Signal open set modulation identification method based on deep clustering

Abstract

The invention discloses a signal open set modulation identification method based on deep clustering, and relates to the technical field of artificial intelligence and signal processing, and the method comprises the steps: obtaining data of at least two different representation forms of a known modulation type signal, and carrying out the processing of a deep neural network, so as to generate a high-order semantic feature vector; and training the deep neural network based on the feature vectors and preset clustering anchor points representing each known modulation category, the training being aimed at increasing the distance between different known categories and reducing the distance in the same category in a feature space, and enabling the feature vectors of each known category to converge to the corresponding clustering anchor points. During identification, the feature vector of a to-be-identified signal is extracted, and the to-be-identified signal is judged to be a certain known modulation category or an unknown modulation category according to the position relation between the to-be-identified signal and each clustering anchor point. According to the method, the known and unknown modulation signals can be efficiently and accurately identified by constructing the structured feature space, and the robustness is high.

Description

Technical Field

[0001] The present invention relates to the fields of communication signal processing and artificial intelligence, and in particular to a signal open-set modulation recognition method based on deep clustering. Background Art

[0002] Automatic Modulation Classification (AMC) is a key technology in wireless communications, spectrum monitoring, electronic countermeasures, and other fields. Its core task is to automatically determine the modulation scheme of a received signal. Traditional modulation classification methods rely primarily on expert experience and signal processing techniques, such as those based on maximum likelihood estimation, feature extraction based on cyclic spectra, and methods based on instantaneous statistical features. These methods achieve good results under ideal channel conditions and high signal-to-noise ratios. However, in complex real-world electromagnetic environments, signals are often severely interfered with by factors such as noise, multipath propagation, and channel fading, causing the performance of traditional methods to plummet. Furthermore, these methods rely on manually designed features, have weak generalization capabilities, and are unable to cope with the increasingly diverse modulation types.

[0003] In recent years, artificial intelligence technologies, particularly deep learning, have demonstrated tremendous potential in signal modulation recognition thanks to their powerful adaptive feature extraction and nonlinear mapping capabilities. Models based on deep neural networks are able to learn deep, more robust features directly from raw signals, achieving remarkable success in "closed-set" recognition tasks. This refers to a model's task of classifying all encountered signals into a predefined, limited set of known modulation types.

[0004] However, closed-set recognition models suffer from a fundamental flaw: they assume that all signals under test belong to a known set of categories. In real-world applications, recognition systems inevitably encounter new or unknown modulation signals never seen during training. In these cases, closed-set models incorrectly force these unknown signals into a known category, severely distorting recognition results and compromising system reliability.

[0005] Therefore, how to enable modulation recognition systems to accurately identify known modulation types while effectively detecting and distinguishing unknown modulation types not present in the training set—that is, achieving "open-set recognition" (OSR)—is a key challenge and technical bottleneck in the current field. Existing open-set recognition efforts, such as the Chinese invention patent publication number CN113052126B, employ a multi-stage approach: first, classifiers or thresholding methods are used to identify known classes and filter out suspected unknown signals. These filtered unknown signals are then subjected to secondary clustering or analysis using independent networks (such as Siamese neural networks). While these approaches achieve some degree of open-set recognition, they often suffer from complex system structures, disconnected training and recognition processes, and difficulty in coordinating optimization across multiple models or stages. More importantly, these approaches can lack a clear and robust feature space boundary between known and unknown classes, which impacts the detection rate and recognition accuracy of unknown signals under complex conditions. Summary of the Invention

[0006] In order to solve the problems existing in the above-mentioned prior art, the present invention provides the following technical solutions, including: A signal open set modulation recognition method based on deep clustering includes the following steps: S1. Obtain data of at least two different representations of a signal of known modulation category and process them through a deep neural network to generate a high-order semantic feature vector; S2. Based on the high-order semantic feature vectors and the cluster anchor points representing each known modulation category, training the deep neural network to increase the distance between different known modulation categories and reduce the distance within the same known modulation category in the feature space, and make the high-order semantic feature vectors of each known modulation category converge to their corresponding cluster anchor points; S3. Obtaining data of at least two different representations of the signal to be identified, and extracting its high-order semantic feature vector using the deep neural network; S4. Based on the positional relationship between the high-order semantic feature vector of the signal to be identified and the cluster anchor point in the feature space, determine whether the category of the signal to be identified is one of the known modulation categories, or determine whether it is an unknown modulation category.

