A radar signal intelligent detection model construction method based on Channel-DeepIQ, a radar signal intelligent detection method, a computer device and a storage medium

Abstract

A method for constructing a radar signal intelligent detection model based on Channel‑DeepIQ, a radar signal intelligent detection method, a computer device and a storage medium. The present invention relates to the field of electronic reconnaissance, and specifically to the field of radar signal intelligent detection. In order to solve the problem that the existing signal detection method has unsatisfactory detection effect under low signal-to-noise ratio conditions. The present invention combines digital channelization with neural network technology to improve the detection performance of low signal-to-noise ratio signals while maintaining a certain false alarm probability. After digital channelization improves the signal-to-noise ratio, a classification network is used for the arrival of signals in the sub-channel, that is, a convolution kernel with a structure similar to that of a matching receiver to complete the matching reception classification of the sub-channel; for pulse edge signals, a jump point detection network is used to detect the pulse edge and divide the signal area and the noise area. The present invention provides possibilities for more complex signal detection tasks.

Description

Technical Field

[0001] The present invention relates to the field of electronic reconnaissance, and in particular to the field of intelligent detection of radar signals. Background Art

[0002] As an important part of electronic warfare, the main function of the electronic reconnaissance receiver is to intercept the signals transmitted by the enemy and extract the parameter information of the signals so that we can make corresponding countermeasures. Since our electronic reconnaissance receiver and the enemy's radiation source are in a non-cooperative working mode, the reconnaissance receiver is required to have a wide monitoring bandwidth. The effective detection of radiation source signals by broadband receivers is an important part of achieving victory in electronic warfare. With the application of more and more radars with low probability of interception (LPI) characteristics in modern battlefields, the radiation source signals are weaker than the background noise, and are easily interfered by a large amount of noise during transmission. Therefore, the signals received by the receiver often show low signal-to-noise ratio characteristics. Conventional adaptive threshold detection methods are difficult to meet the performance requirements of current broadband digital receivers, which makes it increasingly difficult for electronic reconnaissance systems to detect signals under low signal-to-noise ratio conditions.

[0003] Digital channelization technology is proposed in electronic warfare reconnaissance receivers to deal with high-speed broadband signal reception. The most common is an efficient implementation structure based on polyphase filtering. This structure divides the monitoring frequency band into multiple independent sub-channels through a group of polyphase filter banks, which can reduce the data rate and complete effective signal noise reduction to reduce the difficulty of subsequent signal detection. In the increasingly complex non-cooperative electromagnetic environment, the use of constant false alarm rate (CFAR) detection technology can ensure the stable operation of the electronic reconnaissance system. The relevant threshold parameters of conventional radar signal detection methods are artificially set in advance. Under the condition of low signal-to-noise ratio, the detection performance is seriously reduced, which has certain application limitations and human subjectivity. Intelligent signal processing has always been an important direction in the field of signal processing. For example, the current popular neural network technology has the advantages of large-scale parallel processing, strong adaptability, strong learning ability, association and error correction. Compared with traditional signal processing systems, intelligent processing has stronger learning ability and can effectively solve problems such as signal detection and classification, target information extraction and recognition through training. In addition, when the training samples are rich enough, the adaptability to the environment is also strong, providing new possibilities for intelligent signal detection in electronic reconnaissance.

[0004] Through the search of existing technical literature, it was found that the "Blind Detection of Broadband Signal Based on Weighted Bi-Directional Feature Pyramid Network" published in the "SENSORS" journal on January 30, 2023, constructed a broadband signal target detection model based on YOLOv5, and the input signal short-time Fourier transform time-frequency image obtained good detection results. However, under the condition of low signal-to-noise ratio, its false alarm rate has far exceeded 1E-3, and the normal operation of the detection system cannot be guaranteed. "Accurate Estimation of LPI Radar Pulse Train Parametersvia Change Point Detection" published in the "IEEE Access" journal on February 6, 2023 uses change point detection to determine the specific arrival time and pulse width of the signal and other parameters to complete the parameter measurement of the signal. However, the setting of the change point parameters is extremely subjective and can only adapt to a single environment. The "Low Probability of Intercept Radar Signal Detection Algorithm Based on Convolutional Neural Network" in the Journal of Electronics and Information Technology proposes an intelligent detection method for radar signals based on convolutional neural networks, which uses the similarity between the convolution kernel and the matched filter structure to accurately match and receive the signal. However, the network can only detect whether the signal exists, but cannot determine the specific arrival time of the signal. The "Radiation Source Signal Detection Method Based on Sequential Statistical Filtering and Binary Accumulation" published on August 19, 2021 in Systems Engineering and Electronics Technology proposes a parallel pipeline signal detection method suitable for digital channelized receivers, but under low signal-to-noise ratio conditions, the detection performance of this detection method drops sharply. The Chinese invention application "A Passive Radar Signal Intelligent Detection Method Based on Inverted U-Net" with publication number CN116804740A proposes a signal intelligent detection method, which combines digital channelization with neural network technology to still have a good detection effect on the signal under low signal-to-noise ratio conditions. However, the false alarm probability of the system increases sharply with the decrease of the signal-to-noise ratio.

