CN117572376A - Low signal-to-noise ratio weak and small target radar echo signal recognition device and training recognition method - Google Patents

Abstract

The invention relates to the technical field of radio signal identification, and particularly discloses a low signal-to-noise ratio weak and small target radar echo signal identification device and a training identification method, wherein the radar echo signal identification device comprises a data block module; the rear end of the data block module is provided with a multichannel deep neural network module; the rear end of the multichannel deep neural network module is provided with a forward reasoning result storage module; the rear end of the forward reasoning result storage module is provided with a fusion reasoning module; the rear end of the fusion reasoning module is provided with a feedback training access module; the rear end of the feedback training access module is connected to the multichannel deep neural network module; the rear end of the multichannel deep neural network module is also provided with a classification model parameter storage module. By adopting the device and the training and identifying method, not only can a good classification model be obtained, but also an accurate classification and identifying result can be obtained, and echo data of strong noise with low signal-to-noise ratio and radar weak and small targets can be effectively distinguished.

Description

Low signal-to-noise ratio weak and small target radar echo signal recognition device and training recognition method

Technical Field

The invention belongs to the technical field of radio detection and positioning, and particularly relates to a low signal-to-noise ratio weak and small target radar echo signal recognition device and a training recognition method.

Background

The existing radio detection and positioning technology adopts radar echo recognition technology, namely, a recognition device is adopted to recognize an echo signal of a radar signal sent by a transmitter of the active radar and received by an active radar receiver. The existing radar echo signal identification method is mainly divided into two types, namely a radar target detection method based on Constant False Alarm (CFAR), and the method is based on estimation of background noise power in a radar echo unit to distinguish whether targets or noise are identified, so that the function of classifying and identifying radar targets and noise is achieved; the other is a radar target recognition detection method based on various algorithms of deep learning, and the method is to train the accumulated radar echo signals by the existing deep learning algorithm to obtain a trained radar echo signal classification model for classifying and recognizing targets and noise.

The low signal-to-noise ratio refers to low energy ratio of signal energy and noise, and the weak radar target refers to a target with small radar cross section area and weak reflected radar energy, such as an unmanned plane, a small aircraft, a stealth plane and the like. For the situation of low signal-to-noise ratio and weak target, the existing two radar echo identification technologies have respective defects.

Firstly, the detection effect of a radar target detection algorithm based on CFAR on a weak target is difficult to meet the use requirement under the condition of low signal-to-noise ratio, mainly because radar echo data of the weak target is weak, radar echo occupies a small distance unit, and because the signal-to-noise ratio is low and noise signal energy is high, the radar echo signal energy of the weak target is submerged in the noise energy, whether a received signal is a target echo signal or a noise signal cannot be judged by comparing the average power of the signal energy with the estimated average power of the noise, and further classification and identification of the noise signal and the radar echo signal of the target cannot be performed. The conditions of small targets and low signal-to-noise ratio become a major factor limiting the performance of CFAR-based radar target detection algorithms.

Secondly, the radar target recognition detection method based on the deep learning algorithm needs to be subjected to supervised training, has high requirements on training data, and is higher in the distinguishing property of the training data and the signal to noise ratio, and the accuracy of the trained classification model is higher. In general, the existing radar target recognition and detection method based on the deep learning algorithm is based on radar signal data under the condition of higher signal-to-noise ratio or no noise condition as a training data set, but the method is not in accordance with the actual condition, because the radar echo signal data collected under the actual condition often contains unbalanced noise data, and the condition that targets are weak targets and low signal-to-noise ratio often exists, and a classification model obtained by training the radar echo signal data with low signal-to-noise ratio is utilized, because the weak targets and the low signal-to-noise ratio data easily cause overfitting of the classification model in the training process, the accuracy performance of the classification model is low, and the classification recognition of the noise signals and the radar echo signals of the targets cannot be performed.

Disclosure of Invention

The invention provides a low signal-to-noise ratio weak target radar echo signal recognition device and a training recognition method, which are used for effectively training a data set formed by radar echo signal data of a detected radar weak target and a low signal-to-noise ratio under the background of low radar target signal detection.

The device and the method are used for accurately classifying and identifying the radar weak targets such as unmanned aerial vehicles, small aircrafts, stealth aircrafts and the like (the classifying and identifying refers to distinguishing certain radar weak targets from noise signals), real part and imaginary part time domain and frequency domain information of radar echo signal data and time domain and frequency domain information of radar echo signal data after pulse compression are simultaneously used as input of the classifying model by utilizing the classifying model trained by the invention, and are substituted into the trained classifying model to obtain accurate classifying and identifying results, and the results can effectively distinguish strong noise with low signal to noise ratio from echo data of the radar weak targets.

In order to achieve the above purpose, the present invention provides the following technical solutions: the radar echo signal identification device for the weak and small target with low signal-to-noise ratio comprises a data block module; the rear end of the data blocking module is provided with a multichannel deep neural network module; the rear end of the multichannel deep neural network module is provided with a forward reasoning result storage module; the rear end of the forward reasoning result storage module is provided with a fusion reasoning module; the rear end of the fusion reasoning module is provided with a feedback training access module; the rear end of the feedback training access module is connected to the multichannel deep neural network module; the rear end of the multichannel deep neural network module is also provided with a classification model parameter storage module.

