Low signal-to-noise ratio radar emitter signal intra-pulse identification device and training identification method thereof
Through the deep neural network module and the fusion reasoning step, the problem of low recognition accuracy of radar radiation source signals under low signal-to-noise ratio is solved, and the accurate classification of radar radiation source signals is achieved is achieved, thereby improving the recognition accuracy.
Patent Information
- Application Number
- CN202410056328.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-16
- Publication Date
- 2025-01-10
- Estimated Expiration
- 2044-01-16
AI Technical Summary
The existing intra-vibic classification recognition methods for radar radiation source signals are not highly accurate in the case of low signal-to-noise ratio. This is mainly because the training data are prone to confusion, which leads to overfitting of deep learning algorithms training and reducing the recognition accuracy.
The deep neural network module is used for training, and the access module is trained through data chunking, storage of forward inference results, fusion inference and feedback training, avoid overfitting, and training is achieved for classification models with better classification performance, and using fusion inference steps to improve recognition accuracy.
Effective training of classification models under low signal-to-noise ratio improves the accurate classification and recognition ability of radar radiation source signal types, reduces the impact of noise interference, and improves the recognition accuracy.
Smart Images
Figure CN117851827B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of radio detection and positioning, and in particular relates to a low signal-to-noise ratio radar radiation source signal intra-pulse recognition device and a training recognition method thereof. Background Art
[0002] The existing radio detection and positioning technology is implemented by receiving and identifying signals by radar, and the radar emitter signal intra-pulse classification and identification device and training and identification method refer to a device for classifying and identifying the radiation source signal emitted by a non-cooperative radar received by a passive radar receiver, and its training and identification method. The existing radar emitter signal intra-pulse classification and identification methods are mainly divided into two types, one is the identification method based on traditional machine learning, and the other is the identification method based on deep learning. Low signal-to-noise ratio refers to the low ratio of signal energy to noise energy. The above two identification methods have their own defects in signal recognition under low signal-to-noise ratio conditions.
[0003] First, the recognition method based on traditional machine learning is generally divided into two steps: the first step is to extract features using manual methods; the second step is to use traditional machine learning methods to build a classifier for recognition. This method based on machine learning often has the following problems in application: (1) The quality of the manually designed feature extraction algorithm usually depends on the professional knowledge and experience level of the practitioner. The extracted features are often highly targeted. Once the data set changes, the performance of the original feature extraction algorithm will be greatly reduced, making the method less transferable; (2) The extracted features are often "shallow features" and it is impossible to extract more abstract "advanced" features. Moreover, the quality of the features cannot be determined in advance and can only be tried through a large number of experiments; (3) It is very sensitive to noise. Under low signal-to-noise ratio conditions, it is usually difficult to extract effective features, which makes it difficult to meet current needs.
[0004] Secondly, the recognition method based on deep learning mainly uses deep learning algorithms for recognition, and the main idea of deep learning algorithms is to use time-frequency analysis methods to transform the original radar signal into radar time-frequency images. At this time, the signal recognition problem is transformed into an image recognition problem. The deep learning method integrates feature extraction and recognition, in which feature extraction is to extract decisive representations from data by constructing a loss function and using gradient descent and error back propagation algorithms to complete the classification of targets. The deep learning method is a more general, automatic, and more intelligent method. Through the combination of multiple layers of nonlinear modules, it can automatically learn the "good" and "advanced" features in the data set, and in the continuous convolution process, it amplifies the features in the input that are beneficial to target classification, while suppressing irrelevant noise. Therefore, the method based on deep learning has strong noise suppression and feature learning capabilities, and its application has greatly improved the recognition effect of radar radiation source signals. However, the existing deep learning-based radar emitter signal pulse classification and recognition methods have low recognition accuracy in low signal-to-noise ratio conditions. The reason is that the existing methods are trained on radar emitter signal data sets under different signal-to-noise ratio conditions or radar emitter signal data sets under high signal-to-noise ratio conditions, in the hope of obtaining a radar emitter signal classification model applicable to any signal-to-noise ratio. However, for such mixed training data with different signal-to-noise ratio data, the noise signal energy under low signal-to-noise ratio conditions is higher, making the training data easy to be confused with each other. If the same method is used to train with the emitter signal data under high signal-to-noise ratio conditions, it is bound to lead to overfitting of the deep learning algorithm training, thereby reducing the recognition accuracy of radar emitter signals under low signal-to-noise ratio conditions. As for the case of training with high signal-to-noise ratio radiation source data, since the training data itself only considers the high signal-to-noise ratio case, the classification accuracy for the high signal-to-noise ratio data can be higher. However, for the low signal-to-noise ratio case, due to the interference of large noise energy, the classification model cannot effectively classify the low signal-to-noise ratio data, thereby reducing the accuracy of radar radiation source signal recognition under low signal-to-noise ratio conditions. Summary of the invention
[0005] The present invention provides a low signal-to-noise ratio radar radiation source signal pulse identification device and a training identification method thereof. The use of the identification device and method can achieve the following purposes:
[0006] First, in order to effectively train the data set formed by the detected low signal-to-noise ratio radar emitter signal in actual situations, only the data of the low signal-to-noise ratio radar emitter signal is used, the data is converted into the time-frequency domain to obtain the radar emitter time-frequency data, and the deep learning algorithm is trained to obtain the parameters of the classification model. In the training process, the classification model feedback training method is used to avoid overfitting, and a classification model with better classification performance is trained.