[0005] The search results of existing literature show that the existing signal detection methods have certain limitations, such as low algorithm detection probability and high false alarm probability under low signal-to-noise ratio. In general, the current signal detection methods are difficult to detect low signal-to-noise ratio signals. Summary of the invention

[0006] In order to solve the problem that the existing signal detection method has unsatisfactory detection effect under low signal-to-noise ratio conditions, the present invention provides the following solution:

[0007] A method for constructing a radar signal intelligent detection model based on Channel-DeepIQ, the method comprising:

[0008] Step 11: Perform digital channelization preprocessing on the sampled radar pulse signal to obtain the information matrix A IQ , the information matrix A IQ It consists of sub-channel IQ data;

[0009] Step 12: Use the information matrix A IQ Create a dataset;

[0010] Generate labels in the data set, the labels include: labels of the classification network Labels for the transition point detection network and ideal binary detection results

[0011] Step 13: construct a radar signal intelligent detection model based on Channel-DeepIQ, wherein the first layer of the radar signal intelligent detection model based on Channel-DeepIQ is a classification network, and the second layer is a jump point detection network;

[0012] Step 14: train the Channel-DeepIQ-based radar signal intelligent detection model and input the information matrix A into the classification network. IQ , get the classification result C, calculate the classification result C and the label of the classification network The cross entropy loss is used to update the parameters of the classification network;

[0013] Input the subchannel IQ data to the transition point detection network to obtain a classification result L, and calculate the classification result L and the label of the transition point detection network. The cross entropy loss is used to update the parameters of the jump point detection network;

[0014] Using the binary detection results The performance of the Channel-DeepIQ-based radar signal intelligent detection model is analyzed, and finally a trained Channel-DeepIQ-based radar signal intelligent detection model is obtained.

[0015] Further, in step 11, the information matrix A IQ for:

[0016]

[0017] Among them, A Ii,j Represents the I-channel data at the j-th sampling point output by the i-th sub-channel; the number of channels is K; and N is the number of sampling points in the sub-channel signal.

[0018] Further, in step 12, the label of the classification network for:

[0019]

[0020] in represents the classification result of the i-th subchannel;

[0021] The label of the transition point detection network for:

[0022]

[0023] in Refers to the jump point classification result of the jth point;

[0024] The ideal binary detection result for:

[0025]

[0026] in, It represents the ideal binary detection result at the jth sampling point of the i-th subchannel output, and its value is 0 or 1.

[0027] Further, in step 13, the classification network includes: a normalization layer, three dense convolution activation blocks, a fully connected layer module and a classifier.

[0028] Furthermore, the dense convolution activation block includes: a convolution layer, a RELU activation layer, a maximum pooling layer, and a batch normalization layer.

[0029] Further, in step 13, the jump point network includes: a normalization layer, a Bi-LSTM layer, a multi-head attention layer, a fully connected layer and a classifier

[0030] A radar signal intelligent detection method based on Channel-DeepIQ, the method comprising:

[0031] Step 21: Perform digital channelization preprocessing on the radar pulse signal to be detected to obtain the information matrix A of the radar pulse signal to be detected IQ ;

[0032] Step 22: The information matrix A of the radar pulse signal to be detected IQInputting a radar signal intelligent detection model based on Channel-DeepIQ to obtain a detection result D; the radar signal intelligent detection model based on Channel-DeepIQ is constructed by using the method described above;

[0033] The test result D is:

[0034]

[0035] Among them, D i,j Represents the binary detection result at the jth sampling point of the i-th subchannel output, and its value is 0 or 1.

