Radar radiation source signal sorting method based on CNN-LSTM composite depth model

Abstract

The invention discloses a radar radiation source signal sorting method based on a CNN-LSTM composite depth model, and the method comprises the steps: carrying out the simulation generation of aliasing radar radiation source pulse signal data, building a CNN network model, adding an LSTM layer, obtaining a CNN-LSTM composite depth model, carrying out the preprocessing of the aliasing radar radiation source pulse signal data, dividing a training set and a test set, and carrying out the recognition of the aliasing radar radiation source pulse signal data. And training the CNN-LSTM composite depth model by using the training set, finally inputting test set data to the trained CNN-LSTM composite depth model, and carrying out aliasing radar radiation source pulse signal sorting to obtain a radar radiation source sorting result. Compared with an existing method and a sorting method based on a full-connection network, the method has the advantages that the sorting accuracy is higher, the robustness is high under the condition that the leakage pulse rate and the false pulse rate in input pulses are high, a new theoretical support and a technical approach can be provided for radar radiation source signal sorting in a modern complex electromagnetic environment, and the method is suitable for popularization and application. And effective help is provided for radar signal receiving and analysis.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal reception and analysis, and in particular relates to a radar radiation source signal sorting method based on a CNN-LSTM composite depth model. Background Art

[0002] With the continuous development of electronic technology, various radar equipment has been put into use, and radars are everywhere. Therefore, in a sense, radar radiation source signal sorting has become the core link of radar signal reception and analysis technology, providing more possibilities for further processing and analysis of radar radiation source signals. The results can directly affect the quality of radar intelligence, so the relevant technology has important research significance and practical value.

[0003] Existing signal sorting is mostly based on pulse-to-pulse characteristic PDW parameters and PRI parameters. However, with the development of radar electronic technology, the modulation types of radar signals are emerging in an endless stream. The accuracy of existing algorithms can no longer meet the needs of sorting complex radiation source signals in modern environments. At present, deep learning has made major breakthroughs in many fields, and deep learning, as a new machine learning method, provides new ideas for radar pulse signal sorting. The paper "Denoising and Deinterleaving of Pulse Streams with Recurrent Neural Networks.IEEE Transactions on Aerospace and Electronic Systems.2018" uses recurrent neural networks to sort radiation sources with overlapping pulse repetition intervals or pulse width parameters in various noise environments. The paper "Radar Emitter Classification with Attribute-specific Recurrent Neural Networks.arXiv preprint arXiv, 1·911.07683.2019" optimizes the normalization method and uses the optimized network of recurrent neural networks, namely the long short-term memory network, to successfully sort 17 radiation sources.

[0004] Although the above methods can show good results in sequence data processing, the structures of these networks are relatively simple, which limits the model's ability to capture diverse sequence features, resulting in insufficient performance when processing highly complex or nonlinear sequence data. This structural limitation makes it difficult to achieve efficient learning and generalization when processing long sequence dependencies and high-dimensional data. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a radar radiation source signal sorting method based on a CNN-LSTM composite deep model. First, the aliased radar radiation source pulse signal data is preprocessed to adapt it to the network, and then the processed signal data is sent to the CNN-LSTM network for training and learning to finally obtain the sorting result.

[0006] The technical solution adopted by the present invention is: a radar radiation source signal sorting method based on a CNN-LSTM composite depth model, and the specific steps are as follows:

[0007] S1, simulating and generating aliased radar radiation source pulse signal data;

[0008] In a simulation environment, n groups of pulse data are generated according to different radar radiation sources; the pulse data of different radar radiation sources are distinguished by using labels, and the n groups of pulse data are aliased according to the labels.

[0009] Among them, the corresponding pulse characteristic parameters should be set in the simulation environment, including: pulse carrier frequency RF, pulse width PW, pulse repetition interval PRI and the measurement error of the corresponding pulse characteristic parameters; and the missed pulse rate and false pulse rate when the receiver receives the radar radiation source pulse signal in ideal and non-ideal scenarios should be considered.

[0010] Among them, in an ideal scenario, the existence of measurement error, missed pulse rate and false pulse rate is not considered, and these three parameters are set to 0; in a non-ideal scenario, the measurement error, missed pulse rate and false pulse rate are set according to the actual situation.

[0011] S2. Establish a CNN network model, and add an LSTM layer to the CNN network model to obtain a CNN-LSTM composite deep model;

[0012] Establishing a CNN network model includes: constructing the input layer, convolution layer, activation function, pooling layer, fully connected layer and output layer in sequence. The input layer receives the input training data; the convolution layer extracts the local features of the input data through the convolution kernel; the activation function introduces the nonlinear part; the pooling layer reduces the dimension of the feature vector while ensuring the invariance of the input; the fully connected layer performs matrix-vector product on the extracted features for classification operations in the subsequent output layer; and the output layer outputs the results.

