A multi-component communication signal identification method and device, electronic equipment and computer readable storage medium

Abstract

This invention discloses a method, apparatus, electronic device, and computer-readable storage medium for identifying multi-component communication signals. The method includes the following steps: data preprocessing: performing time-frequency transformation on the received communication signal to generate a time-frequency image of a preset size, and normalizing the time-frequency image to obtain time-frequency image features; signal quantity detection: inputting the time-frequency image features into a pre-trained signal quantity detection module and outputting the signal component quantity result; targeted identification: based on the signal component quantity result, routing the time-frequency image features to a corresponding pre-trained identification model, and after feature enhancement processing and category difference amplification processing, outputting the modulation type identification result of the communication signal. This invention achieves efficient identification of multi-component communication signals in complex environments through the logic of "first detecting the number of components → then targeted identification".

Description

Technical Field

[0001] This invention relates to the field of communication signal identification, and more particularly to a method, apparatus, electronic device, and computer-readable storage medium for identifying multi-component communication signals. Background Technology

[0002] Modern electronic reconnaissance, spectrum monitoring, and electronic countermeasures place extremely high demands on communication signal processing capabilities in complex electromagnetic environments. Communication signal recognition, as a core component of these fields, directly determines the efficiency and practical value of electronic reconnaissance systems through its accuracy, adaptability to low signal-to-noise ratios (SNR), and multi-component signal processing capabilities. Especially in low signal-to-noise ratio (SNR) environments, the electromagnetic space is filled with significant Gaussian white noise interference, and received communication signals often exist in single-component, dual-component, or even multi-component mixed forms, with the number of components exhibiting randomness and uncertainty. How to achieve accurate modulation and recognition of multi-component communication signals in such scenarios has become a critical technical problem urgently needing to be solved in this field.

[0003] In recent years, deep learning-based communication signal recognition methods have been widely researched and applied, becoming the mainstream technical solution to replace traditional manual feature extraction and recognition methods. In existing technologies, researchers mostly use deep learning models such as Convolutional Neural Networks (CNNs), Residual Networks (ResNets), and attention mechanisms to generate time-frequency images by performing time-frequency transformations on communication signals, extracting time-frequency domain features to achieve modulation recognition of single-component communication signals. For the need to recognize multi-component signals, some technical solutions focus on dual-component communication signal scenarios, employing multi-instance multi-label learning, multi-scale feature extraction, and a combination of convolutional neural networks and reinforcement learning to complete feature extraction and classification of overlapping dual-component signals. Meanwhile, other technologies improve the structure of convolutional neural networks by introducing modules such as moving inverted bottleneck convolutional blocks and channel-space attention mechanisms to achieve modulation recognition of dual-component and three-component radar signals, thus improving the recognition accuracy of multi-component signals to some extent.

[0004] In addition, in terms of signal preprocessing and noise suppression, existing technologies mostly use two-dimensional Wiener filtering, fixed threshold denoising and other methods to suppress time-frequency image noise, or optimize the time-frequency feature expression of the signal by improving time-frequency distribution algorithms (such as Cohen-type time-frequency distribution, smooth pseudo-Wigner-Ville distribution), in an attempt to find a balance between noise suppression and feature preservation in order to adapt to the signal recognition needs under low signal-to-noise ratio.

[0005] However, existing deep learning-based multi-component signal recognition technologies, whether for communication signals or radar signals, still suffer from numerous insurmountable technical defects in the application of multi-component communication signal recognition in complex electromagnetic environments. These defects result in low recognition accuracy and insufficient robustness, failing to meet the practical engineering needs of electronic reconnaissance and spectrum monitoring. The specific problems are manifested in the following three aspects: Poor adaptability to the number of signal components makes adaptive identification of signals with unknown component counts impossible. Existing technologies all presuppose that the number of signal components is a known fixed value (such as single component, dual component, or triple component), and do not design a dedicated signal component count detection module. They directly perform multi-label classification on the signal. However, the communication signals received in the actual electromagnetic environment may be noise, single component, or multi-component (dual component, triple component, quad component, etc.), and the number of components is random. This assumption leads to a significant decrease in the accuracy of existing technologies in identifying mixed multi-component communication signals with unknown component counts, and may even result in identification failure.

