Deep clustering based signal out-of-set modulation identification method
By training a deep neural network using deep clustering methods, generating high-order semantic feature vectors, and using clustering anchors to distinguish between known and unknown signals, the problem of closed-set models being unable to identify unknown signals is solved, and efficient open-set modulation recognition is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SICHUAN JIUTIAN EMBODIED INTELLIGENT TECHNOLOGY CO LTD
- Filing Date
- 2025-06-13
- Publication Date
- 2026-05-19
AI Technical Summary
Existing closed-set modulation recognition models cannot effectively recognize unknown modulation signals, resulting in severe distortion of recognition results. Furthermore, existing open-set recognition methods have complex system structures and unclear feature boundaries, affecting recognition accuracy and robustness.
A signal open set modulation recognition method based on deep clustering is adopted. High-order semantic feature vectors are generated through deep neural networks, and the model is trained by merging loss function and adaptive clustering loss to form compact clusters of known modulation categories in feature space. Clustering anchors are used to distinguish unknown signals.
It achieves accurate identification of known modulation categories and effective detection of unknown categories in a single model, simplifies the system architecture, improves recognition performance and environmental adaptability, and reduces complexity.
Smart Images
Figure REF-OBJ-1774852663247-000002 
Figure REF-OBJ-1774852663247-000003 
Figure REF-OBJ-1774852663247-000004
Abstract
Description
Technical Field
[0001] This 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 Technology
[0002] Automatic Modulation Classification (AMC) is a key technology in wireless communication, spectrum monitoring, and electronic countermeasures. Its core task is to automatically determine the modulation scheme of the received signal. Traditional modulation identification methods mainly rely on expert experience and signal processing techniques, such as maximum likelihood estimation, cyclic spectrum feature extraction, and instantaneous statistical feature-based methods. These methods can achieve good results under ideal channel conditions and high signal-to-noise ratios. However, in real-world complex electromagnetic environments, signals are often severely interfered with by noise, multipath propagation, and channel fading, leading to a sharp decline in the performance of traditional methods. Furthermore, their reliance on manually designed features results in weak generalization ability, making it difficult to cope with the increasingly diverse modulation types.
[0003] In recent years, artificial intelligence technologies, represented by deep learning, have demonstrated enormous potential in the field of signal modulation recognition due to their powerful adaptive feature extraction and nonlinear mapping capabilities. Models based on deep neural networks can directly learn deeper and more robust features from raw signals, achieving significant success in "closed-set" recognition tasks. "Closed-set" recognition refers to the model's task of classifying all encountered signals into a predefined, finite set of known modulation types.
[0004] However, closed-set recognition models have a fundamental flaw: they assume that all signals to be tested belong to a known set of categories. In practical applications, recognition systems inevitably encounter novel or unknown modulation signals that were never seen during the training phase. In such cases, the closed-set model will incorrectly force these unknown signals to be classified into a known category, leading to severely distorted recognition results and rendering the system unreliable.
[0005] Therefore, how to enable modulation recognition systems to accurately identify known modulation types while effectively detecting and distinguishing unknown modulation types not found in the training set—that is, achieving "open-set recognition" (OSR)—is a key challenge and technical bottleneck in the field. Some existing attempts at open-set recognition, such as the Chinese invention patent with publication number CN113052126B, employ a multi-stage processing approach: first, a classifier or thresholding method is used to identify known classes and filter out suspected unknown signals; then, an independent network (such as a Siamese neural network) is used to perform secondary clustering or analysis on these filtered unknown signals. While such methods achieve the function of open-set recognition to some extent, they often suffer from complex system structures, fragmented training and recognition processes, and difficulties in coordinating and optimizing multiple models or stages. More importantly, they may not be clear and robust enough in constructing the feature space boundary between known and unknown classes, thus affecting the detection rate and recognition accuracy of unknown signals under complex conditions. Summary of the Invention
[0006] To address the problems existing in the prior art, the present invention provides the following technical solution, including:
[0007] The signal open set modulation recognition method based on deep clustering includes the following steps:
[0008] S1. Obtain data in at least two different representations of a known modulation category signal, and process them through a deep neural network to generate a high-order semantic feature vector;
[0009] S2. Based on the higher-order semantic feature vector and the clustering anchor points representing each known modulation category, train the deep neural network to increase the distance between different known modulation categories and decrease the distance within the same known modulation category in the feature space, and make the higher-order semantic feature vector of each known modulation category converge to its corresponding clustering anchor point;
[0010] S3. Acquire data in at least two different representations of the signal to be identified, and extract its high-order semantic feature vector using the deep neural network;
[0011] S4. Based on the positional relationship between the high-order semantic feature vector of the signal to be identified and the clustering anchor point in the feature space, determine that the category of the signal to be identified is one of the known modulation categories, or determine that it is an unknown modulation category.
