Multi-level feature distribution learning method and device for automatic modulation open set recognition

Abstract

The invention provides a multilevel feature distribution learning method and device for automatic modulation open set recognition, and relates to the technical field of wireless communication. Comprising the following steps: processing known modulation type signals of different labels by using a manifold hybrid algorithm to generate a pseudo new modulation placeholder; multiple known modulation type signals and unknown modulation type signals are input into a pre-training modulation signal classification model, training is carried out by using a target loss function to obtain a trained modulation signal classification model, and the target loss function comprises modulation classification loss, sample centroid comparison loss and sample maximum and minimum feature constraint loss; inputting a plurality of test modulation signals into the trained modulation signal classification model, and outputting a plurality of target feature embedding; and classifying the plurality of test modulation signals according to the plurality of target feature embedding, the centroid vectors of the plurality of known modulation type signals and the adjustment coefficients. The problem that the empirical risk of a known modulation type and the open space risk of an unknown modulation type are difficult to balance and optimize is solved.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a multi-level feature distribution learning method and device for automatic modulation open set recognition. Background Art

[0002] With the rapid development of wireless communication technology, its application scenarios are becoming increasingly rich, covering multiple fields such as civil and public safety. As a key technology in wireless communication technology, Automatic Modulation Recognition (AMR) plays a vital role in ensuring the normal operation and security of communication systems. The emergence of deep learning technology has brought new development opportunities for AMR. Deep learning technology can automatically learn the characteristics of communication signals without the need for manual complex feature extraction, thereby improving the accuracy of automatic modulation recognition. However, most traditional AMR methods are based on the closed set scenario assumption, that is, it is assumed that all possible modulation types are known in the training stage, and the training data contains samples of these modulation types. This assumption has great limitations in actual complex communication environments. In order to solve the limitations of the closed set scenario assumption method, Automatic Modulation Open Set Recognition (AMOSR) technology came into being. AMOSR technology aims to maintain accurate recognition of known modulation types while effectively identifying and rejecting unknown modulation types. However, AMOSR technology faces many challenges, the most important of which is the difficulty in identifying known and unknown modulation types. Therefore, it is particularly important to identify modulation types.

[0003] Currently, the modulation type is usually identified by optimizing the feature space, using triplet loss to constrain the sample pairwise distance between different signals, and using sample-centroid similarity measures to enhance network training. However, the method of optimizing the feature space does not fully constrain the feature space optimization. The pairwise distance loss cannot constrain all categories, and only relying on the sample-to-centroid similarity loss ignores the distance relationship between local samples, making it difficult to balance the optimization between the empirical risk of known modulation types and the open space risk of unknown modulation types. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a multi-level feature distribution learning method and device for automatic modulation open set identification, so as to solve the problem of difficult balance and optimization between the empirical risk of known modulation types and the open space risk of unknown modulation types.

[0005] To solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0006] A first aspect of the present invention provides a multi-level feature distribution learning method for automatic modulation open set recognition, the method comprising:

[0007] A known modulation type signal with different labels is processed by a manifold mixing algorithm to generate a pseudo new modulation placeholder, and the pseudo new modulation placeholder is used as an unknown modulation type signal, where the pseudo new modulation placeholder is a placeholder corresponding to the unknown signal forged by the manifold mixing algorithm;

[0008] Input a plurality of known modulation type signals and unknown modulation type signals into a pre-trained modulation signal classification model, and train the pre-trained modulation signal classification model using a target loss function to obtain a trained modulation signal classification model, wherein the target loss function includes modulation classification loss, sample centroid comparison loss, and sample maximum and minimum feature constraint loss;

[0009] Input multiple test modulation signals into the trained modulation signal classification model, and output multiple target feature embeddings;

[0010] According to the embedding of multiple target features, the centroid vectors and adjustment coefficients of multiple known modulation type signals, multiple test modulation signals are classified to obtain corresponding classification results.

[0011] A second aspect of the present invention provides a multi-level feature distribution learning device for automatic modulation open set recognition, the device comprising:

[0012] A generation module is used to process known modulation type signals with different labels using a manifold mixing algorithm to generate a pseudo new modulation placeholder, and use the pseudo new modulation placeholder as an unknown modulation type signal, where the pseudo new modulation placeholder is a placeholder corresponding to the unknown signal forged according to the manifold mixing algorithm;

[0013] An input module is used to input multiple test modulation signals into a trained modulation signal classification model, and output multiple target feature embeddings. The trained modulation signal classification model is to input multiple known modulation type signals and unknown modulation type signals into a pre-trained modulation signal classification model, so as to train the pre-trained modulation signal classification model using a target loss function. The obtained modulation signal classification model has a target loss function including modulation classification loss, sample centroid comparison loss, and sample maximum and minimum feature constraint loss.