[0007] Secondly, in order to be able to accurately classify and identify the radar radiation source signal type under low signal-to-noise ratio conditions under actual circumstances (the classification and identification referred to here refers to distinguishing the various types of radar radiation source signal types), by using the classification model trained by the present invention, the low signal-to-noise ratio radar radiation source signal data is used as the input of the classification model, and substituted into the trained classification model to obtain an accurate classification and identification result, which can effectively distinguish the radar radiation source signal type with a low signal-to-noise ratio.
[0008] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a low signal-to-noise ratio radar radiation source signal intra-pulse identification device, the radar radiation source signal intra-pulse identification device comprising a data segmentation module; a deep neural network module is provided at the back end of the data segmentation module; a forward reasoning result storage module is provided at the back end of the deep neural network module; a fusion reasoning module is provided at the back end of the forward reasoning result storage module; a feedback training access module is provided at the back end of the fusion reasoning module; the feedback training access module is connected to the deep neural network module at the back end; and a classification model parameter storage module is also provided at the back end of the deep neural network module.
[0009] The present invention also provides a training method for a low signal-to-noise ratio radar radiation source signal intra-pulse identification device, comprising the following training steps:
[0010] S1, inputting the time-frequency data of the radar radiation source signal with labels into the data segmentation module as input data to obtain segmented training data;
[0011] S2, inputting the divided training data into the deep neural network module, and obtaining the forward reasoning result through the forward reasoning operation of the network structure and network parameters in the deep neural network module;
[0012] S3, saving the forward reasoning operation results of each type of input data in batches into the forward reasoning result saving module, and taking the average value of each batch of results in each dimension for saving;
[0013] S4, inputting the saved average values of the forward reasoning results of each category into the fusion reasoning module one by one, performing fusion reasoning on the average values of the forward reasoning results and the label results of the input data, obtaining the fusion reasoning results and saving them;
[0014] S5, input the fusion reasoning results into the feedback training access module one by one, subtract the average value of the fusion reasoning results from the forward reasoning results to obtain difference vectors one by one, substitute the difference vectors into the deep neural network model stored in the feedback training access module, perform reverse reasoning layer by layer of the reverse reasoning algorithm between the layers corresponding to the convolutional neural network, and obtain new network model parameters of the deep neural network model;
[0015] S6, inputting new network model parameters of the deep neural network model obtained by reverse reasoning of the feedback training access module into the deep neural network module for saving and updating;
[0016] S7, iterate steps S2 to S6 up to 100 times, and send the network model parameters of the finally trained deep neural network model to the classification model parameter saving module for storage.
[0017] Preferably, in step S1 of the training method, the data segmentation module performs normalized data segmentation processing on the input data with labels, and the specific processing method includes:
[0018] S1a, calculate the two-dimensional data size of the time-frequency input data of radar radiation source signals of various models. Since the input data are all discretized two-dimensional data, calculate the number of rows and columns of various two-dimensional data, and save the maximum number of rows and columns of various types of input data;
[0019] S1b, each type of input data is padded according to the maximum number of rows and columns stored, that is, if a certain type of input data is less than the number of rows and columns of the longest stored two-dimensional data, the number of rows and columns that differ is set to 0 to padded in front of the rows and columns of the input data, and then Gaussian white noise is added to the part where it is taken as 0. The mean of the Gaussian white noise is 0, and the standard deviation is the average power of the signal. Through the above steps, different types of radar radiation source signal time-frequency input data with equal number of rows and columns can be obtained.
[0020] Preferably, the specific model design of the forward reasoning operation of the deep neural network module in step S2 of the training method is: corresponding to the data input of the data segmentation module, the input data is connected to three convolutional pooling layers of the convolutional neural network, one hidden layer of the convolutional neural network, and one output layer of the convolutional neural network.
[0021] Preferably, the specific steps of the forward reasoning result saving module in step S3 of the training method for saving the forward reasoning operation results in batches are:
[0022] S3a, the forward reasoning result saving module saves the label result of each training input data;
[0023] S3b, randomly extract 10 data from the training input data with the same label result, substitute each training input data into the deep neural network module, perform forward reasoning calculation, and take the average value of the N-dimensional vector forward reasoning results of the 10 training input data in each dimension, where N represents the number of radar radiation source types;
[0024] S3c, saves the average value as the forward inference result of 10 randomly selected data, and saves the corresponding relationship with the 10 training input data.