[0036] Further, in step 22, the detection result is specifically: when the sub-channel is pure noise, the corresponding channel detection result D i,1～N =0; when the subchannel is a full pulse signal, the corresponding channel detection result D i,1～N =1;

[0037] When the sub-channel is a signal with a pulse edge, the detection results in the noise area are all 0, and the detection results in the signal area are all 1.

[0038] The present invention also provides a computer device, which includes 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 any one of the above-mentioned model building methods or detection methods.

[0039] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, any one of the above-mentioned model building methods or detection methods is implemented.

[0040] Beneficial effects of the present invention:

[0041] Compared with the prior art, the present invention aims at the problem that low signal-to-noise ratio signal detection is difficult and the false alarm rate of the neural network intelligent detection method is unstable. A radar signal intelligent detection method based on Channel-DeepIQ is designed. After the low signal-to-noise ratio signal within the large instantaneous monitoring bandwidth is slowed down and the signal-to-noise ratio is improved by digital channelization preprocessing, it is combined with neural network technology to achieve accurate detection under low signal-to-noise ratio conditions. When the signal is detected under low signal-to-noise ratio conditions, the signal intelligent detection method proposed by the present invention has better detection effect than the ordinary intelligent detection method and the traditional adaptive threshold detection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a training flow chart of the classification network in implementation mode 4;

[0043] Figure 2 Schematic diagram of the jump point in implementation mode 6;

[0044] Figure 3 The training flow chart of the transition point detection network in the sixth implementation mode;

[0045] Figure 4 is a flow chart of a radar signal intelligent detection method based on Channel-DeepIQ in implementation mode seven;

[0046] Figure 5 526MHz signal in the eleventh embodiment of the digital channelization output amplitude diagram when the signal-to-noise ratio is 0dB, wherein Figure a is the output amplitude diagram of the full pulse signal, Figure b is the output amplitude diagram of the pulse edge signal, and Figure c is the output amplitude diagram of pure noise;

[0047] Figure 6 It is a schematic diagram of detection probability statistical curves of the present invention and the comparative detection method in implementation mode eleven, wherein the comparative detection method includes: IU-Net detection, traditional mean value detection and sequential statistics detection;

[0048] Figure 7 It is a schematic diagram of the false alarm probability statistical curve of the present invention and the comparison detection algorithm in the eleventh implementation mode. DETAILED DESCRIPTION

[0049] Embodiment 1: A method for constructing a radar signal intelligent detection model based on Channel-DeepIQ, the method comprising:

[0050] Step 11: Perform digital channelization preprocessing on the sampled radar pulse signal to obtain the information matrix A IQ , the information matrix A IQ It consists of sub-channel IQ data;

[0051] Step 12: Use the information matrix A IQ Create a dataset;

[0052] Generate labels in the data set, the labels include: labels of the classification network Labels for the transition point detection network and ideal binary detection results

[0053] Step 13: construct a radar signal intelligent detection model based on Channel-DeepIQ, wherein the first layer of the radar signal intelligent detection model based on Channel-DeepIQ is a classification network, and the second layer is a jump point detection network;

[0054] Step 14: train the Channel-DeepIQ-based radar signal intelligent detection model and input the information matrix A into the classification network. IQ , get the classification result C, calculate the classification result C and the label of the classification network The cross entropy loss is used to update the parameters of the classification network;

[0055] Input the subchannel IQ data to the transition point detection network to obtain a classification result L, and calculate the classification result L and the label of the transition point detection network. The cross entropy loss is used to update the parameters of the jump point detection network;

[0056] Using the binary detection results The performance of the Channel-DeepIQ-based radar signal intelligent detection model is analyzed, and finally a trained Channel-DeepIQ-based radar signal intelligent detection model is obtained.

