Deep learning signal open set identification method based on dynamic matching comparison

Abstract

The invention relates to a deep learning signal open-set identification method based on dynamic matching comparison, and the method comprises the steps: processing all signals to obtain a training sample, constructing a data set, training batch data, and transmitting the data to a deep learning network for forward propagation, thereby obtaining 2k high-order semantic features; euclidean distance calculation and activation function nonlinear processing are carried out on the first k high-order semantic features and the last k high-order semantic features, and signals corresponding to the minimum distance are found and sorted; back propagation is carried out on the deep learning network, all weight parameters are updated, the data set training batch number is constructed again, and after training of the deep learning model is stopped, features are extracted to construct a feature matching sample library; and performing feature extraction after processing the to-be-matched signal data, performing Euclidean distance calculation on the extracted features and all features in the feature matching sample library, and performing identification and judgment according to a calculation result. According to the invention, a known signal target can be identified, and an unknown signal target can also be identified.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and in particular to a deep learning signal open set recognition method based on dynamic matching comparison. Background Art

[0002] With the rapid development of wireless communication technology and intelligent equipment, signal open set identification has shown its indispensable importance in both civilian and military fields. In the civilian field, signal open set identification is reflected in the optimization and management of complex communication environments. In modern life, wireless communication devices such as smart phones, wireless routers, and Bluetooth devices are ubiquitous, resulting in increasingly crowded civilian frequency bands. Signal open set identification technology can accurately distinguish signals emitted by different devices, effectively identify and manage interference signals, thereby ensuring the stability and reliability of communication networks. In addition, this technology can also assist spectrum management agencies in the reasonable allocation and utilization of limited spectrum resources, improve spectrum utilization efficiency, and provide strong support for the sustainable development of wireless communications. In the military field, the importance of signal open set identification is even more prominent. In a complex electromagnetic environment, military communication systems need to quickly and accurately identify signals from both sides of the enemy to ensure the confidentiality, integrity, and availability of communications. Signal open set identification technology can achieve real-time monitoring and identification of unknown or potential threat signals, providing key information support for military decision-making. At the same time, this technology can also improve the anti-interference and survivability of military communication systems, ensuring that effective communication can be maintained in harsh environments.

[0003] In recent years, the rapid development of deep learning technology has brought revolutionary breakthroughs in the field of signal recognition. Through powerful feature extraction and pattern recognition capabilities, it has significantly improved the accuracy and efficiency of signal recognition, and provided strong support for technological progress and application innovation in multiple fields such as wireless communications, the Internet of Things, and intelligent monitoring. However, many signal recognition applications are used in non-cooperative scenarios, and current signal recognition technology cannot meet practical scenarios when the number of signal targets is unknown. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a deep learning signal open set recognition method based on dynamic matching comparison to solve the shortcomings of the prior art.

[0005] The object of the present invention is achieved by the following technical solution: a deep learning signal open set recognition method based on dynamic matching comparison, the recognition method comprising:

[0006] S1. Process all signals to obtain training samples, construct data set training batch data and send them into the deep learning network for forward propagation to obtain 2k high-order semantic features;

[0007] S2, perform Euclidean distance calculation and activation function nonlinear processing on the first k high-order semantic features and the last k high-order semantic features, find the corresponding signal with the minimum distance and sort them;

[0008] S3, back-propagate the deep learning network, update all weight parameters, re-construct the number of training batches of the data set, and extract features to build a feature matching sample library after the deep learning model stops training;

[0009] S4. After processing the signal data to be matched, feature extraction is performed, and the Euclidean distance between the extracted features and all features in the feature matching sample library is calculated, and recognition judgment is performed based on the calculation results.