[0012] Preferably, the two different representations of data include: IQ data of the signal, frequency domain data calculated based on the IQ data, and phase data.
[0013] Preferably, the method for training the deep neural network includes: training the deep neural network using a merging loss function; the merging loss function includes:
[0014] Contrast distance loss is used to reduce the feature distance of similar signal samples and increase the feature distance of dissimilar signal samples during training.
[0015] An adaptive clustering loss is used during training to minimize the distance between the higher-order semantic feature vector and the cluster anchor point corresponding to the known modulation category.
[0016] Preferably, the method for calculating the adaptive clustering loss includes:
[0017] For each known modulation category, calculate the center point of all its higher-order semantic feature vectors;
[0018] Repeatedly find the closest pair between all currently unmatched centroids and unmatched cluster anchors, and pair them up until all centroids are matched;
[0019] Based on the pairing relationship, determine the clustering anchor point corresponding to each higher-order semantic feature vector in the training batch;
[0020] The mean distance between all higher-order semantic feature vectors and their corresponding clustering anchors is calculated and used as the adaptive clustering loss.
[0021] Preferably, the method for obtaining the clustering anchor points includes: determining the clustering anchor points for each known modulation category by performing voting statistics on the high-order semantic feature vectors of all training samples.
[0022] Preferably, the method for determining that the category of the signal to be identified is one of the known modulation categories includes:
[0023] If the proportion of samples whose high-order semantic feature vectors correspond to the same cluster anchor point in the signal to be identified exceeds a preset voting threshold, then the modulation category of the signal to be identified is determined to be the known modulation category corresponding to the cluster anchor point.
[0024] Preferably, before determining the positional relationship, step S4 further includes binarizing the high-order semantic feature vector of the signal to be identified.
[0025] Preferably, the binarization post-processing is as follows: the values in the higher-order semantic feature vector are compared with a preset threshold, and those less than the threshold are set to 0, otherwise they are set to 1.
[0026] Beneficial effects
[0027] 1. This invention achieves both accurate identification of known modulation categories and effective detection of unknown categories by designing 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.
[0028] 2. This invention employs a unique dual-objective training method. On one hand, it increases the feature distance between different known categories; on the other hand, it utilizes "clustering anchors" to make the feature distribution of each known class more compact and stable. This training method constructs a feature space with clear boundaries, enabling the feature vector of an unknown signal to be effectively identified due to its significant deviation from the clustering anchors of all known classes, thereby greatly improving the performance of open set recognition.
[0029] 3. Multimodal feature fusion can capture the inherent patterns of signals from different dimensions, forming a more comprehensive and richer feature representation than single-modal fusion. It can better overcome the interference of complex factors such as noise and channel fading in practical applications, further enhancing the overall performance and environmental adaptability of the method. Attached Figure Description
[0030] Figure 1 This is a flowchart illustrating a signal open set modulation recognition method based on deep clustering provided in a preferred embodiment of the present invention. Detailed Implementation
[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described below with reference to the accompanying drawings. In the description of this invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention.