[0014] The classification module is used to classify multiple test modulation signals according to multiple target feature embeddings, centroid vectors and adjustment coefficients of multiple known modulation type signals to obtain corresponding classification results.

[0015] Compared with the prior art, the multi-level feature distribution learning method and device for automatic modulation open set recognition provided by the present invention process known modulation type signals with different labels using a manifold hybrid algorithm to generate a pseudo-new modulation placeholder, and use the pseudo-new modulation placeholder as an unknown modulation type signal; input a plurality of known modulation type signals and an unknown modulation type signal into a pre-trained modulation signal classification model, and train the pre-trained modulation signal classification model using a target loss function to obtain a trained modulation signal classification model, wherein the target loss function includes a modulation classification loss, a sample centroid comparison loss, and a sample maximum and minimum feature constraint loss; input a plurality of test modulation signals into the trained modulation signal classification model, and output a plurality of target feature embeddings; classify a plurality of test modulation signals according to the plurality of target feature embeddings, the centroid vectors of a plurality of known modulation type signals, and the adjustment coefficients to obtain corresponding classification results. In this way, through the pseudo-new modulation placeholders generated by the manifold mixing algorithm, the pre-trained modulation signal classification model can understand the concept of unknown modulation type signals, enhance the prediction ability, and accurately measure the classification effect of the pre-trained modulation signal classification model on known and unknown modulation type signals; and by proposing a multi-level feature distribution constraint including sample centroid contrast loss and sample-level maximum and minimum feature constraint loss, it helps the pre-trained modulation signal classification model to better distinguish between known and unknown modulation type signals, reserve space for unknown modulation type signals and amplify the difference, thereby optimizing the performance of the pre-trained modulation signal classification model in automatic modulation open set recognition, making it easy to balance and optimize the empirical risk of known modulation type signals and the open space risk of unknown modulation type signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] By reading the detailed description below with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0017] Figure 1 A flowchart of a multi-level feature distribution learning method for automatic modulation open set recognition is schematically shown;

[0018] Figure 2 Schematically showing a schematic diagram of generating a pseudo new modulation placeholder;

[0019] Figure 3 The diagram schematically shows the training process of the pre-trained modulation signal classification model;

[0020] Figure 4 The training diagram of the pre-trained modulation signal classification model is schematically shown;

[0021] Figure 5The structure diagram of a multi-level feature distribution learning device for automatic modulation open set recognition is schematically shown. DETAILED DESCRIPTION

[0022] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0023] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in the present invention should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0024] The method in the embodiment of the present invention is described in detail below.

[0025] Figure 1 The flowchart of the multi-level feature distribution learning method for automatic modulation open set recognition in an embodiment of the present invention is schematically shown. Figure 1 As shown, the method may include:

[0026] S101. Process known modulation type signals with different labels using a manifold mixing algorithm to generate pseudo new modulation placeholders, and use the pseudo new modulation placeholders as unknown modulation type signals.

[0027] The pseudo-new modulation placeholder is a placeholder corresponding to an unknown signal forged according to the manifold mixing algorithm.

[0028] Pseudo-new modulation placeholder is a technique that generates fake unknown signal features through a manifold mixing algorithm, which is used to expand the feature space and improve the unknown category rejection ability of the pre-trained modulation signal classification model in open set modulation recognition.

[0029] Different labels are used to mark the modulation modes of known modulation type signals, and the modulation modes may include 8 digital modulation modes and 3 analog modulation modes. The eight digital modulation modes are: 8Phase ShiftKeying (8PSK), Binary Phase Shift Keying (BPSK), Continuous Phase Frequency Shift Keying (CPFSK), Gaussian Frequency Shift Keying (GFSK), 4-Level Pulse Amplitude Modulation (PAM4), 16-ary Quadrature Amplitude Modulation (16QAM), (64-ary Quadrature Amplitude Modulation, 64QAM) and Quadrature Phase Shift Keying (QPSK). The three analog modulation modes are: Double-Sideband Amplitude Modulation (AM-DSB), Single-Sideband Amplitude Modulation (AM-SSB) and Wide Band Frequency Modulation (WBFM). Modulation, WBFM).