[0057] The training set production data under different signal-to-noise ratios and various signal arrival scenarios are shown in Table 1:

[0058] Table 1

[0059]

[0060] Implementation method 2: This implementation method further limits the method for constructing a radar signal intelligent detection model based on Channel-DeepIQ described in Implementation method 1. In this implementation method, in step 11, the information matrix A IQ pass:

[0061]

[0062] Get, where A Ii,j Represents the I-channel data at the j-th sampling point output by the i-th sub-channel; the number of channels is K; and N is the number of sampling points in the sub-channel signal.

[0063] Implementation method 3: This implementation method further limits the method for constructing a radar signal intelligent detection model based on Channel-DeepIQ described in Implementation method 1. In step 12, the label of the classification network pass:

[0064]

[0065] in represents the classification result of the i-th subchannel;

[0066] The label of the transition point detection network pass:

[0067]

[0068] in Refers to the jump point classification result of the jth point;

[0069] The ideal binary detection result pass:

[0070]

[0071] Obtain, among which, It represents the ideal binary detection result at the jth sampling point of the i-th subchannel output, and its value is 0 or 1.

[0072] Making labels for classification networks There are three types of label classification results:

[0073] Label 0 means the subchannel is pure noise;

[0074] Label 1 indicates that the subchannel is a full pulse signal;

[0075] Label 2 indicates that the subchannel is a pulse edge signal.

[0076] Making labels for the transition point detection network When the rising or falling edge of the pulse edge signal is regarded as a jump point, there are two types of label classification results:

[0077] Label 0 means that the point is a normal point;

[0078] Label 1 indicates that this point is a transition point.

[0079] Implementation method 4: This implementation method is a further limitation of the method for constructing a radar signal intelligent detection model based on Channel-DeepIQ described in implementation method 1. In step 13, the classification network includes: a normalization layer, three dense convolution activation blocks, a fully connected layer module and a classifier.

[0080] Information Matrix A IQ As the input of the network, after the normalization layer, by normalizing each sub-channel separately, the training speed of the network can be accelerated and the adaptability of the network can be improved.

[0081] The normalized data passes through three dense convolution activation blocks to extract the sub-channel information.

[0082] The classification result C of each sub-channel is output in parallel through multiple fully connected layer modules.

[0083] Calculate the classification result C and label through the classifier The cross entropy loss is calculated and the network parameters are updated.

[0084] The classification network is trained with multiple labels to ensure that the output results of different sub-channels are independent, and has a good detection effect in the case of parallel signal detection.

[0085] Figure 1 The training process of the classification network is shown, where DCEBlock-3 represents three DCEBlocks connected in series.

[0086] Implementation method five: This implementation method is a further limitation of the method for constructing a radar signal intelligent detection model based on Channel-DeepIQ described in implementation method four. The dense convolution activation block includes: a convolution layer, a RELU activation layer, a maximum pooling layer, and a batch normalization layer.

[0087] The RELU activation function is:

[0088]

[0089] Using batch normalization layers can increase the learning rate, ignore the impact of certain network initialization settings, and suppress overfitting.

[0090] Implementation method 6: Combination Figure 4 and Figure 5 This embodiment is described. This embodiment further limits the method for constructing a radar signal intelligent detection model based on Channel-DeepIQ described in Implementation 1. In step 13, the jump point network includes: a normalization layer, a Bi-LSTM layer, a multi-head attention layer, a fully connected layer and a classifier.

[0091] The input of the network is the sub-channel IQ data, which can be accelerated through the normalization layer to improve the network's training speed and adaptability.

[0092] The normalized data passes through the Bi-LSTM layer and the multi-head attention layer.

[0093] The core of LSTM is the selective forgetting and addition of information contained in the cell state, so as to learn long-term dependent information. By learning forward and reverse time series through Bi-LSTM, better detection results can be achieved. However, LSTM lacks the ability to adaptively focus on certain areas, which may lead to redundancy or information loss during the learning process. The multi-head attention mechanism can enable LSTM to focus on key feature vectors, enhancing the ability to capture the most relevant information from the input data, thereby achieving more accurate detection results.

[0094] The transition point classification result L is output through the fully connected layer module.

[0095] Calculate the classification result L and label through the classifier The cross entropy loss is calculated and the network parameters are updated.

[0096] The loss function of the jump point detection network is different from the general cross entropy loss function:

[0097]

[0098] where ω c It is the ratio of the number of normal points to the number of jump points in the sequence.