The invention also provides a training method of the low signal-to-noise ratio weak target radar echo signal recognition device, which comprises the following training steps:

s1, inputting input data with labels into a data partitioning module to obtain partitioned training data;

s2, inputting the segmented training data into the multichannel deep neural network module in parallel, and obtaining a forward reasoning operation result through forward reasoning operation of a network structure and network parameters in the multichannel deep neural network module;

S3, saving forward reasoning operation results of each type of input data in batches into a forward reasoning result saving module, and saving the forward reasoning results of each batch of two-dimensional vectors by taking an average value in each dimension;

s4, inputting the saved average value of each batch forward reasoning result into a fusion reasoning module one by one, carrying out fusion reasoning on the average value of the forward reasoning result and a label result of input data to obtain a fusion reasoning result, and saving the fusion reasoning result;

s5, inputting the fusion reasoning results into the feedback training access module one by one, differencing the fusion reasoning results with the average value of the forward reasoning results to obtain one by one difference vector, substituting the difference vector into the multi-channel deep neural network model stored in the feedback training access module 5, and performing layer by layer reverse reasoning of a reverse reasoning algorithm between layers corresponding to the convolutional neural network to obtain new network model parameters of the multi-channel deep neural network model;

s6, inputting new network model parameters of the multichannel deep neural network model obtained by reverse reasoning of the feedback training access module into the multichannel deep neural network module for storage and update;

and S7, iterating the steps S2 to S6 until 100 times, and sending the network model parameters of the finally trained multichannel deep neural network model to a classification model parameter storage module for storage.

Preferably, the input data with the tag in the training method step S1 includes pulse-compressed radar echo signal time domain data, pulse-compressed radar echo signal frequency domain data, radar echo signal real part time domain data, radar echo signal imaginary part time domain data, and radar echo signal frequency domain data.

Preferably, the data blocking module in the training method step S1 performs normalized data blocking processing on data, and the specific processing method includes:

s1a, performing data length calculation on various types of input data, calculating the number of elements of various discretized vector data because the input data are all discretized vector data, and storing the number with the maximum number of elements;

s1b, filling each type of input data according to the number of the largest stored element number, namely, if one type of input data is smaller than the number of the elements of the longest stored discretization vector data, filling 0 for the number of the elements with phase difference before the input data, and then adding Gaussian white noise into the 0-obtained part, wherein the average value of the Gaussian white noise is 0, and the standard deviation is the average power of signals; through the steps, different types of input data with equal lengths can be obtained.

Preferably, in the training method step S2, the specific model of forward reasoning operation of the multichannel deep neural network module is designed as follows: and (3) carrying out parallel input on data corresponding to the data blocking module of each channel, and accessing the data input in parallel into three layers of a convolution pool layer of the convolution neural network, one layer of a hidden layer of the convolution neural network and one layer of an output layer of the convolution neural network, wherein the number of the maximum elements calculated and stored by the data blocking module is assumed to be N.

Preferably, the specific step of saving the forward reasoning result in batch by the forward reasoning result saving module in the training method step S3 is as follows:

s3a, a forward reasoning result storage module stores the label result of each training input data;

s3b, randomly extracting 10 data from training input data with the same label result, substituting each training input data into a multichannel deep neural network module, performing forward reasoning calculation, and averaging the obtained two-dimensional vector forward reasoning results of the 10 training input data in each dimension;

and S3c, saving the average value as a forward reasoning result of the 10 data extracted randomly, and saving the corresponding relation with the 10 training input data.

Preferably, the specific operation method of the fusion inference module in the training method step S4 is as follows:

s4a, inputting the average value of the forward reasoning results of 10 training input data randomly extracted from the same label result stored by the forward reasoning result storage module into the fusion reasoning module;

s4b, carrying out fusion reasoning based on PCR5 combination rules in the DSmT frame on the average value of the forward reasoning results and the label result of the stored input data, obtaining fusion reasoning results, storing the fusion reasoning results, and storing the correspondence with the grouped 10 training input data.

Preferably, in step S5, the random initial parameters of the multichannel deep neural network model stored in the feedback training access module are: random initial parameters of 16 different convolution kernels of the first convolution layer of the convolution neural network, random initial parameters of 32 different convolution kernels of the second convolution layer of the convolution neural network, random initial parameters of 16 different convolution kernels of the third convolution layer of the convolution neural network, random initial full connection parameters between each node of the third convolution layer of the convolution neural network pool layer and 100 nodes of the hidden layer of the convolution neural network, and random initial full connection parameters between 100 nodes of the hidden layer of the convolution neural network and 2 nodes of the output layer of the convolution neural network;

The specific operation method of the feedback training access module in the step S5 is as follows:

s5a, the average value of the fusion reasoning results of the 10 training input data of the groups stored by the fusion reasoning module is differenced with the average value of the forward reasoning results of the same 10 training input data stored by the forward reasoning result storage module, so that a difference vector corresponding to each group is obtained;

s5b, substituting the one-by-one difference vector corresponding to each group into a multichannel deep neural network model stored by a feedback training access module to perform layer-by-layer reverse reasoning of a reverse reasoning algorithm between layers corresponding to the convolutional neural network, namely performing layer-by-layer reverse reasoning of the reverse reasoning algorithm between layers corresponding to the convolutional neural network according to the sequence of convolutional neural network output layer to the convolutional neural network hidden layer, the convolutional neural network hidden layer to the convolutional neural network pooling layer III, the convolutional neural network pooling layer III to the convolutional neural network convolutional layer III, the convolutional neural network convolutional layer III to the convolutional neural network pooling layer II, the convolutional neural network pooling layer II to the convolutional neural network convolutional layer II, the convolutional neural network pooling layer I to the convolutional neural network convolutional layer I, so as to obtain new multichannel deep neural network model parameters, and storing and updating;

S5c, randomly selecting 10 pieces of input training data, and performing steps S5a to S5b;

and S5d, iterating the step S5c until 100 times, obtaining the network model parameters of the trained multichannel deep neural network model, and sending the network model parameters to a classification model parameter storage module for storage.