[0030] Since the real unknown modulation type signal cannot be obtained in the subsequent training process, it is necessary to use the manifold mixing algorithm to generate a pseudo new modulation placeholder, and use the pseudo new modulation placeholder as the unknown modulation type signal.

[0031] Figure 2 A schematic diagram of generating a pseudo new modulation placeholder is shown schematically, see Figure 2 Specifically, a known modulation type signal with different labels is processed by using a manifold hybrid algorithm to generate a pseudo new modulation placeholder, and the pseudo new modulation placeholder is used as an unknown modulation type signal, including:

[0032] Step A1: Input a first known modulation type signal and a second known modulation type signal into a feature extractor respectively, so that the feature extractor outputs a corresponding first feature embedding and a second feature embedding.

[0033] The known modulation type signals with different labels are any two of the multiple known modulation type signals. The known modulation type signals with different labels include a first known modulation type signal and a second known modulation type signal. For example, the multiple known modulation type signals are K known modulation type signals, and the first known modulation type signal can be recorded as the known modulation type signal x m , the second known modulation type signal can be recorded as the known modulation type signal x n . Given a modulation type signal x m and a known modulation type signal x n are any two of the K known modulation type signals. m and n belong to K, m is the mth type of K known modulation type signals, and n is the nth type of K known modulation type signals. The first feature embedding is recorded as For a known modulation type signal x m The corresponding feature embedding, the second feature embedding is recorded as For a known modulation type signal x n The corresponding feature embedding.

[0034] The feature extractor is a feature extractor in a pre-trained modulation signal classification model. The feature extractor may use a multi-scale convolutional long short-term deep neural network (MCLDN) as the feature extractor. There may be multiple feature extractors, and the feature extractor is not specifically limited here.

[0035] Step A2: The first feature embedding and the second feature embedding are processed according to the manifold mixing algorithm to generate a pseudo new modulation placeholder, and the pseudo new modulation placeholder is used as an unknown modulation type signal.

[0036] Specifically, the first feature embedding and the second feature embedding are processed according to the manifold mixing algorithm to generate a pseudo new modulation placeholder, including:

[0037] For the first feature and the second feature embedding According to the manifold mixing algorithm, the pseudo new modulation placeholder is generated using the following first formula:

[0038]

[0039] in, is a pseudo-new modulation placeholder, λ is a mixing ratio coefficient obtained by random sampling within a preset range, is the first feature embedding, is the second feature embedding, y mis the modulation type of the first known modulation type signal, y n is the modulation type of the second known modulation type signal.

[0040] The preset range may be in the range of [0.98, 0.999], that is, λ may be a mixing ratio coefficient obtained by random sampling in the range of [0.98, 0.999].

[0041] The pseudo-new modulation placeholder is taken as an unknown modulation type signal and recorded as the K+1th input sample of the subsequent pre-trained modulation signal classification model.

[0042] The mixing ratio coefficient obtained by random sampling within a preset range ensures that the pseudo-new modulation placeholders are reasonably distributed.

[0043] S102. Input a plurality of known modulation type signals and unknown modulation type signals into a pre-trained modulation signal classification model, and train the pre-trained modulation signal classification model using a target loss function to obtain a trained modulation signal classification model.

[0044] Among them, the target loss function includes modulation classification loss, sample centroid comparison loss and sample maximum and minimum feature constraint loss. The pre-trained modulation signal classification model includes a feature extractor and a classifier. The classifier is a probability distribution classifier, which contains a fully connected layer and a Softmax layer.

[0045] Among them, a variety of known modulation type signals and unknown modulation type signals are used as input samples of the pre-trained modulation signal classification model.

[0046] Specifically, Figure 3 The schematic diagram of the training process of the pre-trained modulation signal classification model is shown schematically, see Figure 3 As shown in the figure, a variety of known modulation type signals and unknown modulation type signals are first input into the feature extractor, and the feature extractor outputs the corresponding feature representation set, which is a set of extracted features; the extracted features in the feature representation set have multiple types, which can be type 1, type 2, type 3 and type 4; then the extracted features in the feature representation set are input into a classifier containing a fully connected layer and a Softmax layer for training, and a trained modulation signal classification model is obtained. In the training process of the pre-trained modulation signal classification model, the target loss function used is composed of three parts, namely modulation classification loss, sample centroid contrast loss and maximum and minimum feature constraint loss, which form a multi-level feature distribution learning.