[0007] Preferably, the two data in different representation forms include: IQ data of the signal, frequency domain data calculated based on the IQ data, and phase data.

[0008] Preferably, the method for training the deep neural network includes: training the deep neural network by combining a loss function; the combined loss function includes: Contrastive distance loss, which is used to reduce the feature distance between similar signal samples and increase the feature distance between heterogeneous signal samples during training; An adaptive clustering loss is used to minimize the distance between the high-order semantic feature vector and the cluster anchor point corresponding to the known modulation category during training.

[0009] Preferably, the method for calculating the adaptive clustering loss includes: For each known modulation category, calculate the center point of all high-order semantic feature vectors it contains; Repeatedly find the closest pair of all currently unmatched center points and unmatched cluster anchor points, and pair them until all center points are matched; Determining, based on the pairing relationship, a cluster anchor point corresponding to each high-order semantic feature vector in the training batch; The mean distance between all high-order semantic feature vectors and their corresponding cluster anchor points is calculated as the adaptive clustering loss.

[0010] Preferably, the method for acquiring the cluster anchor points includes: performing voting statistics on the high-order semantic feature vectors of all training samples of each known modulation category to determine the cluster anchor points of the category.

[0011] Preferably, the method for determining whether the category of the signal to be identified is one of the known modulation categories includes: If the proportion of high-order semantic feature vectors corresponding to the same cluster anchor point in multiple samples included in the signal to be identified exceeds a preset voting threshold, the modulation category of the signal to be identified is determined to be the known modulation category corresponding to the cluster anchor point.

[0012] Preferably, step S4 further includes binarizing the high-order semantic feature vector of the signal to be identified before determining the positional relationship.

[0013] Preferably, the binarization post-processing is: comparing the numerical value in the high-order semantic feature vector with a preset threshold, setting the value to 0 if it is less than the threshold, and setting it to 1 otherwise.

[0014] Beneficial effects 1. This invention achieves both accurate recognition of known modulation classes and effective detection of unknown classes through the design of a single, end-to-end trained deep neural network. The entire open-set recognition capability is internalized within a single model, eliminating the need for additional processing stages or independent discrimination modules. This simplifies the system architecture, reduces implementation complexity, and improves overall operational efficiency.

[0015] 2. This invention employs a unique dual-objective training approach: on the one hand, it increases the feature distances between different known classes; on the other hand, it utilizes "cluster anchors" to make the feature distribution of each known class more compact and stable. This training approach constructs a well-defined feature space, enabling the feature vectors of unknown signals to be effectively identified based on their significant deviation from the cluster anchors of all known classes, significantly improving the performance of open-set recognition.

[0016] 3. Multimodal feature fusion can capture the inherent laws of the signal from different dimensions, forming a more comprehensive and richer feature representation than a single modality. It can better overcome interference from complex factors such as noise and channel fading in practical applications, further enhancing the overall performance and environmental adaptability of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is a flow chart of a signal open set modulation identification method based on deep clustering provided in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described below with reference to the accompanying drawings. In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description. They do not indicate or imply that the devices or components referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, they should not be understood as limiting the present invention.

[0019] In modern wireless communications, accurately identifying a signal's modulation scheme is crucial for ensuring communication quality, spectrum monitoring, and signal reconnaissance. Traditional recognition methods often suffer from limited performance and generalization capabilities when faced with complex, real-world electromagnetic environments. While deep learning has achieved recent success in "closed-set" modulation recognition (i.e., classifying from a set of pre-defined known classes), it struggles with "open-set" problems. When faced with signals of unknown modulation types not seen during training, closed-set models can produce significant misclassifications, which is unacceptable in practical applications. For example, a closed-set system trained only to recognize BPSK and QPSK signals might incorrectly identify a 16QAM signal as either BPSK or QPSK, leading to complete failure of subsequent demodulation and information retrieval. Existing open-set recognition attempts have limited performance and robustness due to complex system architectures, the need for multi-stage processing, and unclear demarcation of the feature boundaries between known and unknown classes.