[0013] Among them, the convolution layer, the activation function and the pooling layer constitute a convolutional space-time block, and the convolutional space-time block and the fully connected layer together constitute the CNN network model.

[0014] Then, an LSTM layer is added to the CNN network model to obtain a CNN-LSTM composite deep model, that is, an LSTM layer is added between the convolutional spatiotemporal block and the fully connected layer in the CNN network model, and an attention mechanism is added between the LSTM layer and the fully connected layer.

[0015] Among them, the LSTM layer model includes: input gate, forget gate and output gate; the details are as follows:

[0016] (1) The input gate of the LSTM layer first determines the new state that should be added to the long short-term memory cell. The input gate is divided into two parts. The first part is to use the input gate sigmoid layer to determine the data at some specific locations that need to be updated. The second part is to use the tanh layer to generate a vector composed of candidate values ​​that can be added to the state of the long short-term memory cell. The input gate state update expression is as follows:

[0017] i t =σ(W t [h t-1 ,x t ]+b t ) (1)

[0018]

[0019] Where σ(·) represents the sigmoid function, [h t-1 ,x t ] represents the output value h of the LSTM layer at the previous moment t-1 and the input value x of LSTM at the current moment t The vector composed of t represents the output of the input gate, W i represents the weight matrix of the input gate, b i Represents the bias of the input gate. represents the unit state at the current moment, W c represents the weight of the input unit, b c Represents the bias of the input unit.

[0020] (2) The forget gate of the LSTM layer is a sigmoid layer that combines the hidden state of the previous moment and the input data of the current moment to decide whether to choose to forget this information. "0" means completely forgetting, and "1" means completely retaining. The state update expression of the forget gate is as follows:

[0021] f t =σ(W f [h t-1 ,x t ]+b f ) (3)

[0022] Among them, ft represents the output of the forget gate, [h t-1 ,x t ] represents the output value h of the LSTM layer at the previous moment t-1 and the input value x of LSTM at the current moment t The vector composed of f represents the weight matrix of the forget gate, b f Represents the bias of the forget gate.

[0023] (3) The output gate of the LSTM layer is based on the current state of the LSTM cell. First, a tanh layer is used to map the LSTM cell state to the interval (-1, 1). Then, the sigmoid layer of the output gate of the LSTM layer is used to control the information in the LSTM output cell state. The output gate state update expression is as follows:

[0024] o t =σ(W o [h t-1 ,x t ]+b o ) (4)

[0025] Among them, t Represents the output of the current output gate, W o represents the weight matrix of the output gate, b o Represents the bias of the output gate.

[0026] Update long and short-term memory cells C t The cell state is expressed as follows:

[0027]

[0028] Here, * represents the Hadamard product.

[0029] The output expression of the output gate is as follows:

[0030] h t =o t *tanh(C t ) (6)

[0031] The convolutional spatiotemporal block, the LSTM layer, the attention mechanism and the fully connected layer constitute the CNN-LSTM composite deep model.

[0032] S3, preprocessing the aliased radar radiation source pulse signal data obtained in step S1, dividing it into a training set and a test set, and using the training set to train the CNN-LSTM composite deep model constructed in step S2;

[0033] Based on step S1 , pulse data of a plurality of different radar radiation sources are generated, and the pulse carrier frequency, pulse width and pulse arrival time of each radar radiation source under the measurement error are generated according to the set value.

[0034] A characteristic parameter difference method is used for preprocessing. First, X is set to represent the PDW characteristic parameter of the radar pulse signal to be selected. Then, n pulses are selected from the aliased radar radiation source pulse signal generated in step S1. For the i-th pulse signal in the selected pulse, the X characteristic value of the next r pulses is compared with the X characteristic value X of the i-th pulse. i Subtract the difference one by one, and the expression is as follows:

[0035] DX ij =|X i+j -X i | (7)

[0036] Among them, DX ij Represents X i+j With X i The absolute value of the difference, i=1,2,…,n,j=1,2,…,r.

[0037] The matrix DX obtained after the difference is the difference matrix of the X feature. The three features of the selected pulse arrival time, pulse carrier frequency and pulse width are calculated according to the method of formula (7) to obtain the three feature difference matrices of pulse arrival time difference matrix DTOA, carrier frequency difference matrix DRF and pulse width difference matrix DPW. Each row in the matrix represents the correlation difference information between the pulse at the current moment and the pulse at the subsequent moment.