[0006] Feature extraction is ineffective at low signal-to-noise ratios (SNR), and balancing noise suppression and feature preservation is difficult. Existing noise suppression methods often employ fixed filtering rules or fixed thresholds, lacking the ability to dynamically adapt to noise intensity. In extremely low SNR environments (SNR≤-10dB), either excessive noise suppression leads to the loss of effective signal features, or Gaussian white noise interference is not effectively filtered out, causing noise features to be confused with signal features. This makes it difficult for subsequent deep learning models to extract effective signal features, ultimately affecting the recognition results.

[0007] The features of multi-component signals are severely interfered with, the class feature discrimination is insufficient, and the recognition robustness is poor. Existing feature extraction modules mostly use single-scale feature extraction or simple attention mechanisms, lacking targeted multi-component feature discrimination and purification mechanisms. The time-frequency features of different component signals are prone to mutual interference. At the same time, existing models rely solely on positive label supervised training, without strengthening the feature differences between signals of different modulation types. This leads to blurred boundaries between different categories of signals in the feature space, making it easy for class confusion to occur. When signal parameters deviate from the training range, the recognition performance drops sharply and cannot adapt to the dynamic changes of communication signal parameters in real-world scenarios.

[0008] Analysis reveals three core reasons for the aforementioned problems: First, existing technologies lack a detection module designed to address the randomness of the number of communication signal components, making it impossible to adapt signal processing and recognition strategies to the actual number of components, thus fundamentally limiting the ability to recognize signals with unknown component counts. Second, during feature extraction and noise suppression, the denoising threshold is not dynamically adjusted based on signal noise characteristics, nor is deep feature purification achieved through multi-scale attention fusion and adaptive feature fusion, resulting in limited feature purification capabilities for low signal-to-noise ratio communication signals. Third, the model training process employs only a single supervised mode with positive labels, without introducing a training mechanism that actively amplifies differences in class features. This leads to insignificant differences in features learned by the model for different modulation types of signals, blurring class boundaries in the feature space, ultimately causing feature interference and class confusion among multiple component signals.

[0009] Therefore, developing a multi-component communication signal recognition method that can adaptively identify the number of unknown components, adapt to extremely low signal-to-noise ratio environments, and enhance the distinguishability of category features, in order to solve the above-mentioned defects of existing technologies and improve the accuracy and robustness of multi-component communication signal recognition in complex electromagnetic environments, has become an urgent need in this field.

[0010] The disclosure of the above background technical content is only for the purpose of assisting in understanding the concept and technical solution of this application, and does not necessarily provide technical instruction. Summary of the Invention

[0011] The purpose of this invention is to provide a method, apparatus, electronic device, and computer-readable storage medium for identifying multi-component communication signals. By employing the logic of "first detecting the number of components → then identifying them specifically", it enables efficient identification of multi-component communication signals in complex environments.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for identifying multi-component communication signals includes the following steps: Step 1, Data Preprocessing: Perform time-frequency transformation on the received communication signal to generate a time-frequency image of a preset size, and normalize the time-frequency image to obtain time-frequency image features; Step 2, Signal Quantity Detection: The time-frequency image features are input into a pre-trained signal quantity detection module, which outputs the signal component quantity result. The signal component quantity result includes noise, single-component signals, and multi-component signals. The signal quantity detection module consists of multiple convolutional neural network sub-models and an ensemble learning unit. Each convolutional neural network sub-model is trained using focus loss as the loss function. The ensemble learning unit performs weighted fusion of the prediction results of each convolutional neural network sub-model and determines the signal component quantity based on the fusion result. Step 3, Targeted Identification: Based on the signal component count result, the time-frequency image features are routed to the corresponding pre-trained recognition model. After feature enhancement processing and category difference amplification processing, the modulation type identification result of the communication signal is output. If the signal component count result is noise, the noise identification result is directly output. If it is a single-component signal, the time-frequency image features are input into the single-component recognition model. If it is a multi-component signal, the time-frequency image features are input into the multi-component recognition model. The single-component recognition model and the multi-component recognition model are provided with feature enhancement blocks and inverse difference blocks. The feature enhancement blocks perform denoising and purification operations and multi-scale feature fusion operations on time-frequency image features. The inverse difference blocks are used to amplify the category feature differences of communication signals with different modulation types.