Preferably, the network model parameters of the trained multi-channel deep neural network model stored by the classification model parameter storage module in step S7 include: the method comprises the steps of training parameters of 16 different convolution kernels of a first convolution layer of the convolution neural network, training parameters of 32 different convolution kernels of a second convolution layer of the convolution neural network, training parameters of 16 different convolution kernels of a third convolution layer of the convolution neural network, training full-connection parameters between each node of a third pooling layer of the convolution neural network and 100 nodes of a hidden layer of the convolution neural network, and training full-connection parameters between 100 nodes of the hidden layer of the convolution neural network and 2 nodes of an output layer of the convolution neural network.

The invention also provides a recognition method of the low signal-to-noise ratio weak target radar echo signal recognition device, which comprises the following recognition steps:

c1, firstly, performing pulse compression on an echo signal of a radar to obtain time domain data and frequency domain data of the radar echo signal after pulse compression;

Substituting three-dimensional data, namely real part time domain data, imaginary part time domain data and frequency domain data of the radar echo signals which are not compressed by the pulse, and two-dimensional data in the C1 into a data block module to obtain five-dimensional radar input data to be identified, wherein the length of the five-dimensional radar input data is the maximum number N of elements calculated and stored by the data block module;

c3, substituting the five-dimensional radar input data to be identified with the length of the maximum number N of the elements calculated and stored by the data block module into a first convolution layer of the convolution neural network, a first pooling layer of the convolution neural network, a second convolution layer of the convolution neural network, a second pooling layer of the convolution neural network, a third convolution layer of the convolution neural network, a third pooling layer of the convolution neural network, a hidden layer of the convolution neural network and an output layer of the convolution neural network from left to right in sequence, and calculating according to model parameters stored by the classification model parameter storage module to obtain a two-dimensional vector result of the output layer of the convolution neural network;

and C4, comparing the two-dimensional vector results, outputting the signal as a target signal if the value representing the target signal is larger than the value representing the non-target signal, outputting the signal as a non-target signal if the value representing the non-target signal is larger than the value representing the target signal, and outputting the signal as an uncertain signal if the value representing the non-target signal is equal to the value representing the target signal.

Compared with the prior art, the invention has the beneficial effects that:

1. the method adopts a fusion reasoning step, is first applied in the radar echo identification field, can treat the forward reasoning result and the label result in the same level, namely, one result is not true, but in practice, under the condition of low signal-to-noise ratio, the classification of radar echo signals can be wrong, and the radar echo signals of different categories can possibly generate the condition of inter-category interleaving due to the influence of noise, so that the labels are sometimes biased, and the fusion reasoning is adopted, so that the training model reasoning capability can be considered to have certain authority, and can be used as a textbook type expert such as expert and label to discuss each other, thus obtaining the fusion result, and the comparison method of the prior neural network model and the label result is greatly improved;

2. in the aspect of training, a data set formed by radar echo signal data of a detected weak and small radar target and low signal-to-noise ratio can be effectively trained under the actual condition, real part and imaginary part time domain and frequency domain information containing amplitude and phase of the radar echo signal data are comprehensively utilized, meanwhile, the signal-to-noise ratio of the radar echo signal data is improved by combining pulse compression to obtain the time domain and frequency domain information of the radar echo signal data after pulse compression, multiple types of information are converted into multi-dimensional training data of a classification model, a classification model feedback training method avoiding overfitting is adopted in the training process, and a classification model with better classification performance is obtained through training;

3. In the aspect of identification, the method can accurately classify and identify radar weak and small targets such as unmanned aerial vehicles, small aircrafts, stealth aircrafts and the like (the classification and identification refers to distinguishing certain radar weak and small targets from noise signals), and the real part and the imaginary part of radar echo signal data, time domain and frequency domain information of radar echo signal data after pulse compression are simultaneously used as the input of a classification model by utilizing the classification model trained by the method, and are substituted into the trained classification model to obtain an accurate classification and identification result, and the result can effectively distinguish the strong noise with low signal to noise ratio and the echo data of the radar weak and small targets.

Other features of the present disclosure and its advantages will become apparent from the following detailed description of exemplary embodiments of the disclosure, which proceeds with reference to the accompanying drawings.

Drawings

In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the related art, the drawings that are required to be used in the embodiments or the related technical descriptions will be briefly described below, and it is apparent that the drawings in the following description are only embodiments of the present disclosure, and other drawings may be obtained from the provided drawings without inventive effort to those of ordinary skill in the art.