[0020] To address these technical challenges, this paper proposes a method for open-set modulation recognition based on deep clustering. The core idea of ​​this method is to construct a structured feature space through a unique end-to-end training framework, where known modulation classes form highly cohesive, fixed clusters within the space. This allows any unknown signal that does not belong to these clusters to be easily distinguished.

[0021] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the specific embodiments described herein are only used to explain the present invention and do not constitute any form of limitation to its protection scope.

[0022] Example 1 like Figure 1 As shown, this embodiment will show in detail the complete process of a signal open set modulation recognition method based on deep clustering, which mainly includes two stages: model training and open set recognition.

[0023] Phase 1: Model training.

[0024] The purpose of model training is to build a deep neural network model that can map signals of different modulation types to different regions in the feature space.

[0025] S1. Obtain data of at least two different representations of a signal of known modulation category and process them through a deep neural network to generate a high-order semantic feature vector.

[0026] First, a dataset containing a variety of known modulation type signals needs to be prepared for training. In this embodiment, it is assumed that the known modulation type set includes five types: BPSK, QPSK, 8PSK, 16QAM, and 64QAM. In order to enable the neural network to understand the signal from a richer dimension, the present invention preferably adopts a multimodal input strategy. Specifically, for each known modulation type signal, the following operations are performed to obtain data in at least two different representations: 1. Obtain raw IQ data: Directly obtain the time series data of the in-phase (I) and quadrature (Q) components of the signal.

[0027] 2. Calculate frequency domain data: Perform a fast Fourier transform (FFT) on the IQ data to obtain the spectrum representation of the signal, i.e., the frequency domain data.

[0028] 3. Calculate phase data: Solve the phase of the IQ data to obtain the phase information sequence of the signal.

[0029] In this way, the original one-dimensional time-domain signal is converted into a dataset containing three modalities: IQ data, frequency domain data, and phase data, forming the multimodal input data for the training phase. This greatly enhances the representation ability of information, as frequency and phase information can reveal implicit features that are difficult to directly observe in the time-domain signal.

[0030] Before entering the network, these multimodal data typically need to be preprocessed. For example, this involves scaling all data to a uniform scale through minimum and maximum normalization, and slicing them into fixed-length segments to form standardized training samples. For example, each signal file can be split into several sample data groups of 8192 bytes, each containing three components: IQ, frequency, and phase.

[0031] This processed multimodal sample data is then fed into a deep neural network. This network is preferably a multimodal convolutional neural network with parallel internal branches that can process IQ, frequency, and phase data separately, effectively fusing the features of these different modalities deep within the network. After forward propagation, the network generates a high-order semantic feature vector for each input training sample. This vector is a compact, abstract representation of the original signal in the deep feature space.

[0032] S2. Based on the high-order semantic feature vector and the cluster anchor points representing each known modulation category, the deep neural network is trained to increase the distance between different known modulation categories and reduce the distance within the same known modulation category in the feature space, and to make the high-order semantic feature vector of each known modulation category converge to its corresponding cluster anchor point.

[0033] This step is the core of our invention. Rather than using traditional classification losses (such as cross-entropy), the network is trained through a specially designed dual-objective optimization process to finely construct the structure of the feature space. This process is based on the high-order semantic feature vectors of the current batch of samples and pre-defined clustering anchors.

[0034] The cluster anchors are pre-set coordinate points in the feature space, each uniquely representing a known modulation class. For example, for five known classes, five orthogonal one-hot vectors can be pre-set as initial anchor points, such as [1,0,0,0,0] for BPSK, [0,1,0,0,0] for QPSK, and so on.

[0035] The training objective is to use a combined loss function To achieve this, the function is preferably composed of two parts: contrast distance loss and adaptive clustering loss, namely .

