Radar signal processing system based on complex field pulse neural network

Abstract

The invention provides a radar signal processing system based on a complex field spiking neural network. The system is a complex field spiking neural network composed of a plurality of layer units. And the structure of each layer unit is that a layer of LIF neurons is superposed on the complex field spiking neural network layer. The complex number field spiking neural network layer is constructed based on a spiking neural network, input signals of the complex number field spiking neural network layer and weights of the spiking neural network are both in a complex number form, and the multiplication calculation of the input signals and the weights of the complex numbers during the forward propagation calculation of the spiking neural network is converted into the following steps: splitting real parts and imaginary parts of the input signals; and respectively multiplying by the real part and the imaginary part of the complex weight, then forming a new real part and a new imaginary part through addition and subtraction operation, and synthesizing the new real part and the new imaginary part into complex output. The method can improve the feature recognition precision of the complex field radar signals.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing and radar target detection and recognition, and in particular to a radar signal processing system based on a complex domain pulse neural network. Background Art

[0002] Electronic warfare first appeared on the historical stage in 1904 with the Russo-Japanese War, and the emergence of electronic warfare played an important role in the outcome of the Russo-Japanese War. With the acceleration of the modernization of warfare and the development of electronic information technology, electronic warfare has gradually become the main battlefield of modern warfare, playing an increasingly important role in the outcome of war. At the same time, the development of electronic warfare has also increased the importance of adjusting one's own response strategy by identifying the enemy's status. Therefore, the demand for using electronic warfare equipment with cognitive performance in war is also increasing.

[0003] Radar resonance region echo recognition, radar pole feature recognition, and radar radiation source recognition are all radar signal processing capabilities required by electronic warfare equipment. However, when processing complex domain radar signals, existing radar signal processing systems, whether traditional radar signal processing solutions or radar signal processing solutions based on deep learning, do not fully utilize the characteristic that radar echo signals are complex signals. Most of them only use the real part of the signal, or take the modulus value of the real and imaginary parts. This cannot extract all the feature information contained in the signal, resulting in a loss of accuracy. The development of deep learning is likely to increase the number of parameters, making it difficult to deploy models and implement hardware, and it is difficult to adapt to the needs of lightweight equipment in electronic warfare. Taking the radiation source recognition task as an example, the richer the extracted features, the higher the recognition accuracy. The real and imaginary parts of the complex signal have a certain correlation. The lack of any dimension of features is a loss, which reduces the recognition accuracy. Summary of the invention

[0004] In view of this, the present invention provides a radar signal processing system based on a complex domain pulse neural network, which can improve the feature recognition accuracy of complex domain radar signals.

[0005] The present invention is applicable to but not limited to radar resonance region echo identification, radar pole feature identification, radar radiation source identification and other tasks. In addition, the system can also be used for deployment tasks of airborne and satellite-borne identification.

[0006] In order to solve the above technical problems, the present invention is implemented as follows.

[0007] A radar signal processing system based on a complex domain pulse neural network, the system is a complex domain pulse neural network composed of multiple layer units; the structure of each layer unit is a complex domain pulse neural network layer superimposed with a layer of LIF neurons;

[0008] The complex domain pulse neural network layer is constructed based on the pulse neural network. The input signal of the complex domain pulse neural network layer and the weight of the pulse neural network are both in complex form. The multiplication calculation of the input complex signal and the complex weight during the forward propagation calculation of the pulse neural network is converted into: the real part and the imaginary part of the input complex signal are split, and then multiplied with the real part and the imaginary part of the complex weight respectively, and then new real part and imaginary part are formed through addition and subtraction operations, and the new real part and imaginary part are synthesized into a complex number and then output.

[0009] Preferably, the real part and imaginary part of the input complex signal are split, multiplied with the real part and imaginary part of the complex weight respectively, and then formed into new real part and imaginary part by addition and subtraction operations:

[0010] The real part a and imaginary part b of the input complex signal are multiplied by the real part c and imaginary part d of the complex weight respectively, and then the difference is calculated to obtain the new real part a·cb·d;

[0011] The real part a and imaginary part b of the input complex signal are multiplied by the real part c and imaginary part d of the complex weight respectively, and then the sum is obtained to obtain a new imaginary part a·d+c·b.