[0099] The transition point detection network uses a multi-head attention mechanism to focus on the detection of pulse edges, reducing the impact of noise and completing signal detection by learning the characteristics of timing.

[0100] Figure 2 It is a schematic diagram of the jump point, which shows the change of the signal.

[0101] Figure 3 Demonstrates the training process of the transition point detection network.

[0102] Embodiment 7: A radar signal intelligent detection method based on Channel-DeepIQ, the method comprising:

[0103] Step 21: Perform digital channelization preprocessing on the radar pulse signal to be detected to obtain the information matrix A of the radar pulse signal to be detected IQ ;

[0104] Step 22: The information matrix A of the radar pulse signal to be detected IQ Inputting a radar signal intelligent detection model based on Channel-DeepIQ to obtain a detection result; the radar signal intelligent detection model based on Channel-DeepIQ is constructed by using the methods described in embodiments one to six;

[0105] The test result D is:

[0106]

[0107] Among them, D i,j Represents the binary detection result at the jth sampling point of the i-th subchannel output, and its value is 0 or 1.

[0108] By combining digital channelization with neural network technology, the detection performance of low signal-to-noise ratio signals can be improved while maintaining a certain false alarm probability. After digital channelization improves the signal-to-noise ratio, a classification network, that is, a convolution kernel with a structure similar to that of a matching receiver, is used to complete the matching reception classification of the sub-channel based on the arrival of the signal in the sub-channel; for the pulse edge signal, a jump point detection network is used to detect the pulse edge and divide the signal area and the noise area. Channel-DeepIQ further improves the detection probability of the signal under low signal-to-noise ratio conditions and maintains the false alarm rate at a low level, providing the possibility for more complex signal detection tasks to solve the problem that existing detection methods are difficult to detect signals under low signal-to-noise ratio conditions.

[0109] Figure 4 The flowchart of the Channel-DeepIQ model signal intelligent detection method designed in this embodiment is shown.

[0110] Implementation 8: This implementation further limits the radar signal intelligent detection method based on Channel-DeepIQ described in Implementation 7. In step 22, the detection result is specifically: when the sub-channel is pure noise, the corresponding channel detection result D i,1～N =0; when the subchannel is a full pulse signal, the corresponding channel detection result D i,1～N =1;

[0111] When the sub-channel is a signal with a pulse edge, the detection results in the noise area are all 0, and the detection results in the signal area are all 1.

[0112] Use the trained Channel-DeepIQ to detect the signal. The input size of Channel-DeepIQ is [32,1024]. IQ The Channel-DeepIQ is input in parallel pipelines with a step of 1024. The network matches, receives and classifies the signals of multiple sub-channels and outputs the classification results of each sub-channel in parallel. When the sub-channel is a pulse edge signal, the jump point detection network is used to detect the jump point in the sub-channel. The area before and after the jump point can be divided into a signal area and a noise area by the high and low changes in the mean value. The detection results of the noise area are all 0, and the detection results of the signal area are all 1. Finally, the detection result D of the N-point signal in the sub-channel is obtained. i,1～N .

[0113] Embodiment 9. This embodiment proposes a computer device, including a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the model building method described in any one of embodiments 1 to 6, or the detection method described in any one of embodiments 7 to 8.

[0114] Embodiment 10. This embodiment proposes a computer-readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the model building method described in any one of embodiments 1 to 6, or the detection method described in any one of embodiments 7 to 8.

[0115] Implementation 11: The above implementation is described below by way of example

[0116] The simulation parameters for this implementation example are set as follows:

[0117] The sampling rate is 2400MHz, the number of channels is 32, the monitoring frequency band range is [200MHz, 1000MHz], and real channelization only needs to monitor the first 16 channels, using a uniform channelization structure. The input of the Channel-DeepIQ neural network model is the IQ data splicing of 16 channels, each sub-channel samples 1024 points, and the input size is "32×1024"; the output is the signal detection result of each point in each channel, and the size is "16×1024". The network learning rate is 0.0001, which decreases by 25% every 15 rounds, the batch size is 128, the gradient decay factor is 0.9, and the Adam optimizer is used for weight update.