[0032] In modern wireless communication, accurately identifying the modulation scheme of a signal is crucial for ensuring communication quality, spectrum monitoring, and signal reconnaissance. Traditional identification methods often suffer from limitations in performance and generalization ability when faced with complex real-world electromagnetic environments. While deep learning has achieved success in "closed-set" modulation identification (classification from a pre-defined set of known categories) in recent years, it cannot handle the "open-set" problem. When encountering unknown modulation types not seen during training, closed-set models can produce severe misclassifications, which is unacceptable in practical applications. For example, a closed-set system trained only on BPSK and QPSK signals might incorrectly identify a 16QAM signal as either BPSK or QPSK, leading to complete failure of subsequent demodulation and information acquisition. Existing attempts at open-set identification still suffer from limitations in performance and robustness, either due to complex system structures requiring multi-stage processing or unclear boundaries between known and unknown classes.
[0033] To address the aforementioned technical challenges, this invention proposes a signal open-set modulation recognition method based on deep clustering. The core idea of this method is to construct a structured feature space through a special end-to-end training framework, allowing known modulation categories to form highly cohesive and fixed-position clusters within the space. This enables any unknown signal not belonging to these clusters to be easily distinguished.
[0034] The technical solution of the present invention will be 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 for explaining the present invention and are not intended to limit its scope of protection in any way.
[0035] Example 1
[0036] like Figure 1 As shown in the figure, this embodiment will demonstrate in detail the complete process of a signal open set modulation recognition method based on deep clustering. The process mainly includes two stages: model training and open set recognition.
[0037] Phase 1: Model training.
[0038] 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.
[0039] S1. Obtain data in at least two different representations of a known modulation category signal and process them through a deep neural network to generate a high-order semantic feature vector.
[0040] First, a dataset containing signals of various known modulation categories needs to be prepared for training. In this embodiment, it is assumed that the set of known modulation categories includes five types: BPSK, QPSK, 8PSK, 16QAM, and 64QAM. To enable the neural network to understand the signals from a richer perspective, this invention preferably employs a multimodal input strategy. Specifically, for each known modulation category signal, the following operations are performed to obtain data in at least two different representations:
[0041] 1. Obtain raw IQ data: Directly acquire the time series data of the in-phase (I) and quadrature (Q) components of the signal.
[0042] 2. Calculate frequency domain data: Perform a Fast Fourier Transform (FFT) on the IQ data to obtain the spectral representation of the signal, i.e., frequency domain data.
[0043] 3. Calculate phase data: Perform phase calculation on the IQ data to obtain the phase information sequence of the signal.
[0044] In this way, the original one-dimensional time-domain signal is transformed 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 representational power of information, because frequency domain and phase information can reveal implicit features in the time-domain signal that are difficult to observe directly.
[0045] Before being fed into the network, these multimodal data typically need to be preprocessed. For example, they can be normalized to the maximum and minimum values to scale all data to a uniform scale, and then sliced into fixed-length samples to form standardized training samples. For instance, each signal file can be divided into several sample data groups of length 8192, each containing three parts: IQ, frequency domain, and phase.
[0046] Then, these processed multimodal sample data are fed into a deep neural network. This network is preferably a multimodal convolutional neural network with parallel branches designed internally to process IQ, frequency domain, and phase data separately, and to effectively fuse these different modal features at a deeper level. After forward propagation, the network generates a high-order semantic feature vector for each input training sample. This vector is a compact and abstract representation of the original signal in the deep feature space.
[0047] S2. Based on the higher-order semantic feature vectors and the clustering anchors representing each known modulation category, train the deep neural network to increase the distance between different known modulation categories and decrease the distance within the same known modulation category in the feature space, and make the higher-order semantic feature vectors of each known modulation category converge to their corresponding clustering anchors.
[0048] This step is the core of this invention. Instead of using traditional classification loss (such as cross-entropy), the network is trained through a specially designed optimization process with dual objectives 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.
[0049] The clustering anchor points are preset coordinate points in the feature space, and each anchor point uniquely represents a known modulation category. For example, for 5 known categories, 5 orthogonal one-hot vectors can be preset as initial anchor points, such as [1,0,0,0,0] representing BPSK, [0,1,0,0,0] representing QPSK, and so on.
[0050] The training objective is achieved through a pooling loss function. To achieve this, the function preferably consists of two parts: a contrastive distance loss and an adaptive clustering loss. .