FIG. 1 is a diagram showing the overall structure and training flow of a radar echo signal recognition device according to the present invention;

FIG. 2 is a diagram showing a data blocking processing procedure of a data blocking module of the radar echo signal recognition device of the invention;

FIG. 3 is a schematic diagram of a model of a multi-channel deep neural network module of the radar echo signal recognition device of the present invention;

in the figure: 1. the system comprises a data block module, a multichannel deep neural network module, a forward reasoning result storage module, a fusion reasoning module, a feedback training access module, a classification model parameter storage module and a classification model parameter storage module.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. The following description of at least one exemplary embodiment is merely exemplary in nature and is in no way intended to limit the invention, its application, or uses. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

The relative arrangement of the components and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present invention unless it is specifically stated otherwise. Meanwhile, it should be understood that the sizes of the respective parts shown in the drawings are not drawn in actual scale for convenience of description. Techniques, methods, and apparatus known to one of ordinary skill in the relevant art may not be discussed in detail, but are intended to be part of the specification where appropriate. In all examples shown and discussed herein, any specific values should be construed as merely illustrative, and not a limitation. Thus, other examples of the exemplary embodiments may have different values. It should be noted that: like reference numerals and letters denote like items in the following figures, and thus once an item is defined in one figure, no further discussion thereof is necessary in subsequent figures.

Spatially relative terms, such as "above … …," "above … …," "upper surface at … …," "above," and the like, may be used herein for ease of description to describe one device or feature's spatial location relative to another device or feature as illustrated in the figures. It will be understood that the spatially relative terms are intended to encompass different orientations in use or operation in addition to the orientation depicted in the figures. For example, if the device in the figures is turned over, elements described as "above" or "over" other devices or structures would then be oriented "below" or "beneath" the other devices or structures. Thus, the exemplary term "above … …" may include both orientations of "above … …" and "below … …". The device may also be positioned in other different ways (rotated 90 degrees or at other orientations) and the spatially relative descriptors used herein interpreted accordingly.

Referring to fig. 1-3, the present invention provides a technical solution: the low signal-to-noise ratio weak target radar echo signal recognition device is characterized in that all modules are connected according to the setting mode shown in fig. 1, namely a data block module 1, a multichannel deep neural network module 2, a forward reasoning result storage module 3, a fusion reasoning module 4 and a feedback training access module 5 are sequentially arranged, the feedback training access module 5 is reversely connected back to the multichannel deep neural network module 2, and the multichannel deep neural network module 2 is independently connected with a classification model parameter storage module 6 downwards.

Wherein in certain embodiments:

the data blocking module 1 is a data input and data arrangement device based on a USB or other data transmission interfaces, a CPU and a hard disk, can read different types of data from an external storage device through the USB or other data transmission interfaces, find and import the data into an internal storage hard disk of the whole training device, arrange the data through instructions of the CPU while importing the data into the internal storage hard disk, calculate the longest length of the uneven data, then fill the data according to the longest length, and then store the data in the internal storage hard disk;

The multichannel deep neural network module 2 is data processing equipment based on a GPU, a memory and a hard disk, the stored multidimensional data is read into the memory through the hard disk, then the data is sent into the GPU through the memory to forward infer a multichannel deep neural network model, a forward reasoning result is obtained, and the reasoning result is sent into the memory for storage;

the forward reasoning result storage module 3 is a memory device based on a memory and a hard disk, namely, the forward reasoning result calculated in the memory by the multichannel deep neural network module 2 is exported from the memory to the hard disk for storage;

the fusion reasoning module 4 is equipment based on a hard disk, a memory and a CPU, reads the forward reasoning result, the label result and the DSmT related combination rule instruction through the hard disk and reads the forward reasoning result and the label result in the memory into the memory, performs operation based on the DSmT related combination rule instruction through the CPU to obtain a fusion reasoning result, and then outputs the fusion reasoning result in the memory to the hard disk for storage;

the feedback training access module 5 is a data processing device based on a GPU, a memory and a hard disk, reads the stored multidimensional data into the memory through the hard disk, then sends the data into the GPU through the memory to reversely infer a multichannel deep neural network model to obtain a reverse reasoning result, and then sends the reasoning result into the memory for storage;

The classification model parameter storage module 6 is a storage device based on a memory and a hard disk, namely, the reverse reasoning result calculated in the memory by the feedback training access module 5 is exported from the memory to the hard disk for storage.

Before the radar echo signal identification device with the low signal-to-noise ratio and the weak target is used for identifying radar echo signals, the device needs to be trained, namely, a radar echo signal classification model is trained under the condition of the low signal-to-noise ratio and the weak target, the training method is suitable for training the radar echo signal classification model under the condition that the actually acquired radar echo signal data has low signal-to-noise ratio and the target is the weak target, the radar echo signal classification model with high training precision and difficult to enter a fitting state can be obtained, and then the radar echo signal is identified by the identification method.

The training method of the classification model is that the module design sequence of the signal identification device executes step-by-step training steps, firstly, input data of training is to be determined, wherein the input data comprises radar echo signal time domain data after pulse compression, radar echo signal frequency domain data after pulse compression, radar echo signal real part time domain data, radar echo signal imaginary part time domain data and radar echo signal frequency domain data.