[0010] The S1 specifically includes the following contents:

[0011] S101, performing 0-mean normalization processing on all signal data to obtain sample data that conforms to a standard normal distribution;

[0012] S102, performing fixed-scale W segmentation on the normalized sample data to obtain training samples, where the length of each sample is W;

[0013] S103, constructing a data set training batch data, the data set training batch data size is 2k, wherein the first k data samples are different signal data randomly selected from the normalized sample data, and the last k data samples are randomly selected data;

[0014] S104: Send it into the deep learning network for forward propagation to obtain 2k high-order semantic features.

[0015] The random selection method includes: matching the k+t-th data with the t-th data, setting the label to 0 if they are the same signal, and setting the label to 1 if they are different signals, to ensure that the ratio of 0 to 1 labels is 1.

[0016] The S2 specifically includes the following contents:

[0017] S201, performing Euclidean distance calculation on the first k high-order semantic features and the last k high-order semantic features through an L2 regularized network, and performing nonlinear processing using a tanh activation function, and performing mean square error loss calculation on the calculation result and the label;

[0018] S202, calculate the Euclidean distance between the first k high-order semantic features, and find the corresponding signal with the smallest distance, sort the k groups of minimum distance results from small to large, find the first k / 2 signals and the corresponding signals, and record them.

[0019] The S3 specifically includes the following contents:

[0020] S301, perform back propagation on the deep learning network and update all weight parameters;

[0021] S302, constructing a training batch of data sets, the first k data samples are different information data randomly selected from the normalized sample data, if the tth signal is recorded, then the k+tth random sample of the corresponding signal in the record is randomly selected, and the label is set to 1, and other signals are selected corresponding to the same random signal, the label is set to 0, and return to step S104;

[0022] S303. When the deep learning model converges, stop training, select a certain number of sample labels for feature extraction, and build a feature matching sample library.

[0023] The S4 specifically includes the following contents:

[0024] S401, after the signal data to be matched is normalized and sample clipped for preprocessing, feature extraction is performed using a feature extraction module in a deep learning model;

[0025] S402, calculating the Euclidean distance between the features of the signal data to be matched and all the features in the feature sample library. If the minimum distance value in the sample matching is greater than a threshold, it is an unknown signal. If it is less than or equal to the threshold, it is a signal corresponding to the minimum feature matching value.

[0026] The deep learning model uses convolution, standardization and pooling as feature extraction modules. The feature extraction module first uses a convolution kernel to extract features from an input feature matrix, performs matrix superposition of the extraction result and input data, uses standardization processing, and finally performs pooling processing.

[0027] The present invention has the following advantages: a deep learning signal open set recognition method based on dynamic matching comparison, which can not only identify known signal targets, but also identify unknown signal targets; it still has good effects when the sample is small or the sample is unbalanced; it uses the model to perform efficient feature extraction and uses the distance of the feature to confirm the category of the signal. This method avoids the excessive reliance of traditional classification models on category labels, thereby improving the flexibility and accuracy of classification, having a higher classification accuracy, and still performing well when processing complex signals and noise environments, and having strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the process of the present invention;

[0029] Figure 2 This is a schematic diagram of the deep learning convolutional neural network design of the present invention;

[0030] Figure 3 A schematic diagram of a dynamic composition method of a training batch data set of the present invention;

[0031] Figure 4 This is the feature sample library construction method of the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below in conjunction with the drawings is not intended to limit the scope of protection of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.

[0033] The present invention specifically relates to a deep learning signal open set recognition method based on dynamic matching and comparison. The method extracts high-order semantic features of signals through contrastive learning and dynamic data sets, establishes a signal feature library based on the data involved in training, and uses the features of the signals to perform Euclidean distance calculation to complete signal matching and recognition.

[0034] like Figure 1 As shown, specifically including the following:

[0035] Step 1: Perform 0-mean normalization on all signal data to obtain sample data that conforms to the standard normal distribution (mean is zero and variance is one). The normalization method is as follows;

[0036]

[0037] Where X is the original data, X * is the normalized data, μ is the mean of the original data, δ is the standard deviation of the original data, and the method for calculating μ and δ is:

[0038]

[0039]

[0040] Among them, all signals include MS110A, STANAG 4285, 2G-ALE, 3G-ALE, Link11A, KG84, ANDVT, and MS110A-39T, a total of 8 signals.