[0047] The specific calculation of the target loss function is as follows:

[0048] Specifically, the expression of the objective loss function is:

[0049]

[0050] in, is the target loss function, is the modulation classification loss, λ CL is the hyperparameter of the sample centroid contrast loss, is the sample centroid contrast loss, λ MMFC is the hyperparameter of the maximum and minimum feature constraint loss, is the maximum and minimum feature constraint loss.

[0051] Hyperparameter λ of sample centroid contrast loss CL and the hyperparameter λ of the maximum and minimum feature constraint loss MMFC , used to adjust the loss weight.

[0052] The expression of the modulation classification loss is:

[0053]

[0054] in, is the modulation classification loss, N is the total number of signals of various known modulation types and unknown modulation types used to calculate the loss during training, l() is the cross entropy loss, x i is the i-th input sample, f(x i ) is the pre-trained modulation signal classification model for the i-th input sample x i The predicted output, y i is the i-th input sample x i The corresponding true label.

[0055] The modulation type of the pseudo new modulation placeholder, that is, the modulation type of the unknown modulation type signal, is the K+1th type.

[0056] The expression of sample centroid contrast loss is:

[0057]

[0058] in, is the sample centroid contrast loss, Norm is the normalization operation, is the i-th input sample x i The sample features extracted by the feature extractor, is the i-th input sample x i Sample features extracted by feature extractor The center of mass of the object, τ is the temperature parameter, y i is the i-th input sample x i The corresponding true label, K is the total number of signals of various known modulation types, is the centroid vector of the kth known modulation type signal, N is the total number of signals of various known modulation type signals and unknown modulation type signals used to calculate the loss during training, 1[y i =k] is the indicator function, when the i-th input sample x i When it belongs to the kth class, the value of the indicator function is 1. When the i-th input sample x i When it does not belong to the kth category, the value of the indicator function is 0.

[0059] Norm is the value of the i-th input sample x i Sample features extracted by feature extractor Perform normalization operation.

[0060] Figure 4 The training diagram of the pre-trained modulation signal classification model is schematically shown, see Figure 4 As shown in the figure, the left side is the sample feature of the original feature space, that is, the input sample without using the sample centroid contrast loss for learning. The dark green point is the centroid of the three surrounding light green points. The three light green points are far away from the dark green point. Using the sample centroid contrast loss for learning the original feature space can make the i-th input sample x i Sample features extracted by feature extractor Close to the centroid The yellow and green points are operated in the same way, and the description of the green points can be seen in the three light green and dark green points on the right. The three light green points are close to the dark green points and far away from the other centroids. The yellow and green points have the same operation, and the description of the green points can be seen in the description of the green points.

[0061] Using sample centroid contrast loss for learning can make the i-th input sample x i Sample features extracted by feature extractor Close to the centroid And stay away from other centroids, optimize the feature distribution of known modulation type signals, enhance intra-class compactness and inter-class separability, and reserve space for unknown modulation type signals.

[0062] After the feature extractor, the sample centroid contrast loss can aggregate the feature representations of the input samples trained in the current training round, calculate the centroid vector of the modulation type signal based on the true label, and use the sample centroid contrast loss optimization to make the features of the input samples close to their centroids and away from other centroids. The sample centroid contrast loss includes sample features The normalization operation and temperature parameters can enhance the intra-class compactness and inter-class separability, reserve space for unknown modulation type signals, and help the pre-trained modulation signal classification model to better distinguish known and unknown modulation type signals, thereby optimizing the performance of the pre-trained modulation signal classification model in automatic modulation open set recognition.

[0063] The expression of the maximum and minimum feature constraint loss of the sample is:

[0064]

[0065] in, is the maximum and minimum feature constraint loss of the sample, To minimize the eigenvalue, is the most significant eigenvalue, D is the dimension of the sample feature extracted by the feature extractor for the i-th input sample, is the jth element in the sample feature extracted by the feature extractor for the i-th input sample, is the total length of the jth element in the sample feature extracted by the feature extractor for the i-th input sample.