[0118] In order to make the network adapt to the current complex electromagnetic environment, the network needs to be generalized. Therefore, when making the data set, it is necessary to consider multiple situations and make it as close to the real scene as possible. The frequency of each signal in the data set is randomly generated within the monitoring frequency band. When the signal appears at the overlap of the channel, it can be regarded as two signals arriving at the same time. The signal-to-noise ratio range is set to [-10dB, 5dB]. The arrival time and signal pulse width are random. The total number of data sets is 8192, of which the ratio of training set to test set is 9:1. Since 16 channels with 1024 points each are used as an input, there may be three hypotheses for the signal in each input channel:

[0119] Full pulse signal H0;

[0120] Pulse edge signal H1;

[0121] Pure noise H2.

[0122] Figure 5 The output amplitudes of the three hypothetical digital channelizations mentioned above are shown when the signal-to-noise ratio is 0dB and the signal frequency is 526MHz.

[0123] The data set of this implementation example takes into account the actual pulse arrival scenario. In order to ensure the diversity of samples, the signal arrival time, frequency, and pulse width are set to random. The signal frequency is randomly generated within the detection frequency band, and the minimum pulse width is 1us. The data set is relatively rich. The network model is trained using the training set in the above signal-to-noise ratio range to obtain the optimal network model.

[0124] In order to verify the detection performance of the Channel-DeepIQ network model proposed in this invention, it is necessary to produce data sets of random frequencies, random pulse widths, and random arrival times under different signal-to-noise ratios, perform signal detection performance verification, compare with the ideal detection results, and statistically analyze the experimental results. Add the following comparative detection method:

[0125] IU-Net neural network detection method. U-Net has been effectively applied in signal detection and recognition, image segmentation and other fields, while IU-Net adopts a processing method of first increasing the dimension and then reducing the dimension, which can increase the richness of target features and realize end-to-end signal detection.

[0126] Sequential statistical filtering detection method. The method for setting the threshold of sequential statistical filtering is generally to sort the instantaneous amplitudes of all output channels first, with a reference sliding window size of 16; the step is 1; then select the optimal statistic, and the optimal criterion is set to "0.75k"; finally, multiply the optimal statistic with a threshold coefficient α (set according to the false alarm rate, the false alarm rate is set to 1E-3), thereby obtaining the threshold estimation threshold of the sequential statistical filtering.

[0127] Traditional mean threshold detection method. The traditional mean threshold setting method is generally to average the instantaneous amplitude of all output channels, then multiply the average value by a threshold coefficient β (generally β = 2), and then add the minimum noise floor value output from all channels to obtain the traditional mean threshold estimation threshold.

[0128] In actual projects, the traditional adaptive threshold detection method is often combined with the MN detection technology to correct its detection results. That is, there are M successful detections in N detections in each sliding window, which means that there is a signal in the current sliding window, otherwise there is no signal. This time, the "2 / 4 criterion" is adopted.

[0129] Figure 6 It is a schematic diagram of the detection probability statistical curve of the present invention and the comparative detection method. A radiation source signal is generated in the data set here, and its signal carrier frequency is randomly generated within the monitoring frequency range, the arrival time is random, the pulse width is random, and the signal-to-noise ratio range of the data set samples is [-10dB, 5dB].

[0130] Figure 7 It is a schematic diagram of false alarm probability statistical curves of the present invention and the comparison detection algorithm.

[0131] from Figure 6 and Figure 7 From the statistical results, we can see that the detection performance of traditional mean detection and sequential statistics detection methods seriously deteriorates after 0dB; although IU-Net still has good detection capabilities under low signal-to-noise ratio, the false alarm probability increases with the decrease of signal-to-noise ratio. Channel-DeepIQ detection is less affected and can achieve better detection results while maintaining a low false alarm probability. The experimental results show that the Channel-DeepIQ hierarchical network detection model obtained by the example training in this implementation has a significant effect on improving the signal detection performance, stabilizes the false alarm probability, and provides the possibility for more complex signal detection tasks.

