radar signal, x[k] represents the noise-free radar signal, n[k] represents the receiver noise, k represents the sampling index, and K represents the number of samples. The method for obtaining a radar signal sequence includes: obtaining the radar signal using a radar signal mathematical model. In this embodiment, radar signals are simulated and generated, including linear frequency modulation signals, nonlinear frequency modulation signals, BPSK signals, FSK signals, Costas signals, Frank code signals, and P1-P4 code signals. The radar signal is Hilbert transformed to obtain the radar signal sequence:

[0060] s′[k]=s[k]+jH(s[k])

[0061] Where S'[k] represents the radar signal sequence, j represents the imaginary unit, and H() represents the Hilbert transform. The analytical correlation formula for x[k] is as follows:

[0062] x[k]=a[k]e jθ[k]

[0063] Where a[k] represents the pulse interval τ pw A non-zero constant signal envelope, usually a[k]=A, 0≤kT s ≤τ pw , where T s It means that for the time interval of the sampling frequency, the instantaneous phase θ[k] can be defined by the instantaneous frequency f[k] and the phase offset Φ[k] as:

[0064] θ[k]=2πf[k](kT s )+φ[k].

[0065] S2. Process the radar signal sequence to obtain a radar signal I sequence and a radar signal Q sequence.

[0066] Among them, the radar signal I sequence is:

[0067] I[k]=Re(S′[k])

[0068] Where I[k] represents the radar signal I sequence, Re() represents the real part; the radar signal Q sequence is:

[0069] Q[k]=Im(S′[k])

[0070] Wherein, Q[k] represents the radar signal Q sequence, and Im() represents the imaginary part.

[0071] S3. Label the radar signal I sequence and the radar signal Q sequence and construct a training dataset.

[0072] In this embodiment, several long signal sequences are generated at signal-to-noise ratios ranging from -16 to 0. Each signal sequence may contain missing pulses. The resulting signal I / Q is divided into a test set, a validation set, and a training set according to a certain ratio, and labels are set.

[0073] S4. Construct a time-frequency domain feature fusion network, use the training data set to train the time-frequency domain feature fusion network, and obtain a signal detection and recognition model.

[0074] Time-frequency domain feature fusion network Figure 2 As shown, it includes: a time domain feature extraction module, a frequency domain feature extraction module, a cross attention module and a detection task head; the time domain feature extraction module is used to extract the time domain feature matrix in the training data set; the frequency domain feature extraction module is used to extract the frequency domain feature matrix in the training data set; the cross attention module is used to perform attention mechanism calculation and data splicing on the time domain feature matrix and the frequency domain feature matrix to obtain the output result; the detection task head is used to perform radar signal detection based on the output result.

[0075] The time domain feature extraction module includes: a signal sliding window cutting module and a first local feature extraction module; the signal sliding window cutting module is used to segment the radar initial signal in the training data set to obtain n ordered sequences:

[0076]

[0077] Among them, χ represents the sequence set, X n represents the nth sequence, represents the real number domain, n represents the number of ordered sequences, and W represents the sliding window length; the first local feature extraction module adopts the improved DLSTM structure, and the first local feature extraction module is used to extract the features of the sequence set χ to obtain the time domain feature matrix

[0078] In this embodiment, the training dataset is input into the time-domain feature extraction module. This module includes two parts: sliding window segmentation and local feature extraction. Sliding window segmentation divides a long signal sequence into multiple shorter signal sequences. After segmentation using the sliding window technique, the local feature extraction module can be used to extract features from the n segmented sequences of the single input sequence.

[0079] Specifically, the radar initial signal is expressed as is the real number domain, T is the signal length, each time step includes the I and Q of the signal, and the signal is divided into n ordered sequences that can be expressed as W is the sliding window length. The improved DLSTM structure is used in the local feature extraction module, and a CBAM is connected after each layer of LSTM output. The first scale LSTM layer extracts time series features at a short sequence length and converts χ into u1 is the number of hidden units in this layer of LSTM. Then the data is processed into Then, in order to extract features at a larger scale, Each adjacent feature vector is concatenated into a new patch and input into the second layer of LSTM to obtain a larger scale feature vector u2 is the number of hidden layers in the second LSTM layer. Then the data is processed into CBAM is a lightweight convolutional attention module that combines channel and spatial attention mechanism modules. It can focus on data from both spatial and channel aspects to increase the weight of data on more important features. Channel attention helps to enhance the feature representation of different channels, while spatial attention helps to extract key information at different locations in space. The CBAM module can capture the correlation between features in different dimensions, thereby improving the performance of the model. The signal I / Q data is extracted through the time domain feature extraction module to obtain the time domain feature matrix.