[0066] In the sample maximum and minimum feature constraint loss (MMFC) part, based on the assumption that the features of known modulation type signals are significant and insignificant (that is, the amplitudes of the features of known modulation type signals are different) and the amplitudes of unknown signal features are relatively small, the most significant eigenvalue and the minimized eigenvalue of each input sample are calculated, and then the sample maximum and minimum feature constraint loss is used. Optimize and enhance the distinction between known modulation type signals and unknown modulation type signals, making the key feature responses of known modulation type signals more significant and the irrelevant feature responses of known modulation type signals less affected. This helps the pre-trained modulation signal classification model to more accurately identify and classify different modulation type signals, thereby improving the performance of the pre-trained modulation signal classification model in automatic modulation open set recognition tasks.

[0067] S103: Input multiple test modulation signals into the trained modulation signal classification model, and output multiple target feature embeddings.

[0068] Multiple target feature embeddings are recorded as

[0069] S104, classifying multiple test modulation signals according to multiple target feature embeddings, centroid vectors and adjustment coefficients of multiple known modulation type signals to obtain corresponding classification results.

[0070] Specifically, according to the embedding of multiple target features, the centroid vectors and adjustment coefficients of multiple known modulation type signals, multiple test modulation signals are classified to obtain corresponding classification results, including:

[0071] Step B1: Determine the minimum distance between multiple target features and the centroid vectors of multiple known modulation type signals.

[0072] The expression of the minimum distance between multiple target features and the centroid vector of multiple known modulation type signals is:

[0073]

[0074] Among them, dis ti′ is the minimum distance, is the embedding of multiple target features, i.e., the i′th target feature embedding, is the centroid vector of multiple known modulation type signals, that is, the centroid vector of the kth known modulation type signal, and K is the total number of multiple known modulation type signals.

[0075] Step B2: Determine a rejection threshold according to the minimum distance, the total number of multiple test modulation signals and the adjustment coefficient.

[0076] Specifically, the average distance of the multiple test modulation signals is calculated according to the minimum distance and the total number of the multiple test modulation signals; and the rejection threshold is determined according to the average distance and the adjustment coefficient.

[0077] The expression for the average distance of multiple test modulation signals is:

[0078]

[0079] Among them, dis avg is the average distance of multiple test modulation signals, N′ is the total number of multiple test modulation signals, and ti′ is the minimum distance.

[0080] The expression for the rejection threshold is:

[0081] th=β·dis avg ;

[0082] Among them, th is the rejection threshold, which is used to determine whether multiple test modulation signals belong to multiple known modulation type signals, β is the adjustment coefficient, that is, the adjustment coefficient of the rejection threshold, and dis avg is the average distance of multiple test modulation signals.

[0083] Step B3: Determine whether the minimum distance is greater than the rejection threshold. If so, the target feature embedding corresponding to the minimum distance belongs to the unknown modulation type signal. If not, the known modulation type signal to which the centroid vector of the known modulation type signal corresponding to the minimum distance belongs is used as the modulation type of the target feature embedding corresponding to the minimum distance.

[0084] When the minimum distance is greater than the rejection threshold, the target feature embedding corresponding to the minimum distance is determined to be an unknown modulation type signal, and the target feature embedding corresponding to the minimum distance is marked as K+1.

[0085] When the minimum distance is not greater than the rejection threshold, the known modulation type signal to which the centroid vector of the known modulation type signal closest to the target feature embedding corresponding to the minimum distance belongs is used as the modulation type embedded in the target feature corresponding to the minimum distance.

[0086] Based on the above Figure 1 It can be seen from the implementation method that the embodiment of the present invention processes known modulation type signals with different labels using a manifold hybrid algorithm to generate a pseudo-new modulation placeholder, and uses the pseudo-new modulation placeholder as an unknown modulation type signal; multiple known modulation type signals and unknown modulation type signals are input into a pre-trained modulation signal classification model to train the pre-trained modulation signal classification model using a target loss function to obtain a trained modulation signal classification model, wherein the target loss function includes modulation classification loss, sample centroid comparison loss, and sample maximum and minimum feature constraint loss; multiple test modulation signals are input into the trained modulation signal classification model to output multiple target feature embeddings; multiple test modulation signals are classified according to multiple target feature embeddings, the centroid vectors of multiple known modulation type signals, and adjustment coefficients to obtain corresponding classification results. In this way, through the pseudo-new modulation placeholders generated by the manifold mixing algorithm, the pre-trained modulation signal classification model can understand the concept of unknown modulation type signals, enhance the prediction ability, and accurately measure the classification effect of the pre-trained modulation signal classification model on known and unknown modulation type signals; and by proposing a multi-level feature distribution constraint including sample centroid contrast loss and sample-level maximum and minimum feature constraint loss, it helps the pre-trained modulation signal classification model to better distinguish between known and unknown modulation type signals, reserve space for unknown modulation type signals and amplify the difference, thereby optimizing the performance of the pre-trained modulation signal classification model in automatic modulation open set recognition, making it easy to balance and optimize the empirical risk of known modulation type signals and the open space risk of unknown modulation type signals.

