An online automatic modulation classification method based on isolated distribution kernels

Abstract

This invention discloses an online automatic modulation classification method based on isolated distributed kernels. The distributed kernels characterize the baseband signal, and an AMC (Automatic Modulation Classification) technique is introduced, which performs well in online environments under real-world time-varying channel conditions. Extensive experiments in online environments demonstrate the effectiveness of the proposed classifier. The proposed method outperforms existing baseline models, incorporating two advanced deep learning classifiers. Furthermore, this invention implements an online AMC classifier with linear time complexity, significantly improving real-time application efficiency.

Description

Technical Field

[0001] This invention relates to an online automatic modulation classification method based on isolated distribution kernels. Background Technology

[0002] Automatic Modulation Classification (AMC) is a key technology in modern non-cooperative communication networks. AMC algorithms enable RF receivers to automatically identify the modulation format of received signals. In CR networks, limited spectrum resources are allocated flexibly and adaptively. Transmitters can change their modulation format to adapt to current external channel conditions, thereby optimizing transmission quality and bandwidth utilization. Receivers need to know which modulation format is currently in use to correctly decode the signal. Modulation format information can be inserted into the transmitted signal frame, but this method consumes additional spectrum resources, thus reducing spectral efficiency. AMC addresses this problem more effectively. At the receiver, AMC continuously monitors the current modulation format, ensuring that the signal is correctly demodulated and the transmitted data is successfully delivered. AMC eliminates the overhead of transmitting modulation format information in the network protocol, effectively improving the overall capacity of the CR network.

[0003] Traditional AMC methods can be divided into likelihood-based (LB) classifiers and feature-based (FB) classifiers. LB classifiers calculate the likelihood of each modulation hypothesis and select the one with the highest likelihood as the classification result. When the receiver knows perfect channel parameters, LB classifiers have proven to be optimal in terms of classification accuracy. However, LB classifiers have high computational complexity, and their performance degrades significantly in real-world channel environments where precise channel parameters are unavailable. In contrast, FB classifiers determine the modulation format based on some engineering characteristics of the received signal (e.g., spectral features, wavelet transform-based features, and higher-order statistical features). FB classifiers are generally easier to implement, but their accuracy is lower compared to LB classifiers.

[0004] Recently, several AMC (Aspect-Oriented Computation) classifiers based on deep learning (DL) have been proposed. DL classifiers are also a form of FB (Flash Beam) classifiers because they automatically extract features in some way. DL classifiers train artificial neural networks to perform classification tasks. Different neural network architectures have been applied to AMC tasks, such as one-dimensional and two-dimensional convolutional neural networks (CNNs) and recurrent neural networks (RNNs). DL classifiers have been shown to achieve good classification performance even under very weak signal conditions. However, all these DL classifiers share some common drawbacks that impair their effectiveness and practicality in real-world online applications:

[0005] The architecture of deep learning classifiers is extremely complex. As is well known, tuning the hyperparameters of a network is very time-consuming.

[0006] Existing deep learning (DL) models are batch learning algorithms, trained on fixed datasets, and cannot perform online model updates to adapt to constantly changing channel environments. Time-varying channel conditions distort received signals in terms of amplitude, phase, etc. This leads to a channel condition mismatch between training and test data, thus reducing the performance of batch learning algorithms such as DL classifiers in AMC (Aspect-Controlled Mechanism). This problem is best addressed by introducing online model update mechanisms. However, integrating such mechanisms further increases the complexity of DL classifiers. For example, an online retraining mechanism has been proposed to compensate for channel impairments, but it requires additional network training and storage overhead. In summary, the complexity of DL models is a major obstacle to designing practical online AMC schemes.

[0007] Currently, most signal representations used in deep learning are indirect representations of signals because each representation requires some transformation of the original signal. Summary of the Invention

[0008] Objective: To address the problems and shortcomings of existing technologies, this invention provides an online automatic modulation classification method based on the Isolated Distribution Kernel (IDK). The constellation diagram of a baseband signal sequence can be intuitively and effectively represented as a distribution without losing information. This allows the use of distribution metrics to characterize both baseband signals and to measure the similarity between any two sets of baseband signals. A novel kernel function called the Isolated Distribution Kernel (IDK) is used to characterize the distribution shown in the constellation diagram. This invention demonstrates that the IDK is a better representation than existing sequence and image representations, as well as manually designed features.