[0041] Step 2: Perform fixed-scale 16384 segmentation on the normalized sample data to obtain training samples, and the length of each sample is 16384;

[0042] Step 3: Construct training batch data of the dataset. The batch data size of the dataset is 16. The first 8 data samples are different signal data randomly selected from the normalized sample data, and the last 8 data samples are randomly selected data. The selection method is to correspond the 8+tth data with the tth data. If they are the same signal, the label is set to 0, and if they are different signals, the label is set to 1, ensuring that the ratio of 0 to 1 labels is about 1;

[0043] Step 4: Send it to the deep learning network for forward propagation to obtain 16 high-order semantic features;

[0044] Step 5: The first 8 high-order semantic features and the last 8 high-order semantic feature vectors are calculated through the L2 regularized network and nonlinearly processed using the tanh activation function. The calculation results and the labels are solved for the mean square error loss. The loss solution method is as follows, where y t is the true label, y p is the predicted label;

[0045]

[0046] Step 6: Calculate the Euclidean distance between the first 8 high-order semantic feature vectors, and find the corresponding signal with the smallest distance. After sorting the 8 groups of minimum distance results from small to large, find the first 4 signals and the corresponding signals, and record them;

[0047] Step 7: Back propagate the deep learning network and update all weight parameters.

[0048] Step 8: Construct training batch data of the data set. The first 8 data samples are different signal data randomly selected from the normalized sample data. If the tth signal is in the record of step 6, then the 8+tth signal is randomly selected from the corresponding signal random sample in the record of step 6, and the label is set to 1. For other signals, the corresponding same random signal is selected, and the label is set to 0, and then go to step 4;

[0049] Step 9: When the deep learning model converges, stop training and only use the model feature extraction part. Select a certain number of sample labels for feature extraction and build a feature matching sample library.

[0050] Step 10: After the signal data to be matched is normalized and sample clipped for preprocessing, feature extraction is performed using a feature extraction module;

[0051] Step 11: Calculate the Euclidean distance between the features of the signal data to be matched and all the features in the feature sample library. If the minimum distance value in the sample matching is greater than the threshold, it is an unknown signal. If it is less than or equal to the threshold, it is the signal corresponding to the minimum feature matching value.

[0052] like Figure 2 As shown in the figure, the input of the deep learning network is an input size of 16*1*16384*1. The network uses convolution, standardization and pooling as feature extraction modules. The module first uses the convolution kernel to extract features from the input feature matrix, superimposes the extraction result and the input data into a matrix and uses standardization, and finally performs pooling. The network uses 64, 64, 128, 128, 265, and 256 convolution kernel modules to process the data in turn, and performs 1*2 pooling to obtain a feature matrix of size 16*256*512. The global average pooling operation (GAP) is used to obtain a 16*512 feature matrix, and the matrix is ​​split according to the first dimension to obtain two 8*512 feature matrices. The two matrices are solved using L2 regularization to obtain 8 output results, and finally tanh is used for nonlinear operations.

[0053] like Figure 3 As shown, the first 8 training samples are feature extracted through the deep learning feature extraction model to obtain 8 512-dimensional feature vectors, the Euclidean distance between the 8 feature vectors and other feature vectors is calculated respectively, and the corresponding signal with the smallest distance is found. After sorting the 8 groups of minimum distance results from small to large, the first 4 signals and the corresponding signals are found and recorded; 8 different signal data are randomly selected from the normalized sample data as the first 8 items of the training batch data set. If the tth signal is in the comparison record, the 8+tth signal random sample corresponding to the comparison record is randomly selected, and the label is set to 1. Other signals select the same random signal, and the label is set to 0. The training batch data set can be completed.

[0054] like Figure 4 As shown, each of the eight different signals has twenty feature vectors. The selection method is to randomly select ten sample data for each signal, and then compare the features of these data samples with the features of the remaining identical signal samples, find the feature vector farthest from the feature, and add it to the sample library.