[0025] Preferably, the specific operation method of the fusion reasoning module in step S4 of the training method is:
[0026] S4a, inputting the average value of the forward reasoning results of 10 randomly selected training input data with the same label results saved by the forward reasoning result saving module into the fusion reasoning module;
[0027] S4b, the average value of the forward reasoning result and the label result of the saved input data are subjected to fusion reasoning based on the PCR5 combination rule in the DSmT framework, the fusion reasoning result is obtained and saved, and the corresponding relationship with the grouped 10 training input data is saved.
[0028] Preferably, the random initial parameters of the deep neural network model stored in the feedback training access module in step S5 are: random initial parameters of 16 different convolution kernels of convolutional neural network convolutional layer 1, random initial parameters of 32 different convolution kernels of convolutional neural network convolutional layer 2, random initial parameters of 16 different convolution kernels of convolutional neural network convolutional layer 3, random initial full connection parameters between each node of convolutional neural network pooling layer 3 and 100 nodes of convolutional neural network hidden layer, random initial full connection parameters between 100 nodes of convolutional neural network hidden layer and 2 nodes of convolutional neural network output layer;
[0029] The specific operation method of the feedback training access module in step S5 is:
[0030] S5a, performing difference calculation on the average value of the fusion reasoning results of the 10 training input data groups saved by the fusion reasoning module and the forward reasoning results of the same data saved by the forward reasoning result saving module, to obtain the difference vector corresponding to the 10 data of each group;
[0031] S5b, input the difference vector corresponding to each group of 10 data into the deep neural network model saved in the feedback training access module, and perform layer-by-layer reverse reasoning of the reverse reasoning algorithm between the layers corresponding to the convolutional neural network, that is, perform layer-by-layer reverse reasoning of the reverse reasoning algorithm between the layers corresponding to the convolutional neural network in the order 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 3, the convolutional neural network pooling layer 3 to the convolutional neural network convolutional layer 3, the convolutional neural network convolutional layer 3 to the convolutional neural network pooling layer 2, the convolutional neural network pooling layer 2 to the convolutional neural network convolutional layer 2, the convolutional neural network convolutional layer 2 to the convolutional neural network pooling layer 1, and the convolutional neural network pooling layer 1 to the convolutional neural network convolutional layer 1, to obtain new deep neural network model parameters, and save and update;
[0032] S5c, randomly select 10 input training data from a group and perform steps S5a to S5b;
[0033] S5d, iterate step S5c up to 100 times to obtain the trained deep neural network model parameters, and send them to the classification model parameter saving module for storage.
[0034] Preferably, the network model parameters of the trained deep neural network model saved by the classification model parameter saving module in step S7 include: trained parameters of 16 different convolution kernels of convolutional neural network convolutional layer 1, trained parameters of 32 different convolution kernels of convolutional neural network convolutional layer 2, trained parameters of 16 different convolution kernels of convolutional neural network convolutional layer 3, trained fully connected parameters between each node of convolutional neural network pooling layer 3 and 100 nodes of convolutional neural network hidden layer, and trained fully connected parameters between 100 nodes of convolutional neural network hidden layer and N nodes representing the number of radar radiation source types in convolutional neural network output layer.
[0035] The present invention also provides a method for identifying a low signal-to-noise ratio radar radiation source signal pulse identification device, comprising the following identification steps:
[0036] C1, substitute the two-dimensional data of the time-frequency data of the radar radiation source signal data into the data block module, and obtain the time-frequency input data of the radar radiation source signal to be identified with the maximum number of rows and columns;
[0037] C2, substitute the time-frequency input data of the radar radiation source signal to be identified with the maximum number of rows and columns into the convolutional neural network convolutional layer 1, convolutional neural network pooling layer 1, convolutional neural network convolutional layer 2, convolutional neural network pooling layer 2, convolutional neural network convolutional layer 3, convolutional neural network pooling layer 3, convolutional neural network hidden layer, and convolutional neural network output layer from left to right, and calculate according to the model parameters saved by the classification model parameter saving module, and calculate the N-dimensional vector result representing the number of radar radiation source types of the convolutional neural network output layer;
[0038] C3, compares the N-dimensional vector results representing the number of radar radiation source types. If the value representing a certain radar radiation source signal type is the maximum value in the vector, the signal is output as a radar radiation source signal of this type.
[0039] Compared with the prior art, the present invention has the following beneficial effects:
[0040] 1. The present invention adopts a fusion reasoning step, which is the first application in the field of radar emitter signal intra-pulse identification. The forward reasoning result and the label result can be treated at the same level, that is, not all results are believed to be true. In fact, under the condition of low signal-to-noise ratio, the classification of radar emitter signals may be wrong, and radar emitter signals of different categories may also be inter-classified due to the influence of noise. Therefore, labels are sometimes biased. By adopting fusion reasoning, it can be considered that the reasoning ability of the trained model also has a certain authority. It can be used as a textbook expert such as an expert and a label to discuss with each other to obtain a fusion result. This is a great improvement over the existing comparison method of neural network models and label results.