[0009] By combining IDK with online gradient descent (OGD), an online AMC scheme is created. The proposed IDK-OGD classifier continuously updates its model based on new signal samples at a low time cost, making it particularly efficient and effective in time-varying channel environments.

[0010] Technical Solution: An online automatic modulation classification method based on isolated distribution kernels (IDK) treats the in-phase (I) and quadrature (Q) components of the baseband signal as independent identically distributed (iid) points sampled from an unknown distribution. These components are transformed into a set of two-dimensional points. Using these iidally distributed points, IDK automatically creates features to characterize the distribution.

[0011] It includes the following steps:

[0012] Step 1: Assume the signal to be automatically modulated and classified belongs to m different candidate modulation formats M. j , j∈[m] will be the weight vector ω of the corresponding m OGD (Online Gradient Descent) classifiers j Initialize to 0;

[0013] Step 2: Using a batch of initial training data, filter out those belonging to modulation format M. j Signals j∈[m] are used as data points to construct the IDK (Isolation Distribution Kernel Function) space partition, thereby constructing m different IDK feature maps. The subscript j indicates that the partitioning of this IDK feature map is done using a modulation format of M. j The data was constructed from it;

[0014] Step 3: Use the initial training data as the first batch of signals to update the OGD classifier. The specific update process is the same as the next step.

[0015] Step 4, when working in an online environment, each time a new batch of signals B is input:

[0016] 1) Classification stage: For each signal S i ∈B:

[0017] 11) Its corresponding distribution Through feature mapping of each IDK Convert to IDK feature vector

[0018] 12) Calculate signal S respectively i It belongs to the modulation format M j Predicted score

[0019] 13) The one with the highest predicted score As signal S i The prediction results are used to output the predicted modulation format.

[0020] 2) Online model update phase: For each signal S i ∈B, if its corresponding baseline true value label can be obtained That is, the actual modulation format:

[0021] 21) Determine if the prediction is correct. If the predicted modulation format is the same as the actual modulation format, then let k be... i If k is 1, otherwise k i =-1;

[0022] 22) Calculate the prediction score g for each modulation scheme using the loss function L. i,j gradient

[0023] Update the weight vector of each OGD classifier based on the learning rate η:

[0024] Attached Figure Description

[0025] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0026] Figure 2 This is a schematic diagram of a distribution-based signal representation according to an embodiment of the present invention. Detailed Implementation

[0027] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0028] like Figure 1 As shown, an online automatic modulation classification method based on isolated distribution kernels includes the following steps:

[0029] Step 1: Assume the signal to be automatically modulated and classified belongs to m different candidate modulation formats M. j For j∈[m], the weight vector ω of the corresponding m OGD (Online Gradient Descent) classifiers is... j Initialize to 0;

[0030] Step 2: Using a batch of initial training data, filter out those belonging to modulation format M. j Signals j∈[m] are used as data points to construct the IDK (Isolation Distribution Kernel Function) space partition, thereby constructing m different IDK feature maps.

[0031] Step 3: Use the initial training data as the first batch of signals to update the OGD classifier. The specific update process is the same as the next step.

[0032] Step 4, when working in an online environment, each time a new batch of signals B is input:

[0033] 1) Classification stage: For each signal S i ∈B:

[0034] 11) Its corresponding distribution Through feature mapping of each IDK Convert to IDK feature vector

[0035] 12) Calculate signal S respectively i It belongs to the modulation format M j Predicted score

[0036] 13) The one with the highest predicted score As signal S iThe prediction results are used to output the predicted modulation format.