[0012] Preferably, the layer units include a complex pulse convolutional layer, a complex pulse linear layer, and a complex pulse fully connected layer.

[0013] Preferably, when the system is used for radar radiation source identification, it is composed of 2 complex pulse convolutional layers, 2 complex pulse linear layers, and 1 complex pulse fully connected layer connected in sequence.

[0014] Preferably, the system is suitable for radar resonance region echo identification, and / or radar pole feature identification, and / or radar radiation source identification.

[0015] Preferably, the method further comprises a training step: training using an automatic mixed precision method, with the hyperparameters used for training being: a total number of iterations of 100, a batch training size of 30, and a learning rate of 0.001.

[0016] Beneficial effects:

[0017] The present invention combines the complex domain characteristics of radar echo signals and designs a radar signal processing system based on a complex domain pulse neural network. The complex domain pulse neural network has stronger timing correlation, computing power and recognition performance, and can also effectively reduce the computational cost of training the model.

[0018] (1) The radar signal processing system based on complex pulse neural network designed by the present invention fully considers the characteristic that radar echo signal is a complex signal. It does not omit any dimension of the real part and the imaginary part, but analyzes and explores the correlation between the real part and the imaginary part, and uses the correlation between the two to perform channel processing and fusion. The complex domain pulse neural network architecture based on this design combines the powerful ability to process spatiotemporal data with complex signals to obtain a network model with powerful feature expression capabilities. The dual information sources can maximize the utilization of echo signals and improve recognition accuracy.

[0019] (2) The radiation source identification method based on complex domain pulse neural network not only has powerful spatiotemporal processing capabilities, but also uses fewer parameters.

[0020] The spiking neural network is different from the convolutional neural network in that it uses pulse signals to complete parameter transfer. The pulse signals contain rich time information, which gives it the ability to process space and time that the convolutional neural network does not have. Its unique brain-like structure also improves the computing power of the model. In addition, during the parameter transfer process, the spiking neural network only calculates the parameter amount when a single pulse occurs, thus reducing the number of parameters in the model.

[0021] (3) In a preferred embodiment, in order to improve the extraction of signal features, two layers of complex pulse convolution layer and complex pulse linear layer are stacked in the feature processing module. The complex pulse linear layer can expand the feature dimension, thereby solving the problem of reducing the number of features caused by the convolution operation.

[0022] (4) The simulation experiment data proves that the model can achieve a high recognition accuracy and a fast convergence speed. The results show that for tasks such as radar radiation source recognition, the present invention can better achieve radiation source recognition tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a structural diagram of a radar signal processing system based on a complex domain pulse neural network according to the present invention;

[0024] Figure 2 for Figure 1 Schematic diagram of the conversion structure of the complex-domain spiking neural network layer inside the middle-level units. DETAILED DESCRIPTION

[0025] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0026] The present invention provides a radar signal processing system based on a complex spiking neural network (Complex Spiking Neural Network, C-SNN). The basic idea is to transform the spiking neural network into a complex spiking neural network. Among them, the two most important tasks are: first, each spiking neural network layer in the spiking neural network is transformed into a complex spiking neural network layer to adapt to the input of complex radar signals. The transformed network layer is divided into two paths to process the real and imaginary parts of the complex signal, and then fuse them, so as not to omit any dimension of the real and imaginary parts, so as to maximize the utilization of the echo signal and improve the recognition accuracy of the complex signal. The second important task is to stack a layer of LIF (Leaky Integrate and Fire) neurons as an excitation function after each layer of the transformed complex domain spiking neural network layer. Through the above transformation, a complex domain spiking neural network model with multiple layer units is obtained.

[0027] It can be seen that the present invention fully considers the characteristic that radar echo signals are complex signals, does not omit any dimension of the real part and the imaginary part, but analyzes and explores the correlation between the real part and the imaginary part, and uses the correlation between the two to perform channel processing and fusion. Based on the complex domain pulse neural network architecture of this design, the powerful ability to process spatiotemporal data is combined with complex signals to obtain a network model with powerful feature expression capabilities. The dual information sources can maximize the use of echo signals and improve recognition accuracy.

[0028] Generally speaking, layer units may include convolutional layers, linear layers, and fully connected layers. When these layer units are constructed based on a spiking neural network and undergo the above-mentioned transformation, they may form a complex spiking convolutional layer (C-SCNN layer), a complex spiking linear layer (C-SLNN layer), and a complex spiking fully connected network layer (C-SFCN layer).