[0041] 2. It can effectively train the data set formed by the time-frequency data of the detected low-SNR radar emitter signal in actual situations. By using the time-frequency data of the low-SNR radar emitter signal, a classification model feedback training method that avoids overfitting is adopted during the training process to train a classification model with better classification performance;
[0042] 3. In actual situations, the type of radar radiation source signal with a low signal-to-noise ratio can be accurately classified and identified (the classification and identification referred to here refers to identifying the type of the radar radiation source signal). By using the classification model trained by the present invention, the time-frequency data of the low signal-to-noise ratio radar radiation source signal is used as the input of the classification model and substituted into the trained classification model to obtain an accurate classification and identification result, which can effectively distinguish the type of radar radiation source signal with a low signal-to-noise ratio.
[0043] Other features and advantages of the present disclosure will become apparent from the following detailed description of exemplary embodiments of the present disclosure with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure or related technologies, the drawings required for use in the embodiments or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0045] Figure 1 The overall structure and training flow chart of the radar radiation source signal intra-pulse identification device of the present invention;
[0046] Figure 2 This is a diagram of the data block processing process of the data block module of the radar radiation source signal pulse identification device of the present invention;
[0047] Figure 3 This is a design diagram of a deep neural network module model of a radar radiation source signal intra-pulse identification device of the present invention;
[0048] In the figure: 1. Data segmentation module, 2. Deep neural network module, 3. Forward reasoning result storage module, 4. Fusion reasoning module, 5. Feedback training access module, 6. Classification model parameter storage module. DETAILED DESCRIPTION
[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0050] Unless otherwise specifically stated, the relative arrangement of the parts and steps described in these embodiments, numerical expressions and numerical values do not limit the scope of the present invention. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the accompanying drawings are not drawn according to the actual proportional relationship. The technology, method and equipment known to ordinary technicians in the relevant field may not be discussed in detail, but in appropriate cases, the technology, method and equipment should be regarded as a part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary, rather than as a limitation. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once a certain item is defined in an accompanying drawing, it does not need to be further discussed in subsequent drawings.
[0051] For ease of description, spatially relative terms such as "above", "above", "on the upper surface of", "above", etc. may be used here to describe the spatial positional relationship between a device or feature and other devices or features as shown in the figure. It should be understood that spatially relative terms are intended to include different orientations of the device in use or operation in addition to the orientation described in the figure. For example, if the device in the accompanying drawings is inverted, the device described as "above other devices or structures" or "above other devices or structures" will be positioned as "below other devices or structures" or "below other devices or structures". Thus, the exemplary term "above" can include both "above" and "below". The device can also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatially relative descriptions used here are interpreted accordingly.
[0052] See also Figure 1-3The present invention provides a technical solution: a low signal-to-noise ratio radar radiation source signal pulse identification device, according to Figure 1 The setting method shown connects each module, that is, the data segmentation module 1, the deep neural network module 2, the forward reasoning result storage module 3, the fusion reasoning module 4, and the feedback training access module 5 are set in sequence; and the feedback training access module 5 is then reversely connected back to the deep neural network module 2; and the deep neural network module 2 is also separately provided with a classification model parameter storage module 6 at the back end.
[0053] Wherein, in certain embodiments:
[0054] The data block module 1 is a data input and data sorting device based on USB or other data transmission interface, CPU and hard disk, that is, it can read different types of data from an external storage device through USB or other data transmission interface, find and import the data into the internal storage hard disk of the entire training device, and sort the data through the instructions of the CPU while importing it into the internal storage hard disk, calculate the longest length of the uneven data, and then fill it according to the longest number of rows and columns, and then put it into the internal storage hard disk for storage;
[0055] The deep neural network module 2 is a data processing device based on GPU, memory and hard disk. It reads the stored data into the memory through the hard disk, and then sends the data to the GPU through the memory to perform forward reasoning on the deep neural network model to obtain the forward reasoning result, and then sends the reasoning result to the memory for storage;
[0056] The forward reasoning result saving module 3 is equivalent to a storage device based on memory and hard disk, that is, the forward reasoning result calculated by the deep neural network module 2 in the memory is exported from the memory to the hard disk for storage;
[0057] The fusion reasoning module 4 is based on a hard disk, a memory and a CPU. The forward reasoning result, the label result and the fusion reasoning instruction based on the PCR5 combination rule in the DSmT framework are read through the hard disk and read into the memory. The CPU performs the fusion reasoning instruction based on the PCR5 combination rule in the DSmT framework on the forward reasoning result and the label result in the memory to obtain the fusion reasoning result, and then outputs the fusion reasoning result in the memory to the hard disk for storage.