[0013] Furthermore, following any one or a combination of the aforementioned technical solutions, the time-frequency transformation in step 1 is implemented using a Choi-Williams distribution.

[0014] Furthermore, based on any one or a combination of the aforementioned technical solutions, the formula for calculating the focus loss is as follows:

[0015] Among them, the category weight coefficient =[0.3, 0.25, 0.2, 0.15, 0.1] (corresponding to components 0 to 4 respectively), focusing parameter r=2, Let be the predicted probability of each convolutional neural network sub-model for the number of the c-th class component. K The maximum number of categories in terms of component quantity; The weighted fusion formula for the integrated learning unit is: ; in, Predict the probability of the number of the c-th component after fusion. Let be the weight of the m-th convolutional neural network sub-model, and the sum of the weights of all convolutional neural network sub-models is 1. Let be the predicted probability of the m-th convolutional neural network sub-model for the number of c-th class components; The integrated learning unit, through Determine the final number of signal components.

[0016] Furthermore, following any one or a combination of the aforementioned technical solutions, the feature enhancement block includes a denoising sub-block and an attention sub-block, and the specific processing procedure is as follows: The denoising sub-block is constructed based on an autoencoder framework and learns the noise threshold dynamically through global average pooling. τ = GAP (X rec ),in X rec For the reconstruction features of the autoencoder, effective features in the time-frequency image features are selected through preset rules; The attention sub-block employs a multi-scale channel attention module and an adaptive feature fusion strategy to extract multi-scale features of time-frequency image features and complete adaptive fusion. The fused features are added to the denoised features via residual connections to output enhanced features. As a feature of fusion, , X These are the features after denoising.

[0017] Furthermore, following any one or a combination of the aforementioned technical solutions, the preset rule is as follows: Features in the time-frequency image with pixel values ​​greater than the noise threshold τ are retained, while the remaining features are set to 0. The formula is:

[0018] in, These are the feature pixel values ​​after denoising. These are the pixel values ​​of the original time-frequency image features. The coordinates of the feature pixel.

[0019] Furthermore, following any one or a combination of the aforementioned technical solutions, the calculation formula for the adaptive feature fusion is as follows: ,in, A For adaptive learning fusion weights, X m , X n Features of different scales extracted by the multi-scale channel attention module.

[0020] Furthermore, following any one or a combination of the aforementioned technical solutions, the processing procedure for the reverse difference block is as follows: Forward training: Based on the original labels of communication signals y The feature extraction network is trained using cross-entropy loss to obtain positive feature output. The cross-entropy loss formula is as follows: ,in c The number of modulation types for communication signals. One-hot encoding for the original label, Predict the probability for the positive label; Reverse training: based on reverse labels The feature extraction network is trained using cross-entropy loss to obtain inverse feature output. The cross-entropy loss formula is as follows: ,in Predict probabilities for inverse labels; Feature difference calculation: Calculate the feature difference between forward training and backward training, input the feature difference into the fully connected classification layer, and output the modulation type prediction result after processing by the activation function.