[0087] Based on the same inventive concept, as an implementation of the above-mentioned multi-level feature distribution learning method for automatic modulation open set identification, an embodiment of the present invention also provides a multi-level feature distribution learning device for automatic modulation open set identification. Figure 5 is a structural diagram of the device in the embodiment of the present invention, see Figure 5 As shown, the device may include:

[0088] A generating module 501 is used to process known modulation type signals with different labels using a manifold mixing algorithm to generate a pseudo new modulation placeholder, and use the pseudo new modulation placeholder as an unknown modulation type signal, where the pseudo new modulation placeholder is a placeholder corresponding to the unknown signal forged according to the manifold mixing algorithm;

[0089] An input module 502 is used to input a plurality of test modulation signals into a trained modulation signal classification model, and output a plurality of target feature embeddings. The trained modulation signal classification model is to input a plurality of known modulation type signals and unknown modulation type signals into a pre-trained modulation signal classification model, so as to train the pre-trained modulation signal classification model using a target loss function. The obtained modulation signal classification model has a target loss function including a modulation classification loss, a sample centroid comparison loss, and a sample maximum and minimum feature constraint loss.

[0090] The classification module 503 is used to classify multiple test modulation signals according to multiple target feature embeddings, centroid vectors and adjustment coefficients of multiple known modulation type signals to obtain corresponding classification results.

[0091] The generation module 501 is specifically used to input the first known modulation type signal and the second known modulation type signal into the feature extractor respectively, so that the feature extractor outputs the corresponding first feature embedding and second feature embedding; the first feature embedding and the second feature embedding are processed according to the manifold mixing algorithm to generate a pseudo-new modulation placeholder, and the pseudo-new modulation placeholder is used as the unknown modulation type signal; the known modulation type signals with different labels include the first known modulation type signal and the second known modulation type signal.

[0092] The generating module 501 processes the first feature embedding and the second feature embedding according to the manifold mixing algorithm to generate a pseudo new modulation placeholder, including: processing the first feature embedding and the second feature embedding according to the manifold mixing algorithm, and generating the pseudo new modulation placeholder using the following first formula:

[0093]

[0094] in, is a pseudo-new modulation placeholder, λ is a mixing ratio coefficient obtained by random sampling within a preset range, is the first feature embedding, is the second feature embedding, y m is the modulation type of the first known modulation type signal, y n is the modulation type of the second known modulation type signal.

[0095] Input module 502, the expression of modulation classification loss is:

[0096]

[0097] in, is the modulation classification loss, N is the total number of signals of various known modulation types and unknown modulation types used to calculate the loss during training, l() is the cross entropy loss, x iis the i-th input sample, f(x i ) is the predicted output of the pre-trained modulation signal classification model for the i-th input sample, y i is the true label corresponding to the i-th input sample; a variety of known modulation type signals and unknown modulation type signals are used as input samples of the pre-trained modulation signal classification model.

[0098] Input module 502, the expression of sample centroid contrast loss is:

[0099]

[0100] in, is the sample centroid contrast loss, Norm is the normalization operation, is the sample feature extracted by the feature extractor for the i-th input sample, is the centroid of the sample feature extracted by the feature extractor for the i-th input sample, τ is the temperature parameter, y i is the true label corresponding to the i-th input sample, K is the total number of signals of multiple known modulation types, is the centroid vector of the kth known modulation type signal, N is the total number of signals of various known modulation type signals and unknown modulation type signals used to calculate the loss during training, 1[y i =k] is the indicator function. When the ith input sample belongs to the kth class, the value of the indicator function is 1. When the ith input sample does not belong to the kth class, the value of the indicator function is 0.