[0055] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used for various other combinations, modifications and improvements, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not deviate from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A deep learning signal open set recognition method based on dynamic matching comparison, characterized by: The identification method comprises: S1. Process all signals to obtain training samples, construct data set training batch data and send them into the deep learning network for forward propagation to obtain 2k high-order semantic features; S2, perform Euclidean distance calculation and activation function nonlinear processing on the first k high-order semantic features and the last k high-order semantic features, find the corresponding signal with the minimum distance and sort them; S3, back-propagate the deep learning network, update all weight parameters, re-construct the number of training batches of the data set, and extract features to build a feature matching sample library after the deep learning model stops training; S4. After processing the signal data to be matched, feature extraction is performed, and the Euclidean distance between the extracted features and all features in the feature matching sample library is calculated, and recognition judgment is performed based on the calculation results.

2. The method for deep learning signal open set recognition based on dynamic matching comparison according to claim 1, characterized in that: The S1 specifically includes the following contents: S101, performing 0-mean normalization processing on all signal data to obtain sample data that conforms to a standard normal distribution; S102, performing fixed-scale W segmentation on the normalized sample data to obtain training samples, where the length of each sample is W; S103, constructing a data set training batch data, the data set training batch data size is 2k, wherein the first k data samples are different signal data randomly selected from the normalized sample data, and the last k data samples are randomly selected data; S104: Send it into the deep learning network for forward propagation to obtain 2k high-order semantic features.

3. The method for deep learning signal open set recognition based on dynamic matching comparison according to claim 2, characterized in that: The random selection method includes: matching the k+t-th data with the t-th data, setting the label to 0 if they are the same signal, and setting the label to 1 if they are different signals, to ensure that the ratio of 0 to 1 labels is 1.

4. The method for deep learning signal open set recognition based on dynamic matching comparison according to claim 1, characterized in that: The S2 specifically includes the following contents: S201, performing Euclidean distance calculation on the first k high-order semantic features and the last k high-order semantic features through an L2 regularized network, and performing nonlinear processing using a tanh activation function, and performing mean square error loss calculation on the calculation result and the label; S202, calculate the Euclidean distance between the first k high-order semantic features, and find the corresponding signal with the smallest distance, sort the k groups of minimum distance results from small to large, find the first k / 2 signals and the corresponding signals, and record them.

5. The method for deep learning signal open set recognition based on dynamic matching comparison according to claim 3, characterized in that: The S3 specifically includes the following contents: S301, perform back propagation on the deep learning network and update all weight parameters; S302, constructing a training batch of data sets, the first k data samples are different information data randomly selected from the normalized sample data, if the tth signal is recorded, then the k+tth random sample of the corresponding signal in the record is randomly selected, and the label is set to 1, and other signals are selected corresponding to the same random signal, the label is set to 0, and return to step S104; S303. When the deep learning model converges, stop training, select a certain number of sample labels for feature extraction, and build a feature matching sample library.

6. The method for deep learning signal open set recognition based on dynamic matching comparison according to claim 3, characterized in that: The S4 specifically includes the following contents: S401, after the signal data to be matched is normalized and sample clipped for preprocessing, feature extraction is performed using a feature extraction module in a deep learning model; S402, calculating the Euclidean distance between the features of the signal data to be matched and all the features in the feature sample library. If the minimum distance value in the sample matching is greater than a threshold, it is an unknown signal. If it is less than or equal to the threshold, it is a signal corresponding to the minimum feature matching value.

7. A deep learning signal open set recognition method based on dynamic matching comparison according to any one of claims 1 to 6, characterized in that: The deep learning model uses convolution, standardization and pooling as feature extraction modules. The feature extraction module first uses a convolution kernel to extract features from an input feature matrix, performs matrix superposition of the extraction result and input data, uses standardization processing, and finally performs pooling processing.

Images