[0037] 2) Online model update phase: For each signal S i ∈B, if its corresponding baseline true value label can be obtained That is, the actual modulation format:

[0038] 21) Determine if the prediction is correct. If the predicted modulation format is the same as the actual modulation format, then let k be... i If k is 1, otherwise k i =-1;

[0039] 22) Calculate the prediction score g for each modulation scheme using the loss function L. i,j gradient

[0040] Update the weight vector of each OGD classifier based on the learning rate η:

[0041]

[0042] The signal is characterized as a distribution in the I / Q signal space.

[0043] like Figure 2 As shown in section II, IDK treats the in-phase (I) and quadrature (Q) components of the baseband signal as independent and identically distributed (iid) points sampled from an unknown distribution. These components are transformed into a set of two-dimensional points. Using these iidally distributed points, IDK automatically creates features to characterize the distribution.

[0044] Signals in the I / Q signal space: Generally speaking, the baseband signal r(t) received from a wireless channel can be expressed as:

[0045] r(t)=s(t)*c(t)+w(t))(1)

[0046] Where s(t) represents the transmitted baseband signal, c(t) is the channel impulse response, w(t) represents additive noise, and * is the convolution operator.

[0047] The transmitted digital amplitude-phase (A / P) modulated baseband signal s(t) with N symbols can be expressed as:

[0048]

[0049] Where j represents the imaginary sign, a n and These represent the amplitude and phase of the nth symbol, respectively; g(t) represents the pulse shaping filter; T s It is the duration of the symbol; I n and Q nThese represent the in-phase (I) and quadrature (Q) components of the nth symbol, respectively.

[0050] Essentially, the difference between the two different A / P modulation formats lies in the A / P pairs for all N symbols. The arrangement.

[0051] As shown in equation (3), the A / P modulated signal consists of a series of complex data points (referred to as the sequence), where the real and imaginary parts are the in-phase (I) and quadrature (Q) components of the signal, respectively (e.g., ...). Figure 2 (As shown in I). This sequence can be represented as a set of points in a two-dimensional complex space of I and Q axes. This set of points in the I / Q space is usually called a constellation diagram. It can be clearly seen from equation (3) that the constellation diagram of the A / P modulated signal contains all the amplitude and phase information of the signal. Therefore, the constellation diagram is equivalent to the original signal sequence under the AMC environment, without information loss.

[0052] In this embodiment, the points in the I / Q space (or constellation diagram) can be regarded as iid points x generated by the unknown distribution P in the I / Q space, that is, x ~ P.

[0053] The signal s(t) in the I / Q component is a set of points S in the I / Q space (hereinafter referred to as signal S); a signal S of a certain modulation format corresponds to an unknown distribution P. S .

[0054] Therefore, the task of AMC to identify the unknown modulation format of signal S can be accomplished by maximizing distribution P. S and m candidate modulation formats {M1, M2, ..., M m The distribution P of modulation scheme α in} Sα This is achieved through similarity. It can be formally represented as:

[0055]

[0056] Where K is the kernel function that measures the similarity between two distributions; and S α It is the iid sample set of the signal with modulation format α.

[0057] Equation (4) returns to P S Most similar P Sα The modulation format α.

[0058] The above is the definition of the AMC task. An algorithm that can complete the AMC task should be able to "given a signal S whose modulation format needs to be identified (assuming its distribution is P)". S From the candidate modulation formats (each modulation format corresponds to a different distribution), select the distribution and P. SThe most similar one is the output of the modulation format of the predicted S (because ideally, all signals belonging to the same modulation format have the same distribution).

[0059] The IDK-OGD proposed in this invention is an algorithm capable of performing this AMC task.

[0060] Use distributed kernel functions to perform distribution-based representations / mappings.

[0061] Point-to-point kernel functions measure the similarity between two points. Each kernel uniquely determines a feature map that maps points in the input space. The point Ф(x) is mapped to a high-dimensional feature space, which is called the reproducing kernel Hilbert space.

[0062] The Isolation Kernel (IK) is a point-to-point kernel function. The IK feature map is constructed by randomly partitioning the data using an isolation mechanism. The parameters of the IK feature map include ψ and t, where ψ is the number of partitions in each partition, and t is the number of partitions.