[0101] Input module 502, the expression of the maximum and minimum feature constraint loss of the sample is:

[0102]

[0103]

[0104] in, is the maximum and minimum feature constraint loss of the sample, To minimize the eigenvalue, is the most significant eigenvalue, D is the dimension of the sample feature extracted by the feature extractor for the i-th input sample, is the jth element in the sample feature extracted by the feature extractor for the i-th input sample, is the total length of the jth element in the sample feature extracted by the feature extractor for the i-th input sample.

[0105] Input module 502, the expression of the target loss function is:

[0106]

[0107] in, is the target loss function, is the modulation classification loss, λ CL is the hyperparameter of the sample centroid contrast loss, is the sample centroid contrast loss, λ MMFC is the hyperparameter of the maximum and minimum feature constraint loss, is the maximum and minimum feature constraint loss.

[0108] The classification module 503 is specifically used to determine the minimum distance of the centroid vectors of multiple target features embedded in multiple known modulation type signals; determine the rejection threshold according to the minimum distance, the total number of multiple test modulation signals and the adjustment coefficient; judge whether the minimum distance is greater than the rejection threshold, if so, the target feature embedding corresponding to the minimum distance belongs to the unknown modulation type signal, if not, the known modulation type signal to which the centroid vector of the known modulation type signal corresponding to the minimum distance belongs is used as the modulation type of the target feature embedding corresponding to the minimum distance.

[0109] It should be pointed out here that the above description of the embodiment of the multi-level feature distribution learning device for automatic modulation open set identification is similar to the description of the embodiment of the multi-level feature distribution learning method for automatic modulation open set identification, and has similar beneficial effects as the embodiment of the multi-level feature distribution learning method for automatic modulation open set identification. For technical details not disclosed in the embodiment of the multi-level feature distribution learning device for automatic modulation open set identification of the embodiment of the present invention, please refer to the description of the embodiment of the multi-level feature distribution learning method for automatic modulation open set identification of the present invention for understanding.

[0110] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A multi-level feature distribution learning method for automatic modulation open set recognition, characterized in that: include: A known modulation type signal with different labels is processed by a manifold mixing algorithm to generate a pseudo new modulation placeholder, and the pseudo new modulation placeholder is used as an unknown modulation type signal, wherein the pseudo new modulation placeholder is a placeholder corresponding to an unknown signal forged by the manifold mixing algorithm; Inputting a plurality of known modulation type signals and the unknown modulation type signal into a pre-trained modulation signal classification model, and training the pre-trained modulation signal classification model using a target loss function to obtain a trained modulation signal classification model, wherein the target loss function includes a modulation classification loss, a sample centroid comparison loss, and a sample maximum and minimum feature constraint loss; Inputting a plurality of test modulation signals into the trained modulation signal classification model, and outputting a plurality of target feature embeddings; According to the embedding of the multiple target features, the centroid vectors and adjustment coefficients of the multiple known modulation type signals, the multiple test modulation signals are classified to obtain corresponding classification results.

2. The multi-level feature distribution learning method for automatic modulation open set identification according to claim 1, characterized in that: The pre-trained modulation signal classification model includes a feature extractor and a classifier, the classifier is a probability distribution classifier, and the probability distribution classifier includes a fully connected layer and a Softmax layer.

3. The multi-level feature distribution learning method for automatic modulation open set identification according to claim 2, characterized in that: The known modulation type signals with different labels include a first known modulation type signal and a second known modulation type signal, and the known modulation type signals with different labels are processed by using a manifold mixing algorithm to generate a pseudo new modulation placeholder, and the pseudo new modulation placeholder is used as an unknown modulation type signal, including: Inputting the first known modulation type signal and the second known modulation type signal into the feature extractor respectively, so that the feature extractor outputs the corresponding first feature embedding and second feature embedding; The first feature embedding and the second feature embedding are processed according to the manifold mixing algorithm to generate the pseudo new modulation placeholder, and the pseudo new modulation placeholder is used as the unknown modulation type signal.

4. The multi-level feature distribution learning method for automatic modulation open set identification according to claim 3, characterized in that: The first feature embedding and the second feature embedding are processed according to the manifold mixing algorithm to generate the pseudo new modulation placeholder, including: The first feature embedding and the second feature embedding are processed according to the manifold mixing algorithm, and the pseudo new modulation placeholder is generated using the following first formula: in, is the pseudo new modulation placeholder, λ is the mixing ratio coefficient obtained by random sampling within a preset range, is the first feature embedding, is the second feature embedding, y m is the modulation type of the first known modulation type signal, y n is the modulation type of the second known modulation type signal.