[0058] The feedback training access module 5 is a data processing device based on a GPU, a memory and a hard disk. The stored multi-dimensional data is read into the memory through the hard disk, and then the data is sent to the GPU through the memory to perform reverse reasoning on the multi-channel deep neural network model to obtain the reverse reasoning result, and then the reasoning result is sent to the memory for storage;
[0059] The classification model parameter storage module 6 is equivalent to a storage device based on memory and hard disk, that is, the reverse reasoning result calculated by the feedback training access module 5 in the memory is exported from the memory to the hard disk for storage.
[0060] Before using the low signal-to-noise ratio radar radiation source signal intra-pulse identification device of the present invention to identify the radar radiation source signal, the device needs to be trained first, that is, a classification model training is performed on the radar radiation source signal when the signal-to-noise ratio of the radar radiation source signal data actually obtained is low, so that a radar radiation source signal classification model with high training accuracy and not easy to enter an overfitting state can be obtained, and then the classification model is used to identify the radar radiation source signal using the identification method of the present invention.
[0061] The specific training method of the classification model is to sequentially execute step-by-step training steps according to the module design of the signal recognition device of the present invention. First, data is input. The input data of the radar radiation source signal classification model training method is the time-frequency data of the radar radiation source signal under the condition of low signal-to-noise ratio radar radiation source signal.
[0062] The specific training method includes the following training steps:
[0063] S1, inputting the time-frequency data of the radar radiation source signal with labels into the data segmentation module 1 as input data to obtain segmented training data;
[0064] In order to facilitate the training of the classification model training method, the input data with labels is normalized and divided into blocks. For the specific method of the data block module 1 in the classification model training method, see Figure 2 , which includes the following steps:
[0065] S1a, calculate the two-dimensional data size of the time-frequency input data of radar radiation source signals of various models. Since the input data are all discretized two-dimensional data, calculate the number of rows and columns of various two-dimensional data, and save the maximum number of rows and columns of various types of input data;
[0066] S1b, each type of input data is padded according to the maximum number of rows and columns stored, that is, if a certain type of input data is less than the number of rows and columns of the longest stored two-dimensional data, the number of rows and columns that differ is set to 0 to padded in front of the rows and columns of the input data, and then Gaussian white noise is added to the part where it is taken as 0. The mean of the Gaussian white noise is 0, and the standard deviation is the average power of the signal. Through the above steps, different types of radar radiation source signal time-frequency input data with equal number of rows and columns can be obtained.
[0067] S2, inputting the divided training data into the deep neural network module 2, and obtaining the forward reasoning result through the forward reasoning operation of the network structure and network parameters in the deep neural network module 2;
[0068] See also Figure 3 ,Taking comprehensive consideration of the depth of the ,dark network model into consideration, the specific model of the ,deep neural network module 2 is designed as follows: ,corresponding to the data input of the data ,segmentation module 1, the input data is connected to three layers of convolutional ,pooling layer, one layer of convolutional neural network hidden layer, and one layer of convolutional ,neural network output layer.
[0069] S3, saving the forward reasoning operation results of each type of input data in batches into the forward reasoning result saving module 3, and taking the average value of each batch of results in each dimension for saving;
[0070] The forward reasoning result saving module 3 completes the forward reasoning result saving of the deep neural network module 2. The forward reasoning result saving process is as follows:
[0071] S3a, the forward reasoning result saving module 3 saves the label result of each training input data;
[0072] S3b, randomly extract 10 data from the training input data with the same label result, substitute each training input data into the deep neural network module 2, perform forward reasoning calculation, and take the average value of the N-dimensional vector forward reasoning results of the 10 training input data in each dimension, where N represents the number of radar radiation source types;
[0073] S3c, saves the average value as the forward inference result of 10 randomly selected data, and saves the corresponding relationship with the 10 training input data.
[0074] S4, inputting the saved average values of the forward reasoning results of each category into the fusion reasoning module 4 one by one, performing fusion reasoning on the average values of the forward reasoning results and the label results of the input data, obtaining the fusion reasoning results and saving them;
[0075] The specific operation method of the fusion reasoning module 4 is:
[0076] S4a, inputting the average value of the forward reasoning results of 10 randomly selected training input data with the same label results saved by the forward reasoning result saving module 3 into the fusion reasoning module 4;
[0077] S4b, the average value of the forward reasoning result and the label result of the saved input data are subjected to fusion reasoning based on the PCR5 combination rule in the DSmT framework, the fusion reasoning result is obtained and saved, and the corresponding relationship with the grouped 10 training input data is saved.