[0021] Furthermore, following any one or a combination of the aforementioned technical solutions, the formula for calculating the feature difference is as follows: Calculate the difference between the positive and negative features, where, The output of the feature extraction network is used for positive training. The output of the reverse training feature extraction network, z For characteristic differences, x Enhanced features for the input.

[0022] Furthermore, following any or a combination of the aforementioned technical solutions, the signal quantity detection module, the single-component recognition model, and the multi-component recognition model are all trained based on a gradient descent optimizer. During the training process, a learning rate decay strategy and an early stopping strategy are set. The early stopping strategy is to stop training when the validation set loss does not decrease for a preset number of consecutive rounds, and save the model weight with the highest validation set accuracy.

[0023] Furthermore, following any or a combination of the aforementioned technical solutions, the normalization process in step 1 is to map the pixel values ​​of the time-frequency image to the [0,1] interval; the communication signal is a modulated communication signal containing Gaussian white noise interference, and the modulated communication signal includes a combination of one or more modulation types such as phase shift keying, frequency shift keying, amplitude shift keying, quadrature amplitude modulation, and continuous phase modulation.

[0024] According to another aspect of the present invention, a multi-component communication signal identification device is provided, comprising a data preprocessing module, a signal quantity detection module, and a signal identification module connected in sequence; The data preprocessing module is used to perform time-frequency transformation on the received communication signal, generate a time-frequency image of a preset size, complete normalization processing, and output time-frequency image features. The signal quantity detection module is used to receive the time-frequency image features and output the signal component quantity result. The signal quantity detection module consists of multiple convolutional neural network sub-models and an ensemble learning unit. Each convolutional neural network sub-model is trained with focus loss as the loss function. The ensemble learning unit is used to weight and fuse the prediction results of each convolutional neural network sub-model and determine the signal component quantity. The signal recognition module integrates a feature enhancement submodule and an inverse difference submodule, and also includes independent single-component recognition units and multi-component recognition units. The feature enhancement submodule performs denoising and purification operations and multi-scale feature fusion operations on time-frequency image features, outputting enhanced features. The inverse difference submodule performs category difference amplification operations on the enhanced features, outputting high-discrimination features. The single-component recognition unit and the multi-component recognition unit are selectively activated based on the number of signal components to classify the modulation type of the high-discrimination features and output the recognition result.

[0025] Furthermore, following any or a combination of the aforementioned technical solutions, the convolutional neural network sub-model adopts a network structure of "convolutional layer, batch normalization layer, activation function connected in sequence, followed by pooling layer, then convolutional layer, batch normalization layer, activation function connected in sequence, and finally connected to fully connected layer"; the activation function is ReLU or its improved function, and the pooling layer is max pooling layer or average pooling layer.

[0026] According to another aspect of the present invention, an electronic device is provided, comprising: at least one processor, at least one memory, and computer-executable instructions stored in the memory; the processor is electrically connected to the memory, and when the computer-executable instructions are executed by the processor, the processor causes the processor to perform the steps of the multi-component communication signal identification method as described in any of the preceding claims.

[0027] According to another aspect of the present invention, a computer-readable storage medium is provided, wherein computer-executable instructions are stored thereon, and when executed by a processor, the computer-executable instructions cause the processor to perform the steps of the multi-component communication signal identification method as described in any of the preceding claims.