[0038] For the TOA feature in the feature parameters, instead of directly performing a difference preprocessing operation on it, a difference preprocessing operation is performed on the difference obtained by subtracting the TOA parameters item by item, as shown in equations (8) and (9):

[0039] ΔTOA i =|TOA i+1 -TOA i | (8)

[0040] DΔTOA i+j =|ΔTOA i+j -ΔTOA i | (9)

[0041] Where, i = 1, 2, ..., n; j = 1, 2, ..., r, ΔTOA i represents the absolute difference between TOAs, that is, the absolute value of the difference between the i+1th TOA and the ith TOA; DΔTOA i+j Represents the difference between the i+jth absolute difference and the ith absolute difference.

[0042] Finally, the preprocessed pulse data is normalized and flattened into one-dimensional data to adapt it to the input of the network.

[0043] The processed pulse data are set as a training set and a test set according to the ratio set by the requirements, and finally the training set is sent to the CNN and LSTM composite deep model for model training. The convolutional spatiotemporal block in the CNN-LSTM composite deep model extracts local features of the pulse data, and the LSTM layer in the CNN-LSTM composite deep model extracts time series features of the pulse data. The attention mechanism in the CNN-LSTM composite deep model adaptively fuses the local features and time series features of the sequence data, and finally the classification output is performed through the fully connected layer in the CNN-LSTM composite deep model, and the network weight is adjusted according to the accuracy of the test set until the model loss value meets the requirements.

[0044] S4. Based on step S3, input the test set data into the trained CNN-LSTM composite deep model to perform aliasing radar radiation source pulse signal sorting to obtain radar radiation source sorting results;

[0045] The training set divided in step S3 is sent to the CNN-LSTM composite deep model trained in step S3 for testing. The output result of the model is the sorted radar emitter label. The radar emitter sorting result is obtained according to the label, and the accuracy rate acc of the radar emitter sorting result is calculated. The calculation expression is as follows:

[0046]

[0047] Among them, right means that the number of labels output by radar radiation source pulse data sorting is the same as the number of actual radar radiation sources, and all means the number of all radar radiation source pulses that need to be sorted.

[0048] Beneficial effects of the present invention: The method of the present invention first simulates and generates aliased radar radiation source pulse signal data, then establishes a CNN network model and adds an LSTM layer to obtain a CNN-LSTM composite depth model, then pre-processes the aliased radar radiation source pulse signal data, divides the training set and the test set, uses the training set to train the CNN-LSTM composite depth model, and finally inputs the test set data into the trained CNN-LSTM composite depth model to perform aliased radar radiation source pulse signal sorting to obtain radar radiation source sorting results. The method of the present invention first establishes a CNN model, and adds an LSTM layer to the CNN model, and uses the CNN-LSTM composite depth model to extract and classify the aliased radar radiation source pulse signal. In the one-dimensional CNN model, the fully connected layer cannot effectively utilize historical information and can only consider the characteristics of the current time. In contrast, the LSTM model can better capture the long-term dependency in the one-dimensional sequence data through a unique gate, so the method of the present invention utilizes the advantages of locality and error adaptability of the convolutional neural network and the long-distance dependency mining ability of the long short-term memory network, thereby improving the accuracy of radar radiation source sorting and the robustness of the classification model. Compared with the existing methods and the sorting methods based on fully connected networks, the method of the present invention has higher sorting accuracy and strong robustness in the face of high leakage pulse rate and false pulse rate in the input pulses. It can provide new theoretical support and technical approach for the sorting of radar radiation source signals in modern complex electromagnetic environments, and provide effective assistance for radar signal reception and analysis, which is of great significance in the field of radar signal reception and analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of a radar radiation source signal sorting method based on a CNN-LSTM composite depth model of the present invention.

[0050] Figure 2 Schematic diagram of setting pulse-to-pulse modulation parameters of radar radiation source signal in an embodiment of the present invention.

[0051] Figure 3 Schematic diagram of the path of the CNN-LSTM composite deep model in an embodiment of the present invention.

[0052] Figure 4 It is a schematic diagram of the training accuracy and loss value curve of the CNN-LSTM composite deep model in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The present invention is mainly verified by simulation experiment, and all steps and conclusions are verified to be correct on Matlab platform. The method of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0054] like Figure 1As shown, a flow chart of a radar radiation source signal sorting method based on a CNN-LSTM composite depth model of the present invention is shown, and the specific steps are as follows:

[0055] S1, simulating and generating aliased radar radiation source pulse signal data;

[0056] In this embodiment, n groups of pulse data are generated according to different radar radiation sources in the Matlab environment; the pulse data of different radar radiation sources are distinguished by using labels, such as Figure 2 As shown, n groups of pulse data are aliased according to the labels.