[0080] The frequency domain feature extraction module includes: a data segmentation unit, a time-frequency conversion unit and a second local feature extraction module; the data segmentation unit is used to segment the original time series in the training data set to obtain segmented data; the time-frequency conversion unit uses EDFT to perform time-frequency conversion on the segmented data to obtain frequency domain data; the second local feature extraction module adopts an improved DLSTM structure, and the second local feature extraction module is used to extract the features of the frequency domain data to obtain the frequency domain feature matrix

[0081] In this embodiment, the training set data is input into the frequency domain feature extraction module. The original time series is segmented, and the segmented data is passed through EDFT to obtain an input spectrum aligned with the length of the complete sequence. Using traditional DFT may introduce frequency misalignment between the spectrum of the input sequence and the entire sequence. Therefore, the model built on the input sequence may lack complete and accurate frequency information, resulting in less accurate prediction of the complete sequence. To solve this problem, this embodiment uses extended EDFT, which overcomes the limitation of input length. This enables us to obtain an input spectrum consistent with the DFT frequency group of the complete sequence. The EDFT calculation formula is as follows:

[0082]

[0083] Among them, F[k] represents the spectrum, L represents the input length, T represents the output length, and j represents the imaginary unit; using this formula, we get a spectrum of length L+T, which is consistent with the DFT spectrum of the complete sequence. After using EDFT to process the cut signal into frequency domain data, the data is input into the local feature extraction module to extract local features and obtain the frequency domain feature matrix

[0084] The crisscross attention module consists of two layers of crisscross multi-head attention, each layer includes two multi-head attentions.

[0085] In this embodiment, the time domain feature matrix and frequency domain feature matrix Input the cross attention module. The cross attention module consists of two layers of cross multi-head attention. In the first cross layer, the inputs of the two multi-head attentions are the time domain features and frequency domain features extracted previously. One of them is input as the query vector, the other is input as the key vector and value vector, and the other is the opposite; the second cross layer uses the output of the previous cross layer as input. Such a structure can increase the model's mining of global similarity features of the data, achieve better recognition performance of long-sequence multi-pulse signals, and also achieve a feature fusion effect. Take one of the multi-head attention layers as an example: the input of the attention layer and is the feature matrix and The query vector q in attention i , key vector k i Sum value vector v i is The matrices of the three projections are Multiplying together gives:

[0086]

[0087]

[0088]

[0089] Among them, d q d k d v Represents the dimensions of the three projection matrices, d q =d k =d v Finally, through the calculation of the attention mechanism, the output result is in Z 11 , Z 12 , Z21 , Z 22 They are the output results of the four multi-head attention mechanisms. After two layers of multi-head attention operations and data splicing, the result is Z s .

[0090] In this embodiment, the data is input into the signal detection task head, and the mean square error is used as the loss function to calculate the loss of the model in training. The formula of the mean square error and the detection task head is as follows:

[0091]

[0092] y D =Sigmod(Z S W D +b D )

[0093] Among them, Loss D Represents model loss, i represents N integers from 1, and y represents the actual value. represents the predicted value, N represents the number of samples, y D Represents the detection task head, Sigmod represents the activation function, and the loss W D represents the weight of the dense layer in the detection task head, b D Represents the bias of the dense layer in the detection task head. Input the data into the modulation recognition task head and use cross entropy as the loss function to calculate the loss of the model in training. R Perform calculations and classify it using the softmax classifier:

[0094]

[0095] y R =Soft nmax(Z S W R +b R )

[0096] Among them, LossR represents the model loss, y R Represents the recognition task head, W R represents the weight of the dense layer in the recognition task head, b R Represents the bias of the dense layer in the recognition task head.

[0097] The constructed model is then trained and tested, inferring a fixed-size set of M predictions, one for each pulse. For signal detection tasks, the predicted start and end positions of the pulse in the time domain are obtained, while for modulation recognition tasks, the predicted signal type is obtained. During the training phase, the model is optimized using a linear combination of detection and recognition losses. During the testing phase, for the signal detection task, the start and end positions of the pulse are obtained through regression. A key issue here is matching each pulse in the prediction set with a pulse in the ground truth set.

[0098] In the part with signal spikes, the intersection-over-union (IoU) metric measures the overlap between the predicted region and the true region, that is, the ratio of their intersection length to their union length. The IoU score is calculated for each spike separately to measure their similarity.

[0099]

[0100]

[0101] Where b and Represent the true value and predicted value of the pulse position respectively. Each IoU result is compared with the threshold. If it is greater than the threshold value, it is regarded as a successful detection (TP), and if it is less than the threshold value, it is a missed detection (FN), and the network prediction output result is obtained. For the part without signal pulses, in the data set, the start and end time of this part of the signal are both 0, and the IoU calculation cannot be performed. Therefore, this embodiment uses the threshold value multiplied by the length of the shortest pulse in the data set as the evaluation criterion. If the length of the predicted value of the pulse position obtained is greater than the threshold value multiplied by the length of the shortest pulse in the data set, it is determined to be a false alarm (FP), otherwise it is determined to be a true negative (TN). Then calculate the recall rate (Pd) and the false alarm rate (Pfa), and draw the ROC curve. The formula for calculating the recall rate and the false alarm rate is as follows:

[0102]

[0103]

[0104] For the modulation recognition task, the model uses the argmax operation to find the most likely result in the predicted probability distribution of the test set labels.

[0105]

[0106] Where c is the possible prediction result, P() represents the probability calculation, and a is the true label. The accuracy of modulation recognition is calculated by comparing the predicted result with the true label.