[0029] The radar signal processing system of the present invention can be used to complete tasks such as radar resonance region echo recognition, radar pole feature recognition, radar radiation source recognition, etc. When applied to different tasks, different layer unit functions can be designed and combined to form the required task model.

[0030] Taking radar emitter identification as an example, the composition and working principle of the radiation source identification network (Complex Domain Radiation Source Identification Network, CDRSI-Net) based on complex domain pulse neural network are described below.

[0031] like Figure 1 As shown, the CDRSI-Net constructed in this embodiment includes a complex domain pulse convolution layer (C-SCNN) that can process complex signals based on a pulse neural network (C-SNN), a complex domain pulse linear layer (C-SLNN) and a complex domain pulse fully connected layer (C-SFCN). In this embodiment, the radar signal processing system constructed using the above three network layers includes five network layers, namely five "layer units", which are two layers of C-SCNN layer units and two layers of C-SLNN layer units, and one layer of C-SFCN layer unit in order. The structure of each "layer unit" is a complex domain pulse neural network layer superimposed with a layer of LIF neurons. The complex domain pulse neural network layer among the five "layer units" is the main functional layer, and the activation of the network is achieved by superimposing a layer of LIF neurons after each layer to replace the role of the activation function.

[0032] Among them, the C-SCNN layer unit is the Complex Spiking Convolutional Neural Network Module. This module consists of two parts, a complex spiking convolutional layer and a layer of LIF neurons. The function of the complex spiking convolutional layer is to realize the conversion of complex signals and perform feature extraction at the same time; the LIF neuron completes the conversion of the convolutional neural network into a spiking neural network.

[0033] The C-SLNN layer unit is the Complex Spiking Linear Nerural Network module. This module consists of a complex spiking linear layer and LIF neurons. It solves the problem of too few features after the complex spiking convolution layer by expanding the feature dimension.

[0034] The C-SFCN layer unit is the Complex Spiking Fully Connected Network module. This module consists of a complex spiking fully connected layer and LIF neurons. The complex spiking fully connected layer realizes the transformation of complex signals, while receiving and classifying features from the complex spiking linear layer. The LIF neurons realize the role of the activation function.

[0035] From the above structure, we can see that the first layer unit of the network is used for feature extraction, which is achieved by stacking two layers of complex domain pulse convolution layers and two layers of complex domain pulse linear layers. After stacking two layers of complex domain pulse convolution layers, the number of features will become very small, while the complex domain pulse linear layer can expand the dimension of the features, so two layers of complex domain pulse linear layers are stacked to expand the dimension of the features. Then, the result of feature extraction is sent to the complex domain pulse fully connected layer to obtain the classification result.

[0036] The important structure in the above-mentioned "layer unit" is the complex domain pulse neural network layer, which is transformed from the pulse neural network layer. The key to the transformation is to transform the existing real network into a network structure that can process complex data. The specific process is: According to the characteristics of the radar signal, the radar echo signal input into the network is x=a+bj. The key to realizing the complex domain pulse neural network lies in the processing of the real and imaginary parts of the complex signal. The input signal is forward propagated in the pulse neural network, and the input signal is multiplied by the weight during the calculation. In order to realize the processing of complex radar echo signals, the weights in the present invention are also designed to be in complex form, recorded as c+dj.

[0037] Then, the multiplication of the input signal and the complex weight is shown in formula (1):

[0038]

[0039] Among them, a+bj represents the input complex signal, and c+dj represents the complex weight of the pulse neural network.

[0040] Generally speaking, here c and d are weight value functions f c and f d , then the above formula should be written as:

[0041]

[0042] It can be seen from the above formulas (1) and (2) that the real part and the imaginary part have a certain correlation, and they affect the real part and the imaginary part of the complex signal. Therefore, in order to better utilize the extracted features, it is necessary to transform the network layers (such as convolutional layers, linear layers, and fully connected layers) in the layer units. The network layer in the layer unit of the present invention is constructed based on a pulse neural network, and the specific transformation scheme is to change the weights of the pulse neural network into a complex form to adapt to the complex form of the input signal. After the modification, the multiplication calculation of the input signal and the complex weight during the forward propagation calculation of the pulse neural network is converted into: after the real and imaginary parts of the input complex signal are split, they are multiplied with the real and imaginary parts of the complex weights respectively, and then new real and imaginary parts are formed through addition and subtraction operations, and the new real and imaginary parts are synthesized into a complex output.