[0057] according to Figure 2 The corresponding pulse characteristic parameters are set in the data, including: pulse carrier frequency RF, pulse width PW, pulse repetition interval PRI and the measurement error of the corresponding pulse characteristic parameters; and the missed pulse rate and false pulse rate when the receiver receives the radar radiation source pulse signal under ideal and non-ideal scenarios are considered.

[0058] Among them, the error parameter values ​​of the two scenarios are shown in Table 1. In the ideal scenario, the existence of measurement error, leakage pulse rate and false pulse rate is not considered, and these three parameters are set to 0; in the non-ideal scenario of this embodiment, the measurement error is set to 5%, the leakage pulse rate is set to 20%, and the false pulse rate is set to 10%.

[0059] Table 1

[0060] serial number Scenario Measurement error Leakage pulse rate False pulse rate 1 Ideal scenario 0 0 0 2 Non-ideal scenario 5％ 20％ 10％

[0061] S2. Establish a CNN network model, and add an LSTM layer to the CNN network model to obtain a CNN-LSTM composite deep model;

[0062] Establishing a CNN network model includes: constructing the input layer, convolution layer, activation function, pooling layer, fully connected layer and output layer in sequence. The input layer receives the input training data; the convolution layer extracts the local features of the input data through the convolution kernel.

[0063] Among them, CNN is a deep feedforward neural network with local connections and full weight sharing, which is widely used in algorithm research. They are named after convolution, which is a mathematical operation applied to input data. The linear operation expression of convolution between two time series is as follows:

[0064]

[0065] Among them, B(t) and C(t) represent two time series, A(t) represents the time series after the convolution of B(t) and C(t), B(τ) represents the convolved function, C(t-τ) represents the convolution kernel, t represents the variable of the output function, and τ represents the integral variable, that is, the position where the convolution kernel slides.

[0066] Filters are two-dimensional networks of discrete numbers that have the ability to recognize specific patterns in the input data. Essentially, a convolutional layer performs a matrix multiplication operation between the layer's input and the selected filter by moving the filter over the input data matrix. Specifically, the filter is first adjusted to coincide with the upper left edge of the input matrix, allowing an element-by-element multiplication between the two matrices to be performed. The result of this operation corresponds to the first element of the new matrix representing the output of the layer, and the value of this element is used to determine whether the desired feature is present.

[0067] In the CNN network model, the activation function introduces a nonlinear part; the pooling layer reduces the dimension of the feature vector while ensuring the invariance of the input; the fully connected layer performs matrix-vector product on the extracted features for classification operations in the subsequent output layer; and the output layer outputs the results.

[0068] Among them, the convolution layer, the activation function and the pooling layer constitute a convolutional space-time block, and the convolutional space-time block and the fully connected layer together constitute the CNN network model.

[0069] Then, an LSTM layer is added to the CNN network model to obtain a CNN-LSTM composite deep model, that is, an LSTM layer is added between the convolutional spatiotemporal block and the fully connected layer in the CNN network model, and an attention mechanism is added between the LSTM layer and the fully connected layer. LSTM is a variant of a recurrent neural network, which can effectively solve the gradient explosion or disappearance problem of a simple recurrent neural network. The main improvements of LSTM are in the following two aspects: LSTM introduces a new internal state for linear cyclic information transmission, while nonlinearly outputting information to the external state of the hidden layer, and a new gating mechanism is added, where the gate is a binary variable {0,1}, 0 represents the closed state, and no information is allowed to pass; 1 represents the open state, and all information is allowed to pass. Then the path of information transmission is controlled by the gate.

[0070] Among them, the LSTM layer model includes: input gate, forget gate and output gate; the details are as follows:

[0071] (1) The input gate of the LSTM layer first determines the new state that should be added to the long short-term memory cell. The input gate is divided into two parts. The first part is to use the input gate sigmoid layer to determine the data at some specific locations that need to be updated. The second part is to use the tanh layer to generate a vector composed of candidate values ​​that can be added to the state of the long short-term memory cell. The input gate state update expression is as follows:

[0072] i t =σ(W t [h t-1 ,x t ]+b t ) (2)

[0073]

[0074] Where σ(·) represents the sigmoid function, [h t-1 ,x t ] represents the output value h of the LSTM layer at the previous moment t-1 and the input value x of LSTM at the current moment t The vector composed of t represents the output of the input gate, W i represents the weight matrix of the input gate, b i Represents the bias of the input gate. represents the unit state at the current moment, W c represents the weight of the input unit, b c Represents the bias of the input unit.

[0075] (2) The forget gate of the LSTM layer is a sigmoid layer that combines the hidden state of the previous moment and the input data of the current moment to decide whether to choose to forget this information. "0" means completely forgetting, and "1" means completely retaining. The state update expression of the forget gate is as follows:

[0076] f t =σ(W f [h t-1 ,x t ]+b f ) (4)

[0077] Among them, f t represents the output of the forget gate, [h t-1 ,x t ] represents the output value h of the LSTM layer at the previous moment t-1 and the input value x of LSTM at the current moment t The vector composed of f represents the weight matrix of the forget gate, b f Represents the bias of the forget gate.