Claims

1. A method for constructing a radar signal intelligent detection model based on Channel-DeepIQ, characterized in that: The method comprises: Step 11: Perform digital channelization preprocessing on the sampled radar pulse signal to obtain the information matrix A IQ , the information matrix A IQ It consists of sub-channel IQ data; Step 12: Use the information matrix A IQ Create a dataset; Generate labels in the data set, the labels include: labels of the classification network Labels for the transition point detection network and ideal binary detection results Step 13: construct a radar signal intelligent detection model based on Channel-DeepIQ, wherein the first layer of the radar signal intelligent detection model based on Channel-DeepIQ is a classification network, and the second layer is a jump point detection network; Step 14: train the Channel-DeepIQ-based radar signal intelligent detection model and input the information matrix A into the classification network. IQ , get the classification result C, calculate the classification result C and the label of the classification network The cross entropy loss is used to update the parameters of the classification network; Input the subchannel IQ data to the transition point detection network to obtain a classification result L, and calculate the classification result L and the label of the transition point detection network. The cross entropy loss is used to update the parameters of the jump point detection network; Using the binary detection results Analyze the performance of the Channel-DeepIQ-based radar signal intelligent detection model, and finally obtain a trained Channel-DeepIQ-based radar signal intelligent detection model; According to the signal arrival situation in the sub-channel, the classification network is used for detection. When the sub-channel is pure noise, the corresponding channel detection result D i,1～N =0; when the subchannel is a full pulse signal, the corresponding channel detection result D i,1～N =1; for the pulse edge signal, the jump point detection network is used for detection. When the sub-channel is a pulse edge signal, the detection results in the noise area are all 0, and the detection results in the signal area are all 1.

2. The method for constructing a radar signal intelligent detection model based on Channel-DeepIQ according to claim 1, characterized in that: In step 11, the information matrix A IQ for: Among them, A Ii,j A represents the I-channel data at the j-th sampling point output by the i-th subchannel; Qi,j Represents the Q-channel data of the j-th sampling point of the i-th sub-channel; the number of channels is K; and N is the number of sampling points in the sub-channel signal.

3. The method for constructing a radar signal intelligent detection model based on Channel-DeepIQ according to claim 1, characterized in that: In step 12, the label of the classification network for: in represents the classification result of the i-th subchannel; The label of the transition point detection network for: in Refers to the jump point classification result of the jth point; The ideal binary detection result for: in, It represents the ideal binary detection result at the jth sampling point of the i-th subchannel output, and its value is 0 or 1.

4. The method for constructing a radar signal intelligent detection model based on Channel-DeepIQ according to claim 1, characterized in that: In step 13, the classification network includes: a normalization layer, three dense convolution activation blocks, a fully connected layer module and a classifier.

5. The method for constructing a radar signal intelligent detection model based on Channel-DeepIQ according to claim 4, characterized in that: The dense convolution activation block includes: a convolution layer, a RELU activation layer, a maximum pooling layer, and a batch normalization layer.

6. The method for constructing a radar signal intelligent detection model based on Channel-DeepIQ according to claim 1, characterized in that: In step 13, the jump point detection network includes: a normalization layer, a Bi-LSTM layer, a multi-head attention layer, a fully connected layer and a classifier.

7. A radar signal intelligent detection method based on Channel-DeepIQ, characterized in that: The method comprises: Step 21: Perform digital channelization preprocessing on the radar pulse signal to be detected to obtain the information matrix A of the radar pulse signal to be detected IQ ; Step 22: The information matrix A of the radar pulse signal to be detected IQ Inputting a radar signal intelligent detection model based on Channel-DeepIQ to obtain a detection result D; the radar signal intelligent detection model based on Channel-DeepIQ is constructed by using the method described in any one of claims 1 to 6; The test result D is: Among them, D i,j Represents the binary detection result at the jth sampling point of the i-th subchannel output, and its value is 0 or 1.

8. The radar signal intelligent detection method based on Channel-DeepIQ according to claim 7, characterized in that: In step 22, the detection result is as follows: when the subchannel is pure noise, the corresponding channel detection result D i,1～N =0; when the subchannel is a full pulse signal, the corresponding channel detection result D i,1～N =1; When the sub-channel is a signal with a pulse edge, the detection results in the noise area are all 0, and the detection results in the signal area are all 1.

9. A computer device comprising a memory and a processor, characterized in that A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the model building method described in any one of claims 1 to 6, or the detection method described in any one of claims 7 to 8.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the model building method described in any one of claims 1 to 6, or the detection method described in any one of claims 7 to 8.

Images