[0063] Specifically, let's set Given a dataset, randomly select points {z} from D. i}(i=1,2,…,ψ), each point z i All through the hypersphere θ i Isolated from the remaining ψ-1 points, the radius r of the hypersphere i It is z i The distance between it and its nearest neighbor among ψ-1 points. Thus, each point in space either belongs to only one of the hyperspheres or does not belong to any hypersphere (if a point happens to be in two or more hyperspheres, then it belongs to the hypersphere closest to it).

[0064] The feature of dividing j to IK by ψ points is defined by the following formula (5), where 1 <= j <= t. Let φ j (x|D) is a ψ-dimensional binary column vector, defined as follows:

[0065]

[0066] Where d min Let x be in the point set {z} i In the set of} (i = 1, 2, ..., ψ), the nearest neighbor z k The distance between them.

[0067] Thereafter, '|D' is omitted to avoid ambiguity.

[0068] Using formula (5) t times to obtain t column vectors, we can generate IKφ: x→{0,1} t×ψThe complete feature mapping. For each x∈D, partition according to formula (5) in the j-th partition, and create a ψ-dimensional vector φ. j (x). The complete feature mapping point Ф(x) is φ1(x),···,φ t The concatenation of (x), i.e.

[0069]

[0070] The feature map of the isolated distribution kernel (IDK) will generate a set of points from the distribution PS (sampled). The feature vector is mapped to a point in the Hilbert space H, and this feature mapping is obtained by using kernel mean embedding of IK. The IDK feature vector is as follows:

[0071]

[0072] Each signal point x can be transformed into an IK eigenvector Ф(x) through an IK eigenmap φ. This yields a signal S = {x1, x2, ..., x...}. L The IK eigenvectors of all signal points in} are {φ1(x), ..., φ L After (x)}, the IDK feature vector of this signal is obtained using Equation 7. P here S This is the distribution corresponding to the signal S. Although IDK does not actually model this distribution, it uses the sample set S obtained from the distribution to represent the distribution P. S This is mapped to a feature vector. Thus, a signal (a set of points consisting of many signal points) is transformed into a vector.

[0073] Use IDKκ to measure the two distributions P S and P T The similarity between them is:

[0074]

[0075] Using IDK to characterize the I / Q baseband signal, OGD (Online Gradient Descent) is applied to the AMC task:

[0076] (1) Use IDK to process I / Q baseband signals.

[0077] (2) Convert the binary classifier to use an m-class classifier (consisting of m binary classifiers) to handle m modulation formats.

[0078] (3) Each batch of incoming I / Q baseband signals has a clear classification stage and model update stage, so that the model is updated only when there is a reference true value label.

[0079] The working principle of the proposed IDK-OGD classifier will be introduced here. The functional flowchart is as follows: Figure 1 As shown. Assume there are m different candidate modulation schemes, denoted as M1, M2, ..., Mn. m First, a batch of initial training data is needed to construct m different IDK feature maps for each modulation format. That is, based on equations (5) to (7), the modulation samples M in the initial training data are used. j Construct feature mapping

[0080] After obtaining each IDK mapping point for each modulation format, all binary classifiers are used to generate a prediction score, and the final classification result is the predicted category with the highest score.

[0081] When working in an online environment, after a batch of samples (i.e., I / Q baseband signals) are classified by the OGD classifier, the classifier model is only updated when the true modulation format labels of these samples are available.

[0082] Algorithm provides pseudocode for the proposed IDK-OGD classifier in this embodiment, which details the technical aspects of the classification and online update phases of the proposed scheme. Notably, since the computations corresponding to each modulation format (i.e., the two for loops in lines 4-6 and 12-14) are independent, the proposed IDK-OGD can be implemented using a parallel architecture, which significantly improves efficiency.