[0078] (3) The output gate of the LSTM layer is based on the current state of the LSTM cell. First, a tanh layer is used to map the LSTM cell state to the interval (-1, 1). Then, the sigmoid layer of the output gate of the LSTM layer is used to control the information in the LSTM output cell state. The output gate state update expression is as follows:

[0079] o t =σ(W o [h t-1 ,x t ]+b o ) (5)

[0080] Among them, t Represents the output of the current output gate, W o represents the weight matrix of the output gate, b o Represents the bias of the output gate.

[0081] Update long and short-term memory cells C t The cell state is expressed as follows:

[0082]

[0083] Here, * represents the Hadamard product.

[0084] The output expression of the output gate is as follows:

[0085] h t =o t *tanh(C t ) (7)

[0086] An attention mechanism is added between the LSTM layer and the fully connected layer, as follows:

[0087] The attention mechanism is a data processing method in machine learning. The degree of attention (importance) to different information is reflected by weights. The attention mechanism can be regarded as a query matrix (Query), key (key) and weighted average constitute a multilayer perceptron (MLP). The working principle of attention is similar to addressing. Given a query matrix of an element in the target, the weight coefficient of each key corresponding value is obtained by calculating the similarity or correlation between the query matrix and each key, and then the value is weighted summed to obtain the final attention value. Therefore, in essence, the attention mechanism is a weighted sum of the values ​​of the elements, and the query matrix and key are used to calculate the weight coefficients of the corresponding values. The attention mechanism aims to enable the model to automatically learn important features in the signal and focus more attention on the signal segments that are critical to the classification task. By introducing the attention mechanism, the model can better learn and utilize the key information in the signal and optimize the performance of the classification task. In the context of radar emitter pulse signal sorting, applying this technology to the time series data of radar pulse signals can help the model focus on the decisive part of the radar pulse signal, thereby improving the sorting accuracy.

[0088] The hidden state in a recurrent neural network stores historical information and represents a kind of memory. In a simple recurrent network, the hidden state is rewritten at every moment, so it can be regarded as a short-term memory. In a neural network, long-term memory can be regarded as a network parameter, which implies the experience learned from the training data. Its update cycle is much shorter than that of short-term memory. In LSTM, the memory unit can capture a certain key information at a certain moment and has the ability to save the less critical information for a certain time interval. The life cycle of information stored in a memory unit is longer than short-term memory, but much shorter than long-term memory, which makes it very suitable for situations where information must be maintained for a long time for accurate prediction. And due to the existence of the forget gate, input gate and output gate, the LSTM network solves the problem of long-term dependency and alleviates the problem of gradient disappearance and explosion, thus achieving good results in various sequence processing fields.

[0089] The convolutional spatiotemporal block, the LSTM layer, the attention mechanism and the fully connected layer constitute a path diagram of the CNN-LSTM composite deep model as shown in Figure 3 shown.

[0090] S3, preprocessing the aliased radar radiation source pulse signal data obtained in step S1, dividing it into a training set and a test set, and using the training set to train the CNN-LSTM composite deep model constructed in step S2;

[0091] This embodiment generates 210,000 pulse data of 4 different radar radiation sources based on step S1, and generates the pulse carrier frequency, pulse width and pulse arrival time of each radar radiation source under the measurement error according to the setting values ​​in Table 1.

[0092] This embodiment uses a characteristic parameter difference method for preprocessing. First, X is set to represent the PDW characteristic parameter of the radar pulse signal to be selected. Then, n pulses are selected from the aliased radar radiation source pulse signal generated in step S1. For the i-th pulse signal in the selected pulse, the X characteristic values ​​of the next r pulses are compared with the X characteristic value X of the i-th pulse. i Subtract the difference one by one, and the expression is as follows:

[0093] DX ij =|X i+j -X i | (8)

[0094] Among them, DX ij Represents X i+j With X i The absolute value of the difference, i=1,2,...,n,j=1,2,...,r.

[0095] The matrix DX obtained after the difference is the difference matrix of the X feature. The three features of the selected pulse arrival time, pulse carrier frequency and pulse width are calculated according to the method of formula (8) to obtain the three feature difference matrices of pulse arrival time difference matrix DTOA, carrier frequency difference matrix DRF and pulse width difference matrix DPW. Each row in the matrix represents the correlation difference information between the pulse at the current moment and the pulse at the subsequent moment. While making full use of the existing available data, it can also effectively reflect the information contained in the feature parameter difference.