[0028] The beneficial effects of the technical solution provided by this invention are as follows: a. Adaptive identification of signal component quantity to suit real-world signal scenarios in complex electromagnetic environments: The system employs a two-level processing framework of "signal component quantity detection + targeted identification". Through a detection module integrating focus loss and multi-convolutional neural network sub-models, it can automatically identify different types of signals such as noise, single components, and dual / triple / quadruple components. This overcomes the limitations of existing technologies that pre-assume the number of signal components and solves the problem of low identification accuracy caused by the randomness and uncertainty of the number of signal components in real-world electromagnetic environments, thus achieving adaptive processing of all types of signals. b. Excellent adaptability to low signal-to-noise ratio and robustness of feature recognition, resulting in a significant improvement in recognition performance: This scheme uses an autoencoder to dynamically learn noise threshold denoising sub-blocks, accurately balancing noise suppression and effective feature preservation. It combines multi-scale channel attention and adaptive feature fusion attention sub-blocks to achieve in-depth feature purification. Then, through the forward and reverse dual training mechanism of the reverse difference block, it actively amplifies the category feature differences of signals with different modulation types, effectively solving the problems of noise interference and mutual interference of multi-component signal features under low signal-to-noise ratio. It maintains a stable high recognition accuracy in the signal-to-noise ratio range of -14dB to 10dB, and can still maintain good recognition performance when the signal parameters deviate from the training range. Its robustness is significantly better than existing models. c. The model design balances performance and engineering practicality, and its flexible deployment adapts to various scenarios: The core module adopts a lightweight and efficient network structure design. The training and inference processes of the detection and recognition modules are clearly divided and seamlessly connected, balancing high performance and computational efficiency in signal recognition. At the same time, the hardware adaptation requirements are clearly defined, supporting heterogeneous computing of CPU and GPU. Model training and inference can be implemented based on Python and PyTorch frameworks, and it is compatible with GPU-accelerated inference mode. It can be deployed with low latency on various hardware platforms such as electronic reconnaissance equipment and spectrum monitoring terminals, adapting to various engineering application scenarios such as edge computing, electronic reconnaissance, spectrum monitoring, and electronic countermeasures, making it highly practical. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A schematic diagram of a multi-component communication signal identification method provided as an exemplary embodiment of the present invention; Figure 2 A schematic diagram of the overall technical architecture provided for an exemplary embodiment of the present invention; Figure 3 A schematic diagram of a signal quantity detection process is provided as an exemplary embodiment of the present invention; Figure 4 A schematic diagram of a feature enhancement block provided for an exemplary embodiment of the present invention; Figure 5 A reverse difference diagram is provided for an exemplary embodiment of the present invention; Figure 6 A schematic diagram of a multi-component communication signal identification device provided for an exemplary embodiment of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0032] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0033] In one embodiment of the present invention, such as Figure 1As shown, a method for identifying multi-component communication signals is provided, including the following steps: Step 1, data preprocessing: performing time-frequency transformation on the received communication signal to generate a time-frequency image of a preset size, and normalizing the time-frequency image to obtain time-frequency image features; Step 2, signal quantity detection: inputting the time-frequency image features into a pre-trained signal quantity detection module, and outputting the signal component quantity result, which includes noise, single-component signals, and multi-component signal types; the signal quantity detection module consists of multiple convolutional neural network sub-models and an ensemble learning unit. Each convolutional neural network sub-model is trained using focus loss as the loss function, and the ensemble learning unit performs weighted fusion of the prediction results of each convolutional neural network sub-model, and determines the signal quantity based on the fusion result. Number of signal components; Step 3, Targeted identification: Based on the number of signal components, the time-frequency image features are routed to the corresponding pre-trained identification model. After feature enhancement processing and category difference amplification processing, the modulation type identification result of the communication signal is output. If the number of signal components is noise, the noise identification result is directly output. If it is a single-component signal, the time-frequency image features are input into the single-component identification model. If it is a multi-component signal, the time-frequency image features are input into the multi-component identification model. The single-component identification model and the multi-component identification model are equipped with feature enhancement blocks and inverse difference blocks. The feature enhancement blocks perform denoising and purification operations and multi-scale feature fusion operations on the time-frequency image features. The inverse difference blocks are used to amplify the category feature differences of communication signals with different modulation types.