[0107] S5. Obtain the radar signal to be detected and complete radar signal detection using the signal detection and recognition model.

[0108] Example 2

[0109] In this embodiment, a radar signal detection system based on time-frequency domain feature fusion network multi-task learning includes: a signal sequence acquisition module, a sequence processing module, a labeling module, a network construction module and a detection module.

[0110] The signal sequence acquisition module is used to establish a mathematical model of radar signals and use the mathematical model to obtain various types of radar signal sequences; the sequence processing module is used to process the radar signal sequence to obtain radar signal I sequence and radar signal Q sequence; the labeling module is used to label the radar signal I sequence and radar signal Q sequence and construct a training data set; the network construction module is used to construct a time-frequency domain feature fusion network and train the time-frequency domain feature fusion network using the training data set to obtain a signal detection and recognition model; the detection module is used to obtain the radar signal to be detected and complete radar signal detection using the signal detection and recognition model.

[0111] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A radar signal detection method based on multi-task learning of time-frequency domain feature fusion network, characterized by: The following steps are involved: Establishing a radar signal mathematical model and using the mathematical model to obtain various types of radar signal sequences; Process the radar signal sequence to obtain a radar signal I Sequences and radar signals Q sequence; The radar signal I Sequence and the radar signal Q The sequence is labeled and a training dataset is constructed; Constructing a time-frequency domain feature fusion network, and training the time-frequency domain feature fusion network using the training data set to obtain a signal detection and recognition model; Acquire a radar signal to be detected, and complete radar signal detection using the signal detection and recognition model; The time-frequency domain feature fusion network includes: a time-domain feature extraction module, a frequency-domain feature extraction module, a cross-attention module and a detection task head; The time domain feature extraction module is used to extract the time domain feature matrix in the training data set; The frequency domain feature extraction module is used to extract the frequency domain feature matrix in the training data set; The cross attention module is used to perform attention mechanism calculation and data splicing on the time domain feature matrix and the frequency domain feature matrix to obtain an output result; The detection task head is used to perform radar signal detection based on the output result; The time domain feature extraction module includes: a signal sliding window cutting module and a first local feature extraction module; The signal sliding window cutting module is used to segment the radar initial signal in the training data set to obtain m An ordered sequence: in, represents a set of sequences, X m Indicates the m A sequence, represents the field of real numbers, m Represents the number of ordered sequences, w represents the sliding window length; The first local feature extraction module adopts an improved DLSTM structure, and the first local feature extraction module is used to extract the sequence set The characteristics of the time domain feature matrix are obtained ; The frequency domain feature extraction module includes: a data segmentation unit, a time-frequency conversion unit and a second local feature extraction module; The data segmentation unit is used to segment the original time series in the training data set to obtain segmented data; The time-frequency conversion unit performs time-frequency conversion processing on the segmented data using EDFT to obtain frequency domain data: in, F [ k ] represents the spectrum, L Indicates the input length, T Indicates the output length, j Represents the smallest imaginary unit; The second local feature extraction module adopts an improved DLSTM structure, and the second local feature extraction module is used to extract the features of the frequency domain data to obtain the frequency domain feature matrix ; The cross attention module consists of two layers of cross multi-head attention, each layer includes two of the multi-head attentions.

2. The radar signal detection method based on time-frequency domain feature fusion network multi-task learning according to claim 1 is characterized in that: The radar signal mathematical model includes: in, Represents a radar signal, x [ k ] represents the noise-free radar signal, n [ k ] represents the receiver noise, k represents the sampling index, K Indicates the number of samples.

3. The radar signal detection method based on time-frequency domain feature fusion network multi-task learning according to claim 2 is characterized in that: The method for obtaining the radar signal sequence includes: obtaining a radar signal using the radar signal mathematical model, and performing a Hilbert transform on the radar signal to obtain a radar signal sequence: in, represents the radar signal sequence, j represents the imaginary unit, H ( ) represents Hilbert transform.

4. The radar signal detection method based on time-frequency domain feature fusion network multi-task learning according to claim 3 is characterized in that: radar signals I The sequence is: in, I [ k ] indicates radar signal I Sequence, Re( ) means taking the real part; radar signals Q The sequence is: in, Q [ k ] indicates radar signal Q Sequence, Im( ) means taking the imaginary part.

5. A radar signal detection system based on multi-task learning of time-frequency domain feature fusion network, wherein the detection system applies the detection method according to any one of claims 1 to 4, characterized in that: include: Signal sequence acquisition module, sequence processing module, annotation module, network construction module and detection module; The signal sequence acquisition module is used to establish a radar signal mathematical model and use the mathematical model to acquire various types of radar signal sequences; The sequence processing module is used to process the radar signal sequence to obtain a radar signal I Sequences and radar signals Q sequence; The marking module is used to mark the radar signal I Sequence and the radar signal Q The sequence is labeled and a training dataset is constructed; The network construction module is used to construct a time-frequency domain feature fusion network, and train the time-frequency domain feature fusion network using the training data set to obtain a signal detection and recognition model; The detection module is used to obtain the radar signal to be detected and complete the radar signal detection using the signal detection and recognition model.

Images