[0036] Comparative distance loss : Its purpose is to make the feature space have basic distinguishability. For the training data within a batch, the loss will calculate and punish those situations where "the distance between samples of the same type is too far" and "the distance between samples of different types is too close". A specific implementation method is to find, for each sample, the sample of the same type but the farthest distance (the most difficult positive sample) and the sample of the different type but the closest distance (the most difficult negative sample), and ensure that the distance between different types is significantly greater than the distance between the same type. By minimizing this loss, the network is driven to push the feature vectors of different modulation categories away from each other in space, while pulling the feature vectors of the same category together to form a preliminary cluster structure. Its function expression is: ; in, Represents the feature that is farthest from the same type (positive sample) of the anchor point (i.e., the "hardest positive sample"); Represents the feature that is closest to the anchor point in a different class (negative sample) (i.e., the "hardest negative sample"); Indicates the number of valid anchor samples in the batch; Represents the high-order semantic feature vector of the i-th anchor point sample after being processed by the deep neural network; Indicates the modulation category label corresponding to the i-th anchor point sample; Representation and anchor Samples that do not belong to the same category; α is a preset hyperparameter, a positive constant, used to set a safety margin.

[0037] The goal of this loss function is to make the distance from the hardest negative sample to the anchor point at least α larger than the distance from the hardest positive sample to the anchor point. If this condition is met, the loss of the anchor point is 0.

[0038] Adaptive clustering loss : Its purpose is to find a stable and fixed affiliation for these initially formed clusters. The core of this loss is to minimize the distance between the feature vector of each sample and the cluster anchor point corresponding to its category. In some preferred embodiments, a preferred adaptive clustering loss calculation method is given, which specifically includes: S21. For each known modulation category, calculate the center point of all high-order semantic feature vectors it contains.

[0039] S22. Repeatedly find the closest pair of all currently unmatched center points and unmatched cluster anchor points, and pair them until all center points are matched.

[0040] S23. Determine the cluster anchor point corresponding to each high-order semantic feature vector in the training batch based on the pairing relationship.

[0041] S24. Calculate the mean distance between all high-order semantic feature vectors and their corresponding cluster anchor points as the adaptive clustering loss.

[0042] The preferred function expression is as follows: ; Where N represents the total number of known modulation types; M represents the number of independent modulation types contained in the batch of data currently being processed; Represents the sample feature vectors of all samples belonging to the jth modulation category in this batch of data A collection of is a function whose output is related to the category The only cluster anchor that matches; Represents the sample feature vector The corresponding cluster anchor point the distance between them; Representation category The center point of all feature vectors belonging to the jth class in the batch The average vector of ; A represents the set of all preset cluster anchor points.

[0043] Through back propagation, according to this combined loss function The weight parameters of the deep neural network are trained and optimized using the values ​​of . This process is repeated until the loss converges below a preset threshold, marking the completion of model training. After training, the resulting model has a structured feature space in which each known class is tightly clustered around its unique cluster anchor.

[0044] After training is complete, the locations of cluster anchors can be further optimized. A preferred method for obtaining anchors is to input all training samples of known categories in the dataset into the trained network to obtain their respective high-order semantic feature vectors. Then, a voting statistics is performed on all feature vectors for each category (for example, by calculating the mean or voting based on the result of binarization of the feature vectors) to determine the final cluster anchor that best represents the characteristics of the category group.

[0045] The second stage: open set recognition.

[0046] Once the model training is completed, it can be used to perform open set recognition on new signals with unknown labels.

[0047] S3. Obtain data in at least two different representation forms of the signal to be identified, and use the deep neural network to extract its high-order semantic feature vector.

[0048] For a signal to be recognized, the data acquisition and processing process is identical to that in step S1. This involves acquiring multimodal data, including IQ, frequency domain, and phase, and performing normalization and slicing to ensure data format consistency. The processed data is then fed into a trained deep neural network to extract its high-order semantic feature vectors.

[0049] S4. Based on the positional relationship between the high-order semantic feature vector of the signal to be identified and the cluster anchor point in the feature space, determine whether the category of the signal to be identified is one of the known modulation categories, or determine whether it is an unknown modulation category.

[0050] This step is the core of recognition. It is based on the positional relationship between the high-order semantic feature vector of the signal to be recognized obtained in the previous step and the cluster anchor points of each known category in the feature space to ultimately determine the signal category.

[0051] To ensure clearer and more stable decisions, a preferred approach is to binarize the extracted high-order semantic feature vectors before determining positional relationships. A specific binarization method involves setting a preset threshold (for example, 0.5) and setting feature vector values ​​greater than or equal to the threshold to 1, and values ​​less than the threshold to 0. This process transforms the network's soft output into a hard decision vector that is identical to the one-hot cluster anchor format.