[0078] S5, input the fusion reasoning results one by one into the feedback training access module 5, make a difference between the fusion reasoning results and the average value of the forward reasoning results, obtain the difference vectors one by one, substitute the difference vectors into the deep neural network model stored in the feedback training access module 5, perform layer-by-layer reverse reasoning of the reverse reasoning algorithm between the layers corresponding to the convolutional neural network, and obtain new network model parameters of the deep neural network model;
[0079] See also Figure 3 In step S5, the random initial parameters of the deep neural network model stored in the feedback training access module 5 are: random initial parameters of 16 different convolution kernels of the convolutional neural network convolutional layer 1, random initial parameters of 32 different convolution kernels of the convolutional neural network convolutional layer 2, random initial parameters of 16 different convolution kernels of the convolutional neural network convolutional layer 3, random initial full connection parameters between each node of the convolutional neural network pooling layer 3 and 100 nodes of the convolutional neural network hidden layer, and random initial full connection parameters between 100 nodes of the convolutional neural network hidden layer and 2 nodes of the convolutional neural network output layer;
[0080] The specific operation method of the feedback training access module 5 in step S5 is:
[0081] S5a, performing difference calculation on the average value of the fusion reasoning results of the 10 training input data groups saved by the fusion reasoning module 4 and the forward reasoning results of the same data saved by the forward reasoning result saving module 3, and obtaining the difference vector corresponding to the 10 data of each group;
[0082] S5b, input the difference vector corresponding to each group of 10 data into the deep neural network model saved in the feedback training access module 5, and perform layer-by-layer reverse reasoning of the reverse reasoning algorithm between the layers corresponding to the convolutional neural network, that is, perform layer-by-layer reverse reasoning of the reverse reasoning algorithm between the layers corresponding to the convolutional neural network in the order 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 3, the convolutional neural network pooling layer 3 to the convolutional neural network convolutional layer 3, the convolutional neural network convolutional layer 3 to the convolutional neural network pooling layer 2, the convolutional neural network pooling layer 2 to the convolutional neural network convolutional layer 2, the convolutional neural network convolutional layer 2 to the convolutional neural network pooling layer 1, and the convolutional neural network pooling layer 1 to the convolutional neural network convolutional layer 1, to obtain new deep neural network model parameters, and save and update;
[0083] S5c, randomly select 10 input training data from a group and perform steps S5a to S5b;
[0084] S5d, iterate step S5c up to 100 times to obtain the trained deep neural network model parameters, and send them to the classification model parameter saving module 6 for storage.
[0085] S6, inputting the new network model parameters of the deep neural network model obtained by reverse reasoning of the feedback training access module 5 into the deep neural network module 2 for storage and update;
[0086] S7, iterate steps S2 to S6 up to 100 times, and send the network model parameters of the finally trained deep neural network model to the classification model parameter saving module 6 for storage.
[0087] See also Figure 3 The network model parameters of the trained deep neural network model saved by the classification model parameter saving module 6 include: the trained parameters of 16 different convolution kernels of the convolutional neural network convolutional layer 1, the trained parameters of 32 different convolution kernels of the convolutional neural network convolutional layer 2, the trained parameters of 16 different convolution kernels of the convolutional neural network convolutional layer 3, the trained full connection parameters between each node of the convolutional neural network pooling layer 3 and the 100 nodes of the convolutional neural network hidden layer, and the trained full connection parameters between the 100 nodes of the convolutional neural network hidden layer and the N nodes representing the number of radar radiation source types in the convolutional neural network output layer.
[0088] After the training method of the aforementioned steps S1-S7, the present invention obtains a low signal-to-noise ratio radar radiation source signal classification model. Figure 3 By using this classification model to classify and identify radar emitter signals with low signal-to-noise ratio, a recognition result with higher classification and recognition accuracy can be obtained.
[0089] Specifically, the classification model obtained after training is used to identify the radar radiation source signal in the pulse under low signal-to-noise ratio conditions, including the following identification steps:
[0090] C1, substitute the two-dimensional data of the time-frequency data of the radar radiation source signal data into the data block module 1, and obtain the time-frequency input data of the radar radiation source signal to be identified with the maximum number of rows and columns;
[0091] C2, substitute the time-frequency input data of the radar radiation source signal to be identified with the maximum number of rows and columns into the convolutional neural network convolutional layer 1, convolutional neural network pooling layer 1, convolutional neural network convolutional layer 2, convolutional neural network pooling layer 2, convolutional neural network convolutional layer 3, convolutional neural network pooling layer 3, convolutional neural network hidden layer, and convolutional neural network output layer from left to right, and calculate according to the model parameters saved by the classification model parameter saving module 6, and calculate and obtain the N-dimensional vector result representing the number of radar radiation source types of the convolutional neural network output layer;
[0092] C3, compares the N-dimensional vector results representing the number of radar radiation source types. If the value representing a certain radar radiation source signal type is the maximum value in the vector, the signal is output as a radar radiation source signal of this type.