[0034] This invention, through the construction of a multi-component communication signal recognition system based on "signal quantity detection - targeted feature enhancement - category difference amplification," overcomes the technical bottlenecks of traditional recognition methods, such as poor adaptability to the number of signal components, inadequate feature extraction in low-noise environments, and mutual interference between features of multiple signal components. It possesses several core advantages: First, it achieves accurate classification of noise, single-component, and multi-component signals through weighted fusion of multiple convolutional neural network sub-models, and then routes the signals to the corresponding recognition model, significantly improving the system's adaptability to signals with different component quantities. Second, the feature enhancement block combines dynamic threshold denoising and multi-scale channel attention feature fusion within an autoencoder framework, achieving effective feature fusion even under Gaussian white noise interference. First, it accurately filters and fully extracts multi-dimensional feature information of the signal, significantly improving the feature extraction quality in low signal-to-noise ratio environments. Second, the reverse difference block effectively amplifies the category feature differences of communication signals with different modulation types by calculating the feature difference through forward and reverse dual training, solving the problem of recognition ambiguity caused by mutual interference of multi-component signal features, and greatly improving the accuracy of modulation type recognition. Third, the models of the entire method system are all based on gradient descent optimizers for efficient training and support hardware-accelerated inference. While ensuring high recognition accuracy, it also takes into account inference efficiency. It can be adapted to various complex electromagnetic environments such as radio frequency communication, spectrum monitoring, and electronic reconnaissance, and has practicality, versatility, and engineering implementation value.

[0035] This invention discloses an exemplary embodiment of a method for identifying multi-component communication signals, such as... Figure 2-5 As shown, this method is applicable to electronic reconnaissance, spectrum monitoring, and electronic countermeasures scenarios in complex electromagnetic environments. It can achieve adaptive detection and accurate modulation type identification of noise, single-component, dual-component, three-component, and four-component communication signals, maintaining stable identification performance within a signal-to-noise ratio range of -14dB to 10dB. This method is implemented using the Python 3.9+ programming language and the PyTorch 2.1+ deep learning framework, and can be deployed on computing devices equipped with CPUs (≥2.40GHz) and GPUs (such as NVIDIA GeForce RTX 3090). It supports low-latency inference in edge computing scenarios. Those skilled in the art can implement this method without creative effort based on the content of this embodiment.

[0036] The hardware computing platform used in the hardware environment meets the following configuration requirements: the central processing unit is an Intel Core i9-12900K (3.20GHz, ≥2.40GHz basic requirement), the graphics processor is an NVIDIA GeForce RTX 3090 (24GB VRAM, supporting CUDA 11.7 and above), the memory is 64GB DDR5, and the storage device is a 1TB NVMe solid-state drive; if it is an edge computing deployment, Jetson AGX Orin embedded hardware can be used.

[0037] The software environment consists of Ubuntu 22.04 LTS 64-bit (or Windows 10 / 11 64-bit) as the operating system, Python 3.9.18 as the programming language, PyTorch 2.1.2 as the deep learning framework, torchvision 0.16.2 as the accompanying library, NumPy 1.26.2 and Pandas 2.1.4 as the numerical computing libraries, OpenCV-Python 4.8.1 and Matplotlib 3.8.2 as the image processing libraries, and scikit-learn 1.3.2 (for dataset partitioning and performance evaluation) and tqdm 4.66.1 (for training progress visualization) as the auxiliary libraries.