[0096] However, when the above r value is enlarged, if the TOA parameter is directly used for difference processing, the magnitude of the elements in the DTOA will be too large. This may cause insufficient memory during neural network training, resulting in poor learning effect of the neural network.

[0097] Therefore, the above difference preprocessing method is improved to a certain extent: for the TOA feature in the feature parameter, the difference preprocessing operation is no longer performed directly on it, but the difference preprocessing operation is performed on the difference obtained by subtracting the TOA parameters item by item, as shown in formula (9) and formula (10):

[0098] ΔTOA i =|TOA i+1 -TOA i | (9)

[0099] DΔTOAi+j =|ΔTOA i+j -ΔTOA i | (10)

[0100] Where i = 1, 2, ..., n; j = 1, 2, ..., r; ΔTOA i represents the absolute difference between TOAs, that is, the absolute value of the difference between the i+1th TOA and the ith TOA; DΔTOA i+j Represents the difference between the i+jth absolute difference and the ith absolute difference.

[0101] Using the DΔTOA matrix instead of the DTOA matrix obtained above to participate in the subsequent sorting process can make the data at the same order of magnitude and avoid the problem of insufficient memory allocation that may occur when generating a high-dimensional interpolation matrix. Considering that the receiver in the actual battlefield often receives aliased radar radiation source pulse signals, some pulses may produce aliasing, resulting in some characteristic parameters of pulse signals from different radar radiation sources being the same, but at this time other parameters are still independent and can play a role in sorting. Therefore, performing difference matrix processing on multiple characteristic parameters separately and splicing them together as the basis for sorting can greatly improve the accuracy of the sorting results.

[0102] Finally, the preprocessed pulse data is normalized and flattened into one-dimensional data to adapt it to the input of the network.

[0103] The processed pulse data are set as a training set and a test set in the ratio of 8:2 set according to the requirements of this embodiment, and finally the training set is sent to the CNN and LSTM composite deep model for model training. The convolutional spatiotemporal block in the CNN-LSTM composite deep model extracts local features of the pulse data, and the LSTM layer in the CNN-LSTM composite deep model extracts time series features of the pulse data. The attention mechanism in the CNN-LSTM composite deep model adaptively fuses the local features and time series features of the sequence data, and finally classified and outputted through the fully connected layer in the CNN-LSTM composite deep model, and the network weight is adjusted according to the accuracy of the test set until the model loss value meets the requirements.

[0104] In this embodiment, the CNN-LSTM composite depth model first inputs radar pulse signal data through the input layer, and then passes through the first convolution layer, the number of channels of this layer is 16, and a convolution kernel size of 2*1 is used, and the activation function is ReLU; the first maximum pooling layer uses a convolution kernel of 2*1; the number of channels of the second convolution layer is 32, and the convolution kernel size is 2*1; the second pooling layer configuration is the same as the first pooling layer. Immediately following is a Flatten layer, whose function is to convert a multi-dimensional vector into a one-dimensional vector to ensure that the output after convolution can be directly connected to the LSTM layer. Then comes the LSTM layer, whose activation function is tanh, and then a fully connected layer with an activation function of ReLU. The final output layer is set to 5 to represent the probabilities of different radars in the five-classification task, and then the output values ​​of the multi-classification are converted to the [0,1] range through the Softmax function. The training structure of the CNN-LSTM composite depth model is shown in Table 2.

[0105] Table 2

[0106] serial number Layer Name Activation Function parameter 1 Input Layer - - 2 Convolutional layer 1 ReLU Channel=16, Kernel_size=2*1 3 Max Pooling Layer 2 - Pool_size=2*1 4 Convolutional Layer 2 ReLU Channel=32, Kernel_size=2*1 5 Max Pooling Layer 2 - Pool_size=2*1 6 Flatten layer - - 7 LSTM Layer tanh - 8 Fully connected layer ReLU - 9 Classification layer Softmax -

[0107] S4. Based on step S3, input the test set data into the trained CNN-LSTM composite model to perform aliasing radar radiation source pulse signal sorting to obtain radar radiation source sorting results;

[0108] The training set divided in step S3 is sent to the CNN-LSTM composite deep model trained in step S3 for testing. The output result of the model is the sorted radar emitter label. The radar emitter sorting result is obtained according to the label, and the accuracy rate acc of the radar emitter sorting result is calculated. The calculation expression is as follows:

[0109]

[0110] Among them, right means that the number of labels output by radar radiation source pulse data sorting is the same as the number of actual radar radiation sources, and all means the number of all radar radiation source pulses that need to be sorted.