[0051] Comparison of distance loss The goal is to give the feature space basic discriminative power. For training data within a batch, this loss calculates and penalizes cases where "samples of the same class are too far apart" and "samples of different classes are too close together." A specific implementation involves finding, for each sample, the sample of the same class but furthest distance (the hardest positive sample) and the sample of the different class but closest distance (the hardest negative sample), ensuring that the distance between different classes is significantly greater than the distance between the same class. By minimizing this loss, the network is driven to push feature vectors of different modulation categories apart in space, while simultaneously bringing feature vectors of the same class together, forming a preliminary cluster structure. Its functional expression is:
[0052] ;
[0053] in, This represents the feature that is furthest away from the anchor point (positive sample) (i.e., the "most difficult positive sample"). This represents the feature that is closest to the anchor point (a negative sample) (i.e., the "hardest negative sample"). Indicates the number of valid anchor point samples in the batch; This represents the high-order semantic feature vector obtained by processing the i-th anchor point sample through a deep neural network. This represents the modulation category label corresponding to the i-th anchor sample; Indicates anchor point Samples that do not belong to the same category; α is a preset hyperparameter, which is a positive number and is used to set a safety boundary.
[0054] The goal of this loss function is to ensure that the distance from the hardest negative sample to the anchor point is at least α greater than the distance from the hardest positive sample to the anchor point. If this condition is met, then the loss at that anchor point is 0.
[0055] Adaptive clustering loss The aim 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, specifically including:
[0056] S21. For each known modulation category, calculate the center point of all high-order semantic feature vectors contained therein.
[0057] S22. Repeatedly find the closest pair between all currently unmatched centroids and unmatched cluster anchors, and pair them up until all centroids are matched.
[0058] S23. Based on the pairing relationship, determine the clustering anchor point corresponding to each high-order semantic feature vector in the training batch.
[0059] S24. Calculate the mean distance between all higher-order semantic feature vectors and their corresponding clustering anchors, as the adaptive clustering loss.
[0060] Its preferred function expression is as follows:
[0061] ;
[0062] Where N represents the total number of known modulation categories; M represents the number of independent modulation categories contained in the current batch of data being processed; This represents the feature vector of all samples belonging to the j-th modulation category in this batch of data. A set; For functions, the output result is related to the category. A unique clustering anchor that matches the target cluster. Represents the sample feature vector Its corresponding clustering anchor point The distance between them; Indicate category The center point, i.e., all feature vectors belonging to class j in this batch. The average vector; A represents the set of all pre-defined clustering anchor points.
[0063] Through backpropagation, based on this merging loss function The values are used to train and optimize the weight parameters of the deep neural network. This process is repeated until the loss converges to below a preset threshold, marking the completion of model training. After training, the resulting model possesses a structured feature space, where each known class is tightly clustered around its unique clustering anchor point.
[0064] After training, the location of the clustering anchor points can be further optimized. A preferred method for obtaining anchor points 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 are performed on all feature vectors of each category (e.g., by calculating the mean or by voting based on the binarized result of the feature vectors) to determine a final clustering anchor point that better represents the characteristics of that category.
[0065] Phase Two: Open Set Recognition.
[0066] Once the model is trained, it can be used to perform open set recognition on new signals with unknown labels.
[0067] S3. Obtain data in at least two different representations of the signal to be identified, and extract its high-order semantic feature vector using the deep neural network.
[0068] For a signal to be identified, firstly, the same data acquisition and processing procedure as step S1 is executed, namely, acquiring its IQ, frequency domain, and phase multimodal data, and performing normalization and slicing to ensure data format consistency. Then, the processed data is fed into a pre-trained deep neural network to extract its high-order semantic feature vector.
[0069] S4. Based on the positional relationship between the high-order semantic feature vector of the signal to be identified and the clustering anchor point in the feature space, determine that the category of the signal to be identified is one of the known modulation categories, or determine that it is an unknown modulation category.
[0070] This step is the core of the identification process. It determines the signal's category based on the positional relationship between the high-order semantic feature vector of the signal to be identified obtained in the previous step and the clustering anchor points of each known category in the feature space.
[0071] To make the judgments more explicit and stable, 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 (e.g., 0.5), setting values in the feature vector greater than or equal to this threshold to 1, and values less than the threshold to 0. After this step, the network's soft output becomes a hard decision vector that is completely consistent with the one-hot clustering anchor point format.