[0043] Figure 2The modified complex domain pulse neural network layer structure is shown, Figure 2 As shown, the complex domain spiking neural network layer includes a real processing channel, an imaginary processing channel and a fusion module. Among them, the real processing channel is composed of a real processing module and a subtraction module, and the imaginary processing channel is composed of an imaginary processing module and an addition module. The real processing module and the imaginary processing module correspond to the real part (real function) and imaginary part (imaginary function) of the complex weights in the spiking neural network. After the real and imaginary parts of the input signal enter these two modules for processing, new data components are generated. Among them, the real processing module calculates the real part a and imaginary part b of the input complex signal with the real part c of the complex weight, generates a·c and sends it to the subtraction module, and generates c·b and sends it to the addition module; the imaginary processing module calculates the real part a and imaginary part b of the input complex signal with the imaginary part d of the complex weight, generates a·d and sends it to the addition module, and generates b·d and sends it to the subtraction module. The subtraction module generates a new real part a·cb·d, and the addition generates a new imaginary part a·d+c·b. Finally, the fusion module synthesizes a new complex number a·cb·d+(a·d+c·b)j and outputs it to the subsequent LIF neurons.

[0044] Through transformation, each layer in the convolutional neural network of the present invention is converted into a network layer suitable for complex number processing, and all the inputs, outputs and weights of the model are complex data. The present invention makes the conversion of the pulse convolution layer, pulse linear layer and pulse fully connected layer of each module to satisfy the complex form signal, increases the expression of the signal timing characteristics, and facilitates the subsequent recognition tasks.

[0045] The design process and use scheme of the radar signal processing system based on the complex domain pulse neural network of the present invention are as follows:

[0046] Step 1: Convert the spiking neural network to a complex domain spiking neural network:

[0047] ① Each spiking neural network layer in the spiking neural network is transformed into a complex domain spiking neural network layer. After splitting the real part a and imaginary part b of the input complex signal, the complex domain spiking neural network layer multiplies the real part a and imaginary part b of the input signal with the real part c and imaginary part d of the complex weight respectively, and then calculates the difference to obtain a new real part a·cb·d; the real part a and imaginary part b of the input signal are multiplied with the real part c and imaginary part d of the complex weight respectively, and then sums them to obtain a new imaginary part a·d+c·b; finally, they are merged into a complex form, i.e., a·cb·d+(a·d+c·b)j, as the output of the complex domain spiking neural network layer.

[0048] ② After each layer of the modified complex domain pulse neural network layer, a layer of LIF neurons is stacked as an excitation function to obtain a complex domain pulse neural network model containing a complex domain pulse convolutional layer, a complex domain pulse linear layer and a complex domain pulse fully connected layer.

[0049] Step 2: Train the complex domain spiking neural network model.

[0050] The input signal is used to extract feature information using a complex pulse convolutional layer and a complex pulse linear layer, and then sent to a complex domain pulse fully connected layer for classification.

[0051] Calculate the difference between the feature data processed by CDRSI-Net and the true value of the identified label, optimize it using the cross entropy loss function, and update the gradient and weight of CDRSI-Net. Save the network training model after reaching the maximum number of iterations to test the recognition accuracy of CDRSI-Net for radar emitter signals. Use the automatic mixed precision method to train the model during training, which can reduce the size of the video memory occupied by the model and increase the training speed. The hyperparameters used for network training are: total number of iterations is 100, batch size is 30, and learning rate is 0.001. Use the Adam optimizer for iterative training.

[0052] Step 3: Load the trained network model, send the complex signal of the radar emitter into CDRSI-Net, and use the trained CDRSI-Net network architecture to complete the emitter identification task.

[0053] Table 1 Comparison of the indicators of the complex pulse neural network of the present invention and the two solutions of the prior art

[0054]

[0055] In order to verify the effect of the radar radiation source identification method of the complex domain pulse neural network proposed in the present invention on radar signal identification, the CDRSI-Net network was used for experimental verification, as shown in Table 1 above. This design compares the three schemes of the complex pulse neural network of the present invention, the complex domain convolutional neural network of the prior art, and the real domain pulse neural network. The parameter transfer of the pulse neural network is based on pulse signals one by one, so only the neuron layer that meets the conditions will transfer the parameters when it receives the pulse signal. The parameter transfer of the convolutional neural network is layer-by-layer, which increases the parameter amount to a certain extent.