[0093] The working principle of the training method of the present invention:
[0094] 1. The fusion reasoning module 4 of the present invention takes into account the special situation of the background with low signal-to-noise ratio, that is, the label result of the radar radiation source signal may be wrong, or it may be overfitted data, that is, due to the influence of noise, multiple types of data are individually intertwined. In this way, forward reasoning is performed on randomly selected grouped data, and then the average value of the forward reasoning results of the grouped data is taken as the forward reasoning result, which largely avoids the influence of individual erroneous data on the result. Then, the forward reasoning result is regarded as a result with a certain degree of credibility with the label result. The two are fused based on the PCR5 rule under the DSmT framework. The influence of erroneous data and overfitted data can be avoided, and the contradiction between the two can also be avoided, so as to obtain a more scientific fusion reasoning result;
[0095] 2. The feedback training access module 5 of the present invention considers subtracting the fused reasoning result from the forward reasoning result. Here, the label is first fused with the forward reasoning result, taking into account both the label result and the reasoning result of the trained model, thereby avoiding label result errors. Subtracting the forward reasoning result from the label result will lead to overfitting, resulting in overfitting of the trained classification model, making it difficult for the classification model to achieve a high accuracy. The present invention avoids this problem and can obtain a high-precision classification model.
[0096] The working principle of the identification method of the present invention is as follows:
[0097] Using the neural network model parameters obtained by the training method of the present invention to identify radar emitter signals can avoid the problem of label result data errors or data overfitting in the case of low signal-to-noise ratio radar emitter signals, and obtain accurate classification results of radar emitter signal types.
[0098] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A low signal-to-noise ratio radar radiation source signal pulse identification device, characterized in that: The radar radiation source signal intra-pulse identification device comprises a data segmentation module (1); a deep neural network module (2) is arranged at the back end of the data segmentation module (1); a forward reasoning result storage module (3) is arranged at the back end of the deep neural network module (2); a fusion reasoning module (4) is arranged at the back end of the forward reasoning result storage module (3); a feedback training access module (5) is arranged at the back end of the fusion reasoning module (4); the feedback training access module (5) is connected to the deep neural network module (2) at the back end; and a classification model parameter storage module (6) is also arranged at the back end of the deep neural network module (2); The training method of the low signal-to-noise ratio radar radiation source signal intra-pulse identification device comprises the following training steps: S1, inputting the time-frequency data of the radar radiation source signal with labels into the data segmentation module (1) as input data to obtain segmented training data; S2, inputting the divided training data into the deep neural network module (2), and obtaining the forward reasoning result through the forward reasoning operation of the network structure and network parameters in the deep neural network module (2); S3, saving the forward reasoning operation results of each type of input data in batches into the forward reasoning result saving module (3), and taking the average value of each batch of results in each dimension for saving; S4, inputting the saved average values of the forward reasoning results of each category into the fusion reasoning module (4) one by one, performing fusion reasoning on the average values of the forward reasoning results and the label results of the input data, obtaining the fusion reasoning results and saving them; S5, input the fusion reasoning results one by one into the feedback training access module (5), subtract the fusion reasoning results from the average value of the forward reasoning results to obtain difference vectors one by one, substitute the difference vectors into the deep neural network model stored in the feedback training access module (5), perform layer-by-layer reverse reasoning of the reverse reasoning algorithm between the corresponding layers of the convolutional neural network, and obtain new network model parameters of the deep neural network model; S6, inputting the new network model parameters of the deep neural network model obtained by reverse reasoning of the feedback training access module (5) into the deep neural network module (2) for storage and update; S7, iterate steps S2 to S6 up to 100 times, and send the network model parameters of the finally trained deep neural network model to the classification model parameter saving module (6) for storage; The specific operation method of the fusion reasoning module (4) in step S4 is: S4a, inputting the average value of the forward reasoning results of 10 randomly selected training input data with the same label results saved by the forward reasoning result saving module (3) into the fusion reasoning module (4); S4b, perform fusion reasoning based on the PCR5 combination rule in the DSmT framework on the average value of the forward reasoning result and the label result of the saved input data, save the obtained fusion reasoning result, and save the corresponding relationship with the grouped 10 training input data; The random initial parameters of the deep neural network model stored in the feedback training access module (5) in step S5 are: random initial parameters of 16 different convolution kernels of convolutional neural network convolutional layer 1, random initial parameters of 32 different convolution kernels of convolutional neural network convolutional layer 2, random initial parameters of 16 different convolution kernels of convolutional neural network convolutional layer 3, random initial full connection parameters between each node of convolutional neural network pooling layer 3 and 100 nodes of convolutional neural network hidden layer, and random initial full connection parameters between 100 nodes of convolutional neural network hidden layer and 2 nodes of convolutional neural network output layer; The specific operation method of the feedback training access module (5) in step S5 is: S5a, performing difference calculation on the fusion reasoning results of the 10 training input data of each group saved by the fusion reasoning module (4) and the average value of the forward reasoning results of the same data saved by the forward reasoning result saving module (3), and obtaining a difference vector corresponding to the 10 data of each group; S5b, input the difference vector corresponding to each group of 10 data into the deep neural network model saved by the feedback training access module (5), and perform reverse reasoning of the reverse reasoning algorithm between the layers corresponding to the convolutional neural network layer by layer, that is, perform reverse reasoning of the reverse reasoning algorithm between the layers corresponding to the convolutional neural network layer by layer in the order 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 3, the convolutional neural network pooling layer 3 to the convolutional neural network convolutional layer 3, the convolutional neural network convolutional layer 3 to the convolutional neural network pooling layer 2, the convolutional neural network pooling layer 2 to the convolutional neural network convolutional layer 2, the convolutional neural network convolutional layer 2 to the convolutional neural network pooling layer 1, and the convolutional neural network pooling layer 1 to the convolutional neural network convolutional layer 1, and obtain new deep neural network model parameters, and save and update them; S5c, randomly select 10 input training data from a group and perform steps S5a to S5b; S5d, iterate step S5c up to 100 times to obtain the trained deep neural network model parameters, which are sent to the classification model parameter saving module (6) for storage.