[0083] Algorithm1IDK-OGD

[0084] Input:

[0085] m - Number of modulation formats

[0086] -IDK feature vector of modulation format j

[0087] η - learning rate

[0088] Output:

[0089]

[0090]

Claims

1. An online automatic modulation classification method based on isolated distribution kernels, characterized in that, Includes the following steps: Step 1, assume the signal that needs to be automatically modulated and classified belongs to m Different candidate modulation formats M j Where j∈[1, m], the corresponding m The weight vector of each OGD classifier ω j Initialize to 0; Step 2: Using a batch of initial training data, filter out those belonging to the modulation format. M j The signals, where j∈[1, m], are used as data points to construct the IDK space partition, thereby constructing... Different IDK feature maps ; Step 3: Use the initial training data as the first batch of signals to update the OGD classifier; The OGD classifier update process is as follows: When working in an online environment, each time a new batch of signals is input... B : 1) Classification stage: For each signal S i ∈ B : 11) Its corresponding distribution Through feature mapping of each IDK Convert to IDK feature vector ; 12) Calculate the signals respectively S i It belongs to the modulation format M j Predicted score g i,j = ; 13) The one with the highest predicted score =argmax j g i,j As a signal The prediction results are used to output the predicted modulation format. ; 2) Online model update phase: For each signal S i ∈ B If the corresponding benchmark true value label can be obtained That is, the true modulation format: 21) Determine if the prediction is correct. If the predicted modulation format is the same as the actual modulation format, then assume... k i =1, otherwise k i =-1; 22) Calculate the loss function L Prediction scores for each modulation format g i,j gradient ; Update the weight vector of each OGD classifier based on the learning rate η: 。 2. The online automatic modulation classification method based on isolated distribution kernels according to claim 1, characterized in that, In step 2, the I / Q baseband signal is processed using IDK to convert the baseband signal into an in-phase signal. I and orthogonal Q The components are considered as independent and identically distributed points sampled from an unknown distribution. Using these independent and identically distributed points, IDK will automatically create features to characterize the distribution. Points in the I / Q space are considered as iid points generated by an unknown distribution P in the I / Q space. x ,Right now x ~ P ; Signal in I / Q components s ( t () is the set of points in I / Q space. S ; A modulation format signal S Corresponding to an unknown distribution P S ; AMC identification signal S The task of unknown modulation format is to maximize the distribution P S and m Candidate modulation formats { M 1, M 2, … , M m Modulation format in} α Distribution To achieve, to obtain and P S Most similar modulation format α .

3. The online automatic modulation classification method based on isolated distribution kernels according to claim 1, characterized in that, For each signal S i Distribution corresponding to the signal Through feature mapping of each IDK Where j∈[1,m] is transformed into IDK feature vectors : Each signal point x in the signal is mapped through the IK feature map. φ Transform into an IK eigenvector Ф ( x This serves as the basis for subsequently constructing the IDK feature vector; a signal S={x1, x2, …, x} is obtained. L The IK eigenvectors of all signal points in} {φ1( x ), · · · , φ L ( x Then, kernel mean embedding is performed using formula (7) to obtain the IDK feature vector of this signal. P here S It is the distribution corresponding to signal S; where IDK is the isolation distribution kernel and IK is the isolation kernel; The parameters of the IK feature map include ψ and t Where ψ is the number of partitions in each partitioning, t It is to divide the number of times; Let D Given a dataset; randomly select a set of points from D { z i }, i=1,2,…,ψ, each point z i All through hypersphere θ i Isolated from the remaining ψ-1 points, the radius of the hypersphere r i yes z i The distance between it and its nearest neighbor among ψ-1 points; Divided by ψ points j The obtained IK eigenvectors are defined as follows (5), 1 <= j <= t; let φ j (x|D) is a ψ-dimensional binary column vector, defined as follows: （5） in d min for x In the point set { z i } and its nearest neighbor z k The distance between them; The process of obtaining t column vectors by using formula (5) t times corresponds to the complete IK feature mapping. φ : x → {0,1} t×ψ For each x ∈ D, in the j In this partitioning process, the partitions are determined according to formula (5), and a ψ-dimensional vector is created. φ j ( x ); complete IK eigenvector Ф ( x ) is φ1 ( x ), · · · , φ t ( x ) in series, i.e. （6） Based on the IK feature vector, the IDK feature mapping will distribute P S Mapping to a point in the Hilbert space H, which is obtained using the kernel mean embedding formula, the IDK feature vector is: （7）。

Images