[0056] The experimental results show that when the input is in complex form, the recognition rate of the pulse neural network is higher than that of the convolutional neural network and the number of parameters is smaller, which verifies the advantage of the pulse neural network in the radiation source identification task; when the convolutional neural network is used as the basic framework, the performance when the input data is in the form of real numbers and complex numbers is compared. The average recognition accuracy rate is increased by 0.7%, and the number of parameters is reduced by 3.6 times. This result verifies the effectiveness of the complex domain method to complete the radiation source identification task. The number of parameters of the pulse neural network. In summary, the complex domain pulse neural network can not only improve the accuracy of the radiation source identification task, but also reduce the number of parameters of the model, meet the needs of lightweight networks, and facilitate subsequent hardware deployment.

[0057] The advantage of this embodiment is that by mining the temporal characteristics of the signal in the radar echo, a radar emitter identification method of a complex domain pulse neural network suitable for the radar emitter identification task is designed. The CDRSI-Net architecture uses a complex pulse neural network to directly extract complex features, which better retains the correlation characteristics of the amplitude and phase in the radar echo; the C-SCNN module, C-SLNN module and C-SFCN module are designed to fully mine the correlation of complex data and the temporal characteristics of the radar signal echo; without increasing the complexity of the model, the feature processing performance of CDRSI-Net is further enhanced through feature reuse between different extraction modules, realizing the radar emitter identification task and achieving good recognition accuracy. In addition, the introduction of the pulse network gives the model more powerful computing power and reduces the power consumption of the radiation source identification task.

[0058] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in the description may be different and are not limited. Therefore, those skilled in the art in the field of the present invention may modify or replace the technical solutions recorded in the above embodiments; and these modifications and replacements do not deviate from the creative purpose and technical solutions of the present invention and should all fall within the protection scope of the present invention.

Claims

1. A radar signal processing system based on complex domain pulse neural network, characterized in that: The system is a complex domain spiking neural network composed of multiple layer units; the structure of each layer unit is a complex domain spiking neural network layer superimposed with a layer of LIF neurons; The complex domain pulse neural network layer is constructed based on the pulse neural network. The input signal of the complex domain pulse neural network layer and the weight of the pulse neural network are both in complex form. The multiplication calculation of the input complex signal and the complex weight during the forward propagation calculation of the pulse neural network is converted into: the real part and the imaginary part of the input complex signal are split, and then multiplied with the real part and the imaginary part of the complex weight respectively, and then new real part and imaginary part are formed through addition and subtraction operations, and the new real part and imaginary part are synthesized into a complex number and then output.

2. The radar signal processing system according to claim 1, characterized in that: The complex domain spiking neural network layer splits the real and imaginary parts of the input complex signal, multiplies them with the real and imaginary parts of the complex weights respectively, and then forms new real and imaginary parts through addition and subtraction operations: The real part a and imaginary part b of the input complex signal are multiplied by the real part c and imaginary part d of the complex weight respectively, and then the difference is calculated to obtain the new real part a·cb·d; The real part a and imaginary part b of the input complex signal are multiplied by the real part c and imaginary part d of the complex weight respectively, and then the sum is obtained to obtain a new imaginary part a·d+c·b.

3. The radar signal processing system according to claim 1, characterized in that: The layer units include a complex pulse convolutional layer, a complex pulse linear layer, and a complex pulse fully connected layer.

4. The radar signal processing system according to claim 3, characterized in that: When used for radar emitter identification, the system consists of two complex pulse convolutional layers, two complex pulse linear layers, and one complex pulse fully connected layer connected in sequence.

5. The radar signal processing system according to claim 1, characterized in that: The system is suitable for radar resonance region echo identification, and / or radar pole feature identification, and / or radar radiation source identification.

6. The radar signal processing system according to claim 1, characterized in that: The method further includes a training step: using automatic mixed precision training, the hyperparameters used in the training are: the total number of iterations is 100 times, the batch training size is 30, and the learning rate is 0.001.

Images