[0038] Construction of Multi-Component Communication Signal Dataset: To validate and train the recognition model of this method, a multi-component communication signal dataset covering noise, single-component, dual-component, three-component, and four-component signals was constructed. The dataset includes six typical communication modulation signals: Quadrature Phase Shift Keying (QPSK), Two-Frequency Shift Keying (2FSK), Four-Frequency Shift Keying (4FSK), Amplitude Shift Keying (ASK), 16-QAM, and Continuous Phase Modulation (CPM). The specific construction steps are as follows: Single-component signal generation: Based on the modulation principle of communication signals, the above six single-component signals are generated. The signal sampling frequency is set to 10MHz and the number of sampling points is 2048. The carrier frequency (0.5~2MHz), symbol rate (100k~1Mbaud), and modulation index of each signal are randomly varied within a reasonable range to simulate the dynamics of signal parameters in the actual electromagnetic environment. Multi-component signal synthesis: Different types of single-component signals are superimposed in the time domain to generate two-component, three-component, and four-component mixed communication signals. The amplitude ratio of each component signal during superposition is 0.5~2.0 and the time delay difference is 0~10μs, all of which are randomly set to simulate the energy and time differences of different component signals in real-world scenarios. Noise superposition and sample generation: Gaussian white noise is superimposed on noise, single-component, dual-component, three-component, and four-component signals to generate signal samples with a signal-to-noise ratio (SNR) ranging from -14dB to 10dB, with an SNR interval of 2dB. 300 samples are generated for each SNR point; among them, the noise samples are pure Gaussian white noise with no effective signal components. Dataset partitioning: Of the 300 samples at each signal-to-noise ratio point, 200 are used for model training and 100 for model testing; the dataset as a whole contains 1700 noisy samples, 10200 single-component signal samples, 15300 two-component signal samples, 20400 three-component signal samples, and 15300 four-component signal samples, for a total of 62900 samples; during training, 20% of the training set is extracted as a validation set to monitor model convergence, with no class bias; Time-frequency transformation and preprocessing: All signal samples were subjected to time-frequency transformation using the Choi-Williams distribution (CWD) to convert the one-dimensional time-domain signal into a 256×256 two-dimensional time-frequency image. The Choi-Williams distribution kernel function parameter σ=1, and the horizontal axis of the time-frequency image was the time axis and the vertical axis was the frequency axis, which fully captured the time-frequency energy distribution characteristics of the signal. The pixel values ​​of the generated time-frequency image were normalized to the [0,1] interval to eliminate the influence of scale differences on model training. At the same time, the training set samples were randomly horizontally flipped and rotated at a small angle (±10°) to perform data augmentation operations to improve the generalization ability of the model.

[0039] Model Construction and Parameter Settings: The recognition model of this method consists of a signal quantity detection module and a signal recognition module. The signal recognition module includes a feature enhancement module and an inverse difference module. The signal recognition module is divided into a single-component recognition model and a multi-component recognition model (adapted to two / three / four components). The network structure of the two recognition models is the same, only the training dataset is different. The structure and parameter settings of each module / model are as follows: The signal quantity detection module takes 256×256 CWD time-frequency image features as input and outputs the number of signal components (0 for noise, 1 for a single component, 2 for two components, 3 for three components, and 4 for four components). It consists of three CNN sub-models and an ensemble learning unit. CNN Sub-model Structure: The three sub-models have the same structure, all following the pattern "Conv(3×3)+BN+ReLU→MaxPool→Conv(3×3)+BN+ReLU→Fully Connected Layer"; the first convolutional layer has 64 kernels, a stride of 1, and is padded with the same type of kernel, followed by a BN layer and a ReLU activation function; the first max pooling layer has 2×2 kernels, a stride of 2, and is padded with the same type of kernel; the second convolutional layer has 128 kernels, a stride of 1, and is padded with the same type of kernel, followed by a BN layer and a ReLU activation function; the second max pooling layer has 2×2 kernels, a stride of 2, and is padded with the same type of kernel; the fully connected layer has 8192 input neurons and 5 output neurons, corresponding to the classification results of the 5 component counts; Loss function parameters: Focal loss is used as the loss function, and the formula is as follows:

[0040] Among them, the category weight coefficient =[0.3, 0.25, 0.2, 0.15, 0.1] (corresponding to components 0 to 4 respectively), focusing parameters r =2, effectively solving the class imbalance problem; Ensemble learning strategy: A weighted average is used to fuse the prediction results of three sub-models, with the sub-model weights... =[0.4, 0.35, 0.25] (weight sum is 1), the fusion formula is:

[0041] pass Determine the number of signal components; The signal recognition module takes 56×256 CWD time-frequency image features as input and outputs the communication signal modulation type (single component is a single modulation type, multi-component is a combination of modulation types). The core consists of a feature enhancement module and an inverse difference module, and the end is connected to a fully connected classification layer. Feature enhancement module: Composed of a denoising submodule and an attention submodule, it achieves denoising and purification of signal features and multi-scale fusion. Denoising submodule: Built on an autoencoder framework, the encoder consists of 3 convolutional layers (3×3 kernels, 1 stride, same padding) to progressively compress feature dimensions. The decoder consists of 3 deconvolutional layers (3×3 kernels, 1 stride, same padding) to restore feature dimensions. Global average pooling is used (…). GAP (4×4 window) Dynamically learned noise threshold τ = GAP ( X rec And select valid features using the following formula:

[0042] Attention Sub-block: A multi-scale channel attention module (MS-CAM) and adaptive feature fusion (AFF) strategy are employed. MS-CAM uses 1×1, 3×3, and 5×5 convolutional kernels to extract fine-grained, medium-grained, and coarse-grained features respectively, with a channel attention compression ratio of 16. The AFF module adaptively learns and fuses weights A, and then... Fuse features from different scales; then output enhanced features through residual connections. To avoid feature loss; Reverse Difference Block: Based on a dual training mechanism of forward and reverse directions, it amplifies the differences in class features of signals with different modulation types: Forward Training: Based on the original labels y Using cross-entropy loss Where c=6 (six modulation signals). One-hot encoding for the original label, Predict probabilities for positive labels; back-end training: based on back labels. Using cross-entropy loss ,in Predicting probabilities for inverse labels; Calculating feature differences: through... Calculate the difference between the positive and negative features, where, The output of the feature extraction network is used for positive training. This is the output of the reverse-training feature extraction network; Fully connected classification layer: 2048 input neurons, 6 output neurons, followed by a Softmax activation function, outputting the predicted probability of each modulation type; the network structure and parameters of the single-component recognition model and the multi-component recognition model are completely identical, only the training datasets are single-component signal dataset and dual / triple / quad-component signal dataset respectively. All models used the Adam optimizer with a weight decay factor of 1e-5 for training hyperparameters. =0.9、 =0.999; the learning rate for the signal quantity detection module is 1e-4, and the learning rate for the single-component / multi-component recognition model is 5e-5; the batch size for all models is 32, and a learning rate decay strategy is adopted, with the learning rate decaying to 0.5 every 20 training rounds; an early stopping strategy is set, and if the loss on the validation set does not decrease for 10 consecutive rounds, training is stopped to avoid overfitting; the weight file of the model with the highest accuracy on the validation set is saved for subsequent inference.

[0043] Model training process: The model training of this method is divided into three independent stages: signal quantity detection module, single-component recognition model, and multi-component recognition model. After training, the modules and models are connected in series to form a complete recognition system. The specific steps are as follows: Signal quantity detection module training: Input the constructed multi-component communication signal dataset time-frequency images into 3 initialized CNN sub-models, train them for 100 rounds using the Adam optimizer according to the above parameter settings, calculate the training set loss and validation set accuracy in each round, and save the optimal weights of each sub-model after training is completed; based on the weights, fuse the prediction results of the sub-models according to the weighted average strategy to determine the final output of the module; Single-component recognition model training: Extract time-frequency images of single-component signals from the dataset (training set + validation set), input them into the initialized signal recognition model, and train for 80 rounds with the parameters set above. The total loss is... Balance the training weights for both positive and negative sides, calculate the total loss of the training set and the recognition accuracy of the validation set in each round, and save the optimal weights of the model after training is completed. Multi-component recognition model training: Extract time-frequency images of dual / triple / quad-component signals (training set + validation set) from the