[0111] The accuracy results and effects of sorting the aliased radar radiation source signal using the CNN-LSTM composite depth model calculated by formula (11) are shown in Table 3 (sorting accuracy of different sorting methods in different environments) and Figure 4 shown.

[0112] Table 3

[0113] serial number Simulation scenario Identification method Accuracy 1 Ideal scenario Existing methods 99.0％ 2 Ideal scenario Fully connected network 89.8％ 3 Ideal scenario CNN-LSTM 99.9％ 4 Non-ideal scenario Existing methods 57.3％ 6 Non-ideal scenario Fully connected network 62.8％ 7 Non-ideal scenario CNN-LSTM 99.1％

[0114] By observing the data in Table 3, it can be found that the existing method can achieve a higher recognition accuracy through statistical methods in ideal scenarios, but the existing method is greatly affected by errors, and the accuracy is almost reduced by half in non-ideal scenarios. Due to its single network structure, the fully connected network cannot learn complex repetition characteristics, so the recognition rate is slightly lower, and the robustness to errors is poor. The method of the present invention achieved a higher accuracy in both scenarios of Table 1, achieving an accuracy of 99.9% in ideal scenarios, and maintaining a high accuracy of 99.1% in non-ideal scenarios, and the robustness to errors is also good. Therefore, it can be considered that the method of the present invention is effective and feasible for sorting aliased complex radar radiation source signals, which verifies the correctness and effectiveness of the method of the present invention.

[0115] In summary, the method of the present invention is mainly used to accurately and efficiently sort out each radar radiation source pulse train from the aliased radar radiation source pulse signal in radar signal reception analysis, taking advantage of the locality and error adaptability of CNN and the long-distance dependency mining capability of LSTM, and sorting and identifying four types of radar signals under different error scenarios. The simulation results show that the sorting and identification accuracy of CNN-LSTM is high, and it has good robustness when comparing errors of different ratios, and is superior to existing algorithms.