The specific training method comprises the following training steps:

s1, inputting input data with labels (radar echo signal time domain data after pulse compression, radar echo signal frequency domain data after pulse compression, radar echo signal real part time domain data, radar echo signal imaginary part time domain data and radar echo signal frequency domain data) into a data blocking module 1 to obtain blocked training data;

referring to the specific method of fig. 2, the specific method of the data partitioning module 1 of the classification model training method includes:

s1a, performing data length calculation on various types of input data, calculating the number of elements of various discretized vector data because the input data are all discretized vector data, and storing the number with the maximum number of elements;

s1b, filling each type of input data according to the number of the largest stored element number, namely, if one type of input data is smaller than the number of the elements of the longest stored discretization vector data, filling 0 for the number of the elements with phase difference before the input data, and then adding Gaussian white noise into the 0-obtained part, wherein the average value of the Gaussian white noise is 0, and the standard deviation is the average power of signals; through the steps, different types of input data with equal lengths can be obtained.

S2, inputting the segmented training data into the multichannel deep neural network module 2 in parallel, and obtaining a forward reasoning operation result through forward reasoning operation of a network structure and network parameters in the multichannel deep neural network module 2;

referring to fig. 3, the depth and the computing performance of the deep neural network model are comprehensively considered, a specific model of the multi-channel deep neural network module 2 is designed to be that data of the data block module 1 corresponding to each channel are input in parallel, and the data input in parallel is accessed into three layers of a convolution pool layer, a convolution neural network hiding layer and a convolution neural network output layer of the convolution neural network. The data blocking module 1 is assumed to calculate that the number of the stored elements is the largest number of N.

S3, saving forward reasoning operation results of each type of input data in batches into a forward reasoning result saving module 3, and saving the forward reasoning results of each batch of two-dimensional vectors by taking an average value in each dimension;

the forward direction reasoning result storage module 3 completes the storage of the forward direction reasoning result of the multichannel deep neural network module 2, and the forward direction reasoning result storage process is as follows:

s3a, a forward reasoning result storage module 3 stores the label result of each training input data;

S3b, randomly extracting 10 data from training input data with the same label result, substituting each training input data into the multichannel deep neural network module 2, performing forward reasoning calculation, and averaging the obtained two-dimensional vector forward reasoning results of the 10 training input data in each dimension;

and S3c, saving the average value as a forward reasoning result of the 10 data extracted randomly, and saving the corresponding relation with the 10 training input data.

S4, inputting the average value of the forward reasoning results of each type of batch into a fusion reasoning module 4 one by one, carrying out fusion reasoning on the average value of the forward reasoning results and the label result of the input data, and obtaining and storing the fusion reasoning results;

the specific operation method of the fusion reasoning module 4 is as follows:

s4a, inputting the average value of the forward reasoning results of 10 training input data randomly extracted from the same label result stored in the forward reasoning result storage module 3 into the fusion reasoning module 4;

s4b, carrying out fusion reasoning based on PCR5 combination rules in the DSmT frame on the average value of the forward reasoning results and the label result of the stored input data, obtaining fusion reasoning results, storing the fusion reasoning results, and storing the correspondence with the grouped 10 training input data.

S5, inputting the fusion reasoning results into the feedback training access module 5 one by one, differencing the fusion reasoning results with the average value of the forward reasoning results to obtain one by one difference vector, substituting the difference vector into a multi-channel deep neural network model stored in the feedback training access module 5, and performing layer by layer reverse reasoning of a reverse reasoning algorithm between layers corresponding to the convolutional neural network to obtain new network model parameters of the multi-channel deep neural network model;

referring to fig. 3, in step S5, the random initial parameters of the multichannel deep neural network model stored in the feedback training access module 5 are: random initial parameters of 16 different convolution kernels of the first convolution layer of the convolution neural network, random initial parameters of 32 different convolution kernels of the second convolution layer of the convolution neural network, random initial parameters of 16 different convolution kernels of the third convolution layer of the convolution neural network, random initial full connection parameters between each node of the third convolution layer of the convolution neural network pool layer and 100 nodes of the hidden layer of the convolution neural network, and random initial full connection parameters between 100 nodes of the hidden layer of the convolution neural network and 2 nodes of the output layer of the convolution neural network;

the specific operation method of the feedback training access module 5 in step S5 is as follows:

S5a, the average value of the fusion reasoning results of the 10 training input data of the group stored by the fusion reasoning module 4 is differenced with the average value of the forward reasoning results of the same 10 training input data stored by the forward reasoning result storage module 3, so that a difference vector corresponding to each group is obtained;

s5b, substituting the one-by-one difference vector corresponding to each group into a multi-channel deep neural network model stored by the feedback training access module 5 to perform layer-by-layer reverse reasoning of a reverse reasoning algorithm between layers corresponding to the convolutional neural network, namely performing layer-by-layer reverse reasoning of the reverse reasoning algorithm between layers corresponding to the convolutional neural network according to the sequence of the output layer to the convolutional neural network hiding layer, the convolutional neural network hiding layer to the convolutional neural network pooling layer III, the convolutional neural network pooling layer III to the convolutional neural network convolutional layer III, the convolutional neural network pooling layer III to the convolutional neural network pooling layer II, the convolutional neural network pooling layer II to the convolutional neural network convolutional layer II, the convolutional neural network pooling layer I to the convolutional neural network pooling layer I, obtaining new multi-channel deep neural network model parameters, and storing and updating;

S5c, randomly selecting 10 pieces of input training data, and performing steps S5a to S5b;

and S5d, iterating the step S5c until 100 times, obtaining the network model parameters of the trained multichannel deep neural network model, and sending the network model parameters to the classification model parameter storage module 6 for storage.