2. The low signal-to-noise ratio radar radiation source signal intra-pulse identification device according to claim 1, characterized in that: In step S1, the data segmentation module (1) performs normalized data segmentation processing on the input data with labels, and the specific processing method includes: S1a, calculate the two-dimensional data size of the time-frequency input data of radar radiation source signals of various models. Since the input data are all discretized two-dimensional data, calculate the number of rows and columns of various two-dimensional data, and save the maximum number of rows and columns of various types of input data; S1b, each type of input data is padded according to the maximum number of rows and columns stored, that is, if a certain type of input data is less than the number of rows and columns of the longest stored two-dimensional data, the number of rows and columns that differ is set to 0 to padded in front of the rows and columns of the input data, and then Gaussian white noise is added to the part where it is taken as 0. The mean of the Gaussian white noise is 0, and the standard deviation is the average power of the signal. Through the above steps, different types of radar radiation source signal time-frequency input data with equal number of rows and columns can be obtained.
3. The low signal-to-noise ratio radar radiation source signal intra-pulse identification device according to claim 1, characterized in that: The specific model design of the forward reasoning operation of the deep neural network module (2) in step S2 is as follows: corresponding to the data input of the data segmentation module (1), the input data is connected to three convolutional pooling layers of the convolutional neural network, one hidden layer of the convolutional neural network, and one output layer of the convolutional neural network.
4. The low signal-to-noise ratio radar radiation source signal intra-pulse identification device according to claim 1, characterized in that: The specific steps of the forward reasoning result saving module (3) in step S3 for saving the forward reasoning operation results in batches are: S3a, the forward reasoning result saving module (3) saves the label result of each training input data; S3b, randomly select 10 data from the training input data with the same label result, substitute each training input data into the deep neural network module (2), perform forward reasoning calculation, and take the average value of the N-dimensional vector forward reasoning results of the 10 training input data in each dimension, where N represents the number of radar radiation source types; S3c, saves the average value as the forward inference result of 10 randomly selected data, and saves the corresponding relationship with the 10 training input data.
5. The low signal-to-noise ratio radar radiation source signal intra-pulse identification device according to claim 1, characterized in that: The network model parameters of the trained deep neural network model saved by the classification model parameter saving module (6) in step S7 include: the trained parameters of 16 different convolution kernels of the convolution neural network convolution layer 1, the trained parameters of 32 different convolution kernels of the convolution neural network convolution layer 2, the trained parameters of 16 different convolution kernels of the convolution neural network convolution layer 3, the trained full connection parameters between each node of the convolution neural network pooling layer 3 and 100 nodes of the convolution neural network hidden layer, and the trained full connection parameters between the 100 nodes of the convolution neural network hidden layer and N nodes representing the number of radar radiation source types in the convolution neural network output layer.
6. The identification method of the low signal-to-noise ratio radar radiation source signal pulse identification device according to claim 1, comprising the following identification steps: C1, substituting the two-dimensional data of the time-frequency data of the radar radiation source signal data into the data block module (1), and obtaining the time-frequency input data of the radar radiation source signal to be identified, the number of rows and the number of columns of which are both the maximum number of rows and the maximum number of columns; C2, substitute the time-frequency input data of the radar radiation source signal to be identified with the maximum number of rows and columns into the convolutional neural network convolutional layer 1, convolutional neural network pooling layer 1, convolutional neural network convolutional layer 2, convolutional neural network pooling layer 2, convolutional neural network convolutional layer 3, convolutional neural network pooling layer 3, convolutional neural network hidden layer, and convolutional neural network output layer from left to right, and calculate according to the model parameters saved by the classification model parameter saving module (6), and calculate and obtain the N-dimensional vector result representing the number of radar radiation source types of the convolutional neural network output layer; C3, compares the N-dimensional vector results representing the number of radar radiation source types. If the value representing a certain radar radiation source signal type is the maximum value in the vector, the signal is output as a radar radiation source signal of this type.