[0072] Next, a judgment is made. A preferred judgment method is to use a voting mechanism. Since a signal file to be identified is usually segmented into multiple samples, we obtain multiple high-order semantic feature vectors. For each feature vector, we calculate which cluster anchor point it is closest to (e.g., which anchor point it perfectly matches after binarization). Then, we summarize the judgment results for all samples:
[0073] If the proportion of samples corresponding to the same cluster anchor point exceeds a preset voting threshold (e.g., 50%), then it is considered that most of the evidence for 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.
[0074] Conversely, if no known category receives more votes than the threshold, it indicates that the signal's characteristics are diffuse in space and do not clearly belong to the "safe zone" of any known category, thus classifying it as an unknown modulation category.
[0075] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this 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 the following steps: S1. Obtain data in at least two different representations of a known modulation category signal, and process them through a deep neural network to generate a high-order semantic feature vector; S2. Based on the higher-order semantic feature vector and the clustering anchor points representing each known modulation category, train the deep neural network to increase the distance between different known modulation categories and decrease the distance within the same known modulation category in the feature space, and make the higher-order semantic feature vector of each known modulation category converge to its corresponding clustering anchor point; S3. Acquire data in at least two different representations of the signal to be identified, and extract 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 clustering anchor point in the feature space, determine that the category of the signal to be identified is one of the known modulation categories, or determine that it is an unknown modulation category; The method for training the deep neural network includes: training the deep neural network using a merged loss function; the merged loss function includes: Contrast distance loss is used to reduce the feature distance of similar signal samples and increase the feature distance of dissimilar signal samples during training. An adaptive clustering loss is used to minimize the distance between the higher-order semantic feature vector and the clustering anchor point of its corresponding known modulation category during training. The functional expression for the contrast distance loss is: ； in, This represents the feature that is furthest away from the anchor point among those of the same type. This represents the feature that is closest to the anchor point in a class other than the anchor point. Indicates the number of valid anchor point samples in the batch; This represents the high-order semantic feature vector obtained by processing the i-th anchor point sample through a deep neural network. This represents the modulation category label corresponding to the i-th anchor sample; Indicates anchor point Samples that do not belong to the same category; α is a preset hyperparameter.
2. The signal open set modulation recognition method based on deep clustering as described in claim 1, characterized in that, The two different representations of the data include: IQ data of the signal, frequency domain data calculated based on the IQ data, and phase data.
3. The signal open set modulation recognition method based on deep clustering as described in claim 1, characterized in that, The method for calculating the adaptive clustering loss includes: For each known modulation category, calculate the center point of all its higher-order semantic feature vectors; Repeatedly find the closest pair between all currently unmatched centroids and unmatched cluster anchors, and pair them up until all centroids are matched; Based on the pairing relationship, determine the clustering anchor point corresponding to each higher-order semantic feature vector in the training batch; The mean distance between all higher-order semantic feature vectors and their corresponding clustering anchors is calculated and used as the adaptive clustering loss.
4. The signal open set modulation recognition method based on deep clustering as described in claim 1, characterized in that, The method for obtaining the clustering anchor points includes: determining the clustering anchor points for each known modulation category by performing voting statistics on the high-order semantic feature vectors of all training samples.
5. The signal open set modulation recognition method based on deep clustering as described in claim 1, characterized in that, The method for determining that the category of the signal to be identified is one of the known modulation categories includes: If the proportion of samples whose high-order semantic feature vectors correspond to the same cluster anchor point in the signal to be identified exceeds a preset voting threshold, then the modulation category of the signal to be identified is determined to be the known modulation category corresponding to the cluster anchor point.
6. The signal open set modulation recognition method based on deep clustering as described in claim 1, characterized in that, Before determining the positional relationship, step S4 also includes binarizing the high-order semantic feature vector of the signal to be identified.
7. The signal open set modulation recognition method based on deep clustering as described in claim 6, characterized in that, The binarization post-processing is as follows: the values in the higher-order semantic feature vector are compared with a preset threshold; values less than the threshold are set to 0, otherwise they are set to 1.