S6, inputting new network model parameters of the multichannel deep neural network model obtained by reverse reasoning of the feedback training access module 5 into the multichannel deep neural network module 2 for storage and update.

And S7, iterating the steps S2 to S6 until 100 times, and sending the network model parameters of the finally trained multichannel deep neural network model to a classification model parameter storage module 6 for storage so as to facilitate subsequent identification and adoption.

Referring to fig. 3, the network model parameters of the trained multi-channel deep neural network model stored by the classification model parameter storage module 6 include: the method comprises the steps of training parameters of 16 different convolution kernels of a first convolution layer of the convolution neural network, training parameters of 32 different convolution kernels of a second convolution layer of the convolution neural network, training parameters of 16 different convolution kernels of a third convolution layer of the convolution neural network, training full-connection parameters between each node of a third pooling layer of the convolution neural network and 100 nodes of a hidden layer of the convolution neural network, and training full-connection parameters between 100 nodes of the hidden layer of the convolution neural network and 2 nodes of an output layer of the convolution neural network.

After the foregoing training method steps S1-S7, the present invention obtains a radar echo signal classification model under the condition of a low signal-to-noise ratio and weak target, where the model refers to fig. 3, and the classification model is used as a classification recognition method of the present invention to recognize radar echo signals.

Under the condition of low signal-to-noise ratio and weak object, radar echo signal energy is low and is easy to be submerged by noise, but radar echo signal energy can be accumulated through pulse compressed radar echo signal time domain data and frequency domain data, amplification is carried out on amplitude, but accumulation of existing pulse compressed signals can cause loss of phase information, the effect of radar echo signal classification is lost, and by receiving radar echo signal real part time domain data, imaginary part time domain data and frequency domain data and taking the five types of data as parallel multidimensional data to carry out simultaneous complementary processing, the loss of radar echo signal phase information of the traditional pulse compression method can be avoided, the accuracy of radar echo signal classification can be improved, the detection rate of weak and small objects can be improved under the condition of low signal-to-noise ratio radar weak and small objects, the detection rate of the low signal-to-noise ratio radar weak and small objects can be still found in time under the background of large noise, and the influence of weak and weak object signal energy is reduced.

A method for identifying a classification model completed by training, comprising the following specific steps:

c1, firstly, performing pulse compression on an echo signal of a radar to obtain time domain data and frequency domain data of the radar echo signal after pulse compression;

substituting three-dimensional data, namely real part time domain data, imaginary part time domain data and frequency domain data of the radar echo signals which are not compressed by the pulse, and two-dimensional data in the C1 into the data block module 1 to obtain five-dimensional radar input data to be identified, wherein the length of the five-dimensional radar input data is the maximum number N of elements calculated and stored by the data block module 1;

c3, substituting the five-dimensional radar input data to be identified with the length of the maximum number N of the elements calculated and stored by the data blocking module 1 into a first convolution layer of the convolution neural network, a first pooling layer of the convolution neural network, a second convolution layer of the convolution neural network, a second pooling layer of the convolution neural network, a third convolution layer of the convolution neural network, a third pooling layer of the convolution neural network, a hidden layer of the convolution neural network and an output layer of the convolution neural network from left to right in sequence, and calculating according to the model parameters stored by the classification model parameter storage module 6 to obtain a two-dimensional vector result of the output layer of the convolution neural network;

And C4, comparing the two-dimensional vector results, outputting the signal as a target signal if the value representing the target signal is larger than the value representing the non-target signal, outputting the signal as a non-target signal if the value representing the non-target signal is larger than the value representing the target signal, and outputting the signal as an uncertain signal if the value representing the non-target signal is equal to the value representing the target signal.

The working principle of the training method is as follows:

1. according to the invention, radar echo data before and after pulse compression, real part and imaginary part are taken as parallel multidimensional data to be used as training input data for training, so that the signal energy of a low-signal-to-noise radar weak target is enhanced, the loss of a target signal with small energy is avoided, the phase information of radar target echoes of the radar real part and the imaginary part data is reserved, the comprehensiveness of radar echo signal data characteristics is reserved to the greatest extent, and comprehensive input characteristics are reserved for training of a classification model;

2. the fusion reasoning module 4 considers the special background condition of a low signal-to-noise ratio radar weak target, namely the tag result of a radar echo signal is possibly wrong or is data which is subjected to fitting, namely multiple types of data are subjected to individual interleaving due to the influence of noise, so that forward reasoning is carried out through randomly extracted grouping data, the average value of the forward reasoning result of the grouping data is taken as the forward reasoning result, the influence of individual wrong data on the result is avoided to a great extent, then the forward reasoning result is taken as a result with certain credibility with the tag result, fusion reasoning based on the PCR5 rule under a DSmT frame is carried out on the two results, the influence of wrong data and overfitting data can be avoided, the contradiction between the wrong data and the overfitting data can also be avoided, and a comparatively scientific fusion reasoning result is obtained;

3. The feedback training access module 5 of the invention considers the difference between the fusion reasoning result and the forward reasoning result, wherein the label and the forward reasoning result are fused, the label result is considered, the reasoning result of the training model is also referred, the error of the label result is avoided, the problem that the difference between the forward reasoning result and the label result is subjected to fitting, the trained classification model is subjected to fitting, and the classification model is difficult to achieve higher precision.

The working principle of the identification method of the invention is as follows:

1. the radar echo data before and after pulse compression and the radar echo data of the real part and the imaginary part are taken as parallel multidimensional data to be taken as classification model identification data to be detected for identification, so that the signal energy of a low signal-to-noise radar weak target is enhanced, the loss of a target signal with small energy is avoided, the phase information of the radar target echo of the radar real part and the radar echo data of the imaginary part is reserved, the comprehensiveness of radar echo signal data characteristics is reserved to the greatest extent, and comprehensive input characteristics are reserved for identification of the classification model;

2. the multichannel neural network model parameters obtained by the training method can avoid the problem of wrong label result data or data over-fitting under the condition of weak and small targets of the low signal-to-noise ratio radar, and obtain accurate radar target classification and identification results.

Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims (10)

1. The low signal-to-noise ratio weak target radar echo signal identification device is characterized in that: the radar echo signal identification device comprises a data blocking module (1); the rear end of the data blocking module (1) is provided with a multichannel deep neural network module (2); the rear end of the multichannel deep neural network module (2) is provided with a forward reasoning result storage module (3); the rear end of the forward reasoning result storage module (3) is provided with a fusion reasoning module (4); the rear end of the fusion reasoning module (4) is provided with a feedback training access module (5); the rear end of the feedback training access module (5) is connected to the multichannel deep neural network module (2); the rear end of the multichannel deep neural network module (2) is also provided with a classification model parameter storage module (6).

2. The training method of the low signal-to-noise ratio weak target radar echo signal identification device according to claim 1, wherein: the training method comprises the following training steps:

S1, inputting input data with labels into a data partitioning module (1) to obtain partitioned training data;

s2, inputting the segmented training data into the multichannel deep neural network module (2) in parallel, and obtaining a forward reasoning operation result through forward reasoning operation of a network structure and network parameters in the multichannel deep neural network module (2);

s3, saving forward reasoning operation results of each type of input data in batches into a forward reasoning result saving module (3), and saving average values of forward reasoning results of each batch of two-dimensional vectors in each dimension;

s4, inputting the saved average value of each batch forward reasoning result into a fusion reasoning module (4) one by one, carrying out fusion reasoning on the average value of the forward reasoning result and a label result of input data, obtaining a fusion reasoning result and saving the fusion reasoning result;

s5, inputting the fusion reasoning results into a feedback training access module (5) one by one, making a difference between the fusion reasoning results and the average value of the forward reasoning results to obtain one by one difference vector, substituting the difference vector into a multi-channel deep neural network model stored in the feedback training access module (5), and performing layer-by-layer reverse reasoning of a reverse reasoning algorithm between layers corresponding to the convolutional neural network to obtain new network model parameters of the multi-channel deep neural network model;

S6, inputting new network model parameters of the multichannel deep neural network model obtained by reverse reasoning of the feedback training access module (5) into the multichannel deep neural network module (2) for storage and update;

and S7, iterating the steps S2 to S6 until 100 times, and sending the network model parameters of the finally trained multichannel deep neural network model to a classification model parameter storage module (6) for storage.

3. The training method of the low signal-to-noise ratio weak target radar echo signal identification device according to claim 2, wherein the training method comprises the following steps: the input data with the tag in the step S1 includes pulse-compressed radar echo signal time domain data, pulse-compressed radar echo signal frequency domain data, radar echo signal real part time domain data, radar echo signal imaginary part time domain data, and radar echo signal frequency domain data.

4. The training method of the low signal-to-noise ratio weak target radar echo signal identification device according to claim 2, wherein the training method comprises the following steps: the data blocking module (1) in the step S1 performs normalized data blocking processing on the data, and the specific processing method comprises the following steps:

s1a, performing data length calculation on various types of input data, calculating the number of elements of various discretized vector data because the input data are all discretized vector data, and storing the number with the maximum number of elements;

S1b, filling each type of input data according to the number of the largest stored element number, namely, if one type of input data is smaller than the number of the elements of the longest stored discretization vector data, filling 0 for the number of the elements with phase difference before the input data, and then adding Gaussian white noise into the 0-obtained part, wherein the average value of the Gaussian white noise is 0, and the standard deviation is the average power of signals; through the steps, different types of input data with equal lengths can be obtained.

5. The training method of the low signal-to-noise ratio weak target radar echo signal identification device according to claim 2, wherein the training method comprises the following steps: the specific model of forward reasoning operation of the multichannel deep neural network module (2) in the step S2 is designed as follows: and (3) carrying out parallel input on data corresponding to the data blocking module (1) of each channel, and accessing the data input in parallel into three layers of a convolution pool layer of a convolution neural network, one layer of a hidden layer of the convolution neural network and one layer of an output layer of the convolution neural network, wherein the number of the maximum elements calculated and stored by the data blocking module (1) is assumed to be N.

6. The training method of the low signal-to-noise ratio weak target radar echo signal identification device according to claim 2, wherein the training method comprises the following steps: the specific steps of the forward reasoning result storage module (3) in the step S3 for storing the forward reasoning operation result in batches are as follows:

S3a, a forward reasoning result storage module (3) stores the label result of each training input data;

s3b, randomly extracting 10 data from training input data of the same label result, substituting each training input data into a multichannel deep neural network module (2), performing forward reasoning calculation, and averaging the obtained two-dimensional vector forward reasoning results of the 10 training input data in each dimension;

and S3c, saving the average value as a forward reasoning result of the 10 data extracted randomly, and saving the corresponding relation with the 10 training input data.

7. The training method of the low signal-to-noise ratio weak target radar echo signal identification device according to claim 2, wherein the training method comprises the following steps: the specific operation method of the fusion reasoning module (4) in the step S4 is as follows:

s4a, inputting the average value of the forward reasoning results of 10 training input data randomly extracted from the same label result stored in the forward reasoning result storage module (3) into the fusion reasoning module (4);

s4b, carrying out fusion reasoning based on PCR5 combination rules in the DSmT frame on the average value of the forward reasoning results and the label result of the stored input data, obtaining fusion reasoning results, storing the fusion reasoning results, and storing the correspondence with the grouped 10 training input data.

8. The training method of the low signal-to-noise ratio weak target radar echo signal identification device according to claim 2, wherein the training method comprises the following steps: in the step S5, the random initial parameters of the multichannel deep neural network model stored in the feedback training access module (5) are as follows: random initial parameters of 16 different convolution kernels of the first convolution layer of the convolution neural network, random initial parameters of 32 different convolution kernels of the second convolution layer of the convolution neural network, random initial parameters of 16 different convolution kernels of the third convolution layer of the convolution neural network, random initial full connection parameters between each node of the third convolution layer of the convolution neural network pool layer and 100 nodes of the hidden layer of the convolution neural network, and random initial full connection parameters between 100 nodes of the hidden layer of the convolution neural network and 2 nodes of the output layer of the convolution neural network;

the specific operation method of the feedback training access module (5) in the step S5 is as follows:

s5a, the average value of the fusion reasoning results of the 10 training input data of the group stored by the fusion reasoning module (4) is differenced with the average value of the forward reasoning results of the same 10 training input data stored by the forward reasoning result storage module (3), so that a difference vector corresponding to each group is obtained;

S5b, substituting the one-by-one difference vector corresponding to each group into a multichannel deep neural network model stored by a feedback training access module (5) to perform layer-by-layer reverse reasoning of a reverse reasoning algorithm between layers corresponding to the convolutional neural network, namely performing layer-by-layer reverse reasoning of the reverse reasoning algorithm between layers corresponding to the convolutional neural network according to the sequence of the convolutional neural network output layer to the convolutional neural network hidden layer, the convolutional neural network hidden layer to the convolutional neural network pooling layer III, the convolutional neural network pooling layer III to the convolutional neural network convolutional layer III, the convolutional neural network convolutional layer III to the convolutional neural network pooling layer II, the convolutional neural network pooling layer II to the convolutional neural network convolutional layer II, the convolutional neural network pooling layer I to the convolutional neural network pooling layer I, so as to obtain new multichannel deep neural network model parameters, and storing and updating;

s5c, randomly selecting 10 pieces of input training data, and performing steps S5a to S5b;

and S5d, iterating the step S5c until 100 times, obtaining the network model parameters of the trained multichannel deep neural network model, and sending the network model parameters to a classification model parameter storage module (6) for storage.

9. The training method of the low signal-to-noise ratio weak target radar echo signal identification device according to claim 2, wherein the training method comprises the following steps: the network model parameters of the trained multichannel deep neural network model stored by the classification model parameter storage module (6) in the step S7 include: the method comprises the steps of training parameters of 16 different convolution kernels of a first convolution layer of the convolution neural network, training parameters of 32 different convolution kernels of a second convolution layer of the convolution neural network, training parameters of 16 different convolution kernels of a third convolution layer of the convolution neural network, training full-connection parameters between each node of a third pooling layer of the convolution neural network and 100 nodes of a hidden layer of the convolution neural network, and training full-connection parameters between 100 nodes of the hidden layer of the convolution neural network and 2 nodes of an output layer of the convolution neural network.

10. The method for identifying the low signal-to-noise ratio weak target radar echo signal identifying device according to claim 1, wherein the method comprises the following steps: the identification method comprises the following identification steps:

c1, firstly, performing pulse compression on an echo signal of a radar to obtain time domain data and frequency domain data of the radar echo signal after pulse compression;

Substituting three-dimensional data, namely real part time domain data, imaginary part time domain data and frequency domain data of radar echo signals which are not compressed by pulses, and two-dimensional data in the C1 into a data block module (1) together to obtain five-dimensional radar input data to be identified, wherein the length of the five-dimensional radar input data is the maximum number N of elements calculated and stored by the data block module (1);

c3, substituting the five-dimensional radar input data to be identified with the length of the maximum number N of the elements calculated and stored by the data blocking module (1) into a first convolution neural network convolution layer, a first convolution neural network pooling layer, a second convolution neural network pooling layer, a third convolution neural network convolution layer, a third convolution neural network pooling layer, a third convolution neural network hiding layer and a convolution neural network output layer from left to right in sequence, and calculating according to model parameters stored by the classification model parameter storage module (6) to obtain a two-dimensional vector result of the convolution neural network output layer;

and C4, comparing the two-dimensional vector results, outputting the signal as a target signal if the value representing the target signal is larger than the value representing the non-target signal, outputting the signal as a non-target signal if the value representing the non-target signal is larger than the value representing the target signal, and outputting the signal as an uncertain signal if the value representing the non-target signal is equal to the value representing the target signal.