5. The multi-level feature distribution learning method for automatic modulation open set identification according to claim 1, characterized in that: Using the plurality of known modulation type signals and the unknown modulation type signal as input samples of the pre-trained modulation signal classification model; The expression of the modulation classification loss is: in, is the modulation classification loss, N is the total number of the plurality of known modulation type signals and the unknown modulation type signals used to calculate the loss during the training process, is the cross entropy loss, x i is the i-th input sample, f(x i ) is the predicted output of the pre-trained modulation signal classification model for the i-th input sample, y i is the true label corresponding to the i-th input sample.

6. The multi-level feature distribution learning method for automatic modulation open set identification according to claim 2, characterized in that: The expression of the sample centroid contrast loss is: in, is the sample centroid contrast loss, Norm is the normalization operation, is the sample feature extracted by the feature extractor for the i-th input sample, is the centroid of the sample feature extracted by the feature extractor for the i-th input sample, τ is the temperature parameter, y i is the true label corresponding to the i-th input sample, K is the total number of the multiple known modulation type signals, is the centroid vector of the kth known modulation type signal, N is the total number of the plurality of known modulation type signals and the unknown modulation type signals used to calculate the loss during the training process, 1[y i =k] is an indicator function. When the ith input sample belongs to the kth class, the value of the indicator function is 1; when the ith input sample does not belong to the kth class, the value of the indicator function is 0.

7. The multi-level feature distribution learning method for automatic modulation open set identification according to claim 6, characterized in that: The expression of the maximum and minimum feature constraint loss of the sample is: in, is the maximum and minimum feature constraint loss of the sample, To minimize the eigenvalue, is the most significant eigenvalue, D is the dimension of the sample feature extracted by the feature extractor for the i-th input sample, is the jth element in the sample feature extracted by the feature extractor for the i-th input sample, is the total length of the j-th element in the sample feature extracted by the feature extractor for the i-th input sample.

8. The multi-level feature distribution learning method for automatic modulation open set identification according to claim 5, characterized in that: The expression of the objective loss function is: in, is the objective loss function, is the modulation classification loss, λ CL is the hyperparameter of the sample centroid contrast loss, is the sample centroid contrast loss, λ MMFC is the hyperparameter of the maximum and minimum feature constraint loss, is the maximum and minimum feature constraint loss.

9. The multi-level feature distribution learning method for automatic modulation open set identification according to claim 6, characterized in that: The method of classifying the multiple test modulation signals according to the multiple target feature embeddings, the centroid vectors of the multiple known modulation type signals, and the adjustment coefficients to obtain corresponding classification results includes: Determining the minimum distances between the plurality of target features and the centroid vectors of the plurality of known modulation type signals; determining a rejection threshold according to the minimum distance, the total number of the plurality of test modulation signals and the adjustment factor; Determine whether the minimum distance is greater than the rejection threshold. If so, the target feature embedding corresponding to the minimum distance belongs to the unknown modulation type signal. If not, the known modulation type signal to which the centroid vector of the known modulation type signal corresponding to the minimum distance belongs is used as the modulation type of the target feature embedding corresponding to the minimum distance.

10. A multi-level feature distribution learning device for automatic modulation open set recognition, characterized in that: include: A generating module, used for processing known modulation type signals with different labels by using a manifold mixing algorithm to generate a pseudo new modulation placeholder, and using the pseudo new modulation placeholder as an unknown modulation type signal, wherein the pseudo new modulation placeholder is a placeholder corresponding to an unknown signal forged by the manifold mixing algorithm; An input module is used to input multiple test modulation signals into the trained modulation signal classification model and output multiple target feature embeddings, wherein the trained modulation signal classification model is to input multiple known modulation type signals and the unknown modulation type signal into a pre-trained modulation signal classification model so as to train the pre-trained modulation signal classification model using a target loss function, and the obtained modulation signal classification model, wherein the target loss function includes a sample modulation classification loss, a centroid comparison loss, and a sample maximum and minimum feature constraint loss; The classification module is used to classify the multiple test modulation signals according to the multiple target feature embeddings, the centroid vectors and adjustment coefficients of multiple known modulation type signals to obtain corresponding classification results.

Images