[0116] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A radar radiation source signal sorting method based on CNN-LSTM composite deep model, the specific steps are as follows: S1, simulating and generating aliased radar radiation source pulse signal data; In a simulation environment, n groups of pulse data are generated according to different radar radiation sources; the pulse data of different radar radiation sources are distinguished by using labels, and the n groups of pulse data are aliased according to the labels; in, The corresponding pulse characteristic parameters should be set in the simulation environment, including: pulse carrier frequency RF, pulse width PW, pulse repetition interval PRI and the measurement error of the corresponding pulse characteristic parameters; and the missed pulse rate and false pulse rate when the receiver receives the radar radiation source pulse signal in ideal and non-ideal scenarios should be considered; Among them, in an ideal scenario, the existence of measurement error, missed pulse rate and false pulse rate is not considered, and these three parameters are set to 0; in a non-ideal scenario, the measurement error, missed pulse rate and false pulse rate are set according to the actual situation; S2. Establish a CNN network model, and add an LSTM layer to the CNN network model to obtain a CNN-LSTM composite deep model; Establish a CNN network model, including: constructing the input layer, convolution layer, activation function, pooling layer, fully connected layer and output layer in sequence; the input layer receives the input training data; the convolution layer extracts the local features of the input data through the convolution kernel; the activation function introduces the nonlinear part; the pooling layer reduces the dimension of the feature vector while ensuring the invariance of the input; the fully connected layer performs matrix-vector product on the extracted features for the classification operation of the subsequent output layer; the output layer outputs the results; The convolutional layer, the activation function and the pooling layer constitute a convolutional spatiotemporal block, and the convolutional spatiotemporal block and the fully connected layer together constitute the CNN network model. Then, an LSTM layer is added to the CNN network model to obtain a CNN-LSTM composite deep model, that is, an LSTM layer is added between the convolutional spatiotemporal block and the fully connected layer in the CNN network model, and an attention mechanism is added between the LSTM layer and the fully connected layer; Among them, the LSTM layer model includes: input gate, forget gate and output gate; the details are as follows: (1) The input gate of the LSTM layer first determines the new state that should be added to the long short-term memory cell. The input gate is divided into two parts. The first part is to use the input gate sigmoid layer to determine the data at some specific locations that need to be updated. The second part is to use the tanh layer to generate a vector composed of candidate values ​​that can be added to the state of the long short-term memory cell. The input gate state update expression is as follows: i t =σ(W t [h t-1 ,x t ]+b t ) (1) Where σ(·) represents the sigmoid function, [h t-1 ,x t ] represents the output value h of the LSTM layer at the previous moment t-1 and the input value x of LSTM at the current moment t The vector composed of t represents the output of the input gate, W i represents the weight matrix of the input gate, b i represents the bias of the input gate; represents the unit state at the current moment, W c represents the weight of the input unit, b c represents the bias of the input unit; (2) The forget gate of the LSTM layer is a sigmoid layer that combines the hidden state of the previous moment and the input data of the current moment to decide whether to choose to forget this information; "0" means completely forgetting, and "1" means completely retaining. The state update expression of the forget gate is as follows: f t =σ(W f [h t-1 ,x t ]+b f ) (3) Among them, f t represents the output of the forget gate, [h t-1 ,x t ] represents the output value h of the LSTM layer at the previous moment t-1 and the input value x of LSTM at the current moment t The vector composed of f represents the weight matrix of the forget gate, b f Represents the bias of the forget gate; (3) The output gate of the LSTM layer is based on the current state of the LSTM cell. First, a tanh layer is used to map the LSTM cell state to the interval (-1, 1). Then, the sigmoid layer of the output gate of the LSTM layer is used to control the information in the LSTM output cell state. The output gate state update expression is as follows: the t =σ(W o [h t-1 ,x t ]+b o ) (4) Among them, t Represents the output of the current output gate, W o represents the weight matrix of the output gate, b o represents the bias of the output gate; Update long and short-term memory cells C t The cell state is expressed as follows: Where * represents the Hadamard product; The output expression of the output gate is as follows: h t =o t *tanh(C t ) (6) The convolutional spatiotemporal block, the LSTM layer, the attention mechanism and the fully connected layer constitute the CNN-LSTM composite deep model; S3, preprocessing the aliased radar radiation source pulse signal data obtained in step S1, dividing it into a training set and a test set, and using the training set to train the CNN-LSTM composite deep model constructed in step S2; Based on step S1, pulse data of multiple different radar radiation sources are generated, and pulse carrier frequency, pulse width and pulse arrival time of each radar radiation source under measurement error are generated according to set values; A characteristic parameter difference method is used for preprocessing. First, X is set to represent the PDW characteristic parameter of the radar pulse signal to be selected. Then, n pulses are selected from the aliased radar radiation source pulse signal generated in step S1. For the i-th pulse signal in the selected pulse, the X characteristic value of the next r pulses is compared with the X characteristic value X of the i-th pulse. i Subtract the difference one by one, and the expression is as follows: DX ij =|X i+j -X i | (7) Among them, DX ij Represents X i+j With X i The absolute value of the difference, i = 1, 2, ..., n, j = 1, 2, ..., r; The matrix DX obtained after the difference is the difference matrix of the X feature. The three features of the selected pulse arrival time, pulse carrier frequency and pulse width are calculated according to the method of formula (7) to obtain the three feature difference matrices of pulse arrival time difference matrix DTOA, carrier frequency difference matrix DRF and pulse width difference matrix DPW; each row in the matrix represents the correlation difference information between the pulse at the current moment and the pulse at the subsequent moment; For the TOA feature in the feature parameters, instead of directly performing a difference preprocessing operation on it, a difference preprocessing operation is performed on the difference obtained by subtracting the TOA parameters item by item, as shown in equations (8) and (9): ΔTOA i =|TOA i+1 -TOA i | (8) DΔTOA i+j =|ΔTOA i+j -ΔTOA i | (9) Where, i = 1, 2, ..., n; j = 1, 2, ..., r, ΔTOA i represents the absolute difference between TOAs, that is, the absolute value of the difference between the i+1th TOA and the ith TOA; DΔTOA i+j represents the difference between the i+jth absolute difference and the ith absolute difference; Finally, the preprocessed pulse data is normalized and flattened into one-dimensional data to adapt it to the input of the network; The processed pulse data is set as a training set and a test set according to the ratio set by the requirements, and finally the training set is sent to the CNN and LSTM composite deep model for model training. The convolutional spatiotemporal block in the CNN-LSTM composite deep model extracts local features of the pulse data, and the LSTM layer in the CNN-LSTM composite deep model extracts time series features of the pulse data. The attention mechanism in the CNN-LSTM composite deep model adaptively fuses the local features and time series features of the sequence data, and finally the classification output is performed through the fully connected layer in the CNN-LSTM composite deep model, and the network weight is adjusted according to the accuracy of the test set until the model loss value meets the requirements; S4. Based on step S3, input the test set data into the trained CNN-LSTM composite deep model to perform aliasing radar radiation source pulse signal sorting to obtain radar radiation source sorting results; The training set divided in step S3 is sent to the CNN-LSTM composite deep model trained in step S3 for testing. The output result of the model is the sorted radar emitter label. The radar emitter sorting result is obtained according to the label, and the accuracy rate acc of the radar emitter sorting result is calculated. The calculation expression is as follows: Among them, right means that the number of labels output by radar radiation source pulse data sorting is the same as the number of actual radar radiation sources, and all means the number of all radar radiation source pulses that need to be sorted.

Images