[0052] Then, a decision is made. A preferred decision method is to use a voting mechanism. Since a signal file to be identified is usually divided into multiple samples, we will obtain multiple high-order semantic feature vectors. For each feature vector, calculate which cluster anchor point has the closest "positional relationship" with it (for example, after binarization, which anchor point completely matches it). Then, the decision results of all samples are counted: If the proportion of samples corresponding to the same cluster anchor point exceeds a preset voting threshold (for example, 50%), it is considered that most of the evidence of the signal points to a known class, and the modulation category of the signal to be identified is determined to be the known modulation category corresponding to the cluster anchor point.

[0053] On the contrary, if the vote rate of no known category exceeds the threshold, it means that the characteristics of the signal are diffuse in space and do not clearly belong to the "safe zone" of any known category, then it is determined to be an unknown modulation category.

[0054] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A signal open set modulation recognition method based on deep clustering, characterized in that: Including steps: S1. Obtain data of at least two different representations of a signal of known modulation category and process them through a deep neural network to generate a high-order semantic feature vector; S2. Based on the high-order semantic feature vectors and the cluster anchor points representing each known modulation category, training the deep neural network to increase the distance between different known modulation categories and reduce the distance within the same known modulation category in the feature space, and make the high-order semantic feature vectors of each known modulation category converge to their corresponding cluster anchor points; S3. Obtaining data of at least two different representations of the signal to be identified, and extracting its high-order semantic feature vector using the deep neural network; S4. Based on the positional relationship between the high-order semantic feature vector of the signal to be identified and the cluster anchor point in the feature space, determine whether the category of the signal to be identified is one of the known modulation categories, or determine whether it is an unknown modulation category.

2. The signal open set modulation identification method based on deep clustering according to claim 1, characterized in that: The two data in different representation forms include: IQ data of the signal, frequency domain data calculated based on the IQ data, and phase data.

3. The signal open set modulation identification method based on deep clustering according to claim 1, characterized in that: The method for training the deep neural network includes: training the deep neural network by combining a loss function; the combined loss function includes: Contrastive distance loss, which is used to reduce the feature distance between similar signal samples and increase the feature distance between heterogeneous signal samples during training; An adaptive clustering loss is used to minimize the distance between the high-order semantic feature vector and the cluster anchor point corresponding to the known modulation category during training.

4. The signal open set modulation identification method based on deep clustering according to claim 3, characterized in that: The calculation method of the adaptive clustering loss includes: For each known modulation category, calculate the center point of all high-order semantic feature vectors it contains; Repeatedly find the closest pair of all currently unmatched center points and unmatched cluster anchor points, and pair them until all center points are matched; Determining, based on the pairing relationship, a cluster anchor point corresponding to each high-order semantic feature vector in the training batch; The mean distance between all high-order semantic feature vectors and their corresponding cluster anchor points is calculated as the adaptive clustering loss.

5. The signal open set modulation identification method based on deep clustering according to claim 1, characterized in that: The method for obtaining cluster anchor points includes: performing voting statistics on high-order semantic feature vectors of all training samples of each known modulation category to determine the cluster anchor points of the category.

6. The signal open set modulation identification method based on deep clustering according to claim 1, characterized in that: The method for determining whether the type of the signal to be identified is one of the known modulation types includes: If the proportion of high-order semantic feature vectors corresponding to the same cluster anchor point in multiple samples included in the signal to be identified exceeds a preset voting threshold, the modulation category of the signal to be identified is determined to be the known modulation category corresponding to the cluster anchor point.

7. The signal open set modulation identification method based on deep clustering according to claim 1, characterized in that: Before determining the positional relationship, step S4 further includes performing binarization post-processing on the high-order semantic feature vector of the signal to be identified.

8. The signal open set modulation identification method based on deep clustering according to claim 7, characterized in that: The binarization post-processing is: comparing the numerical value in the high-order semantic feature vector with a preset threshold, setting the value to 0 if it is less than the threshold, and setting it to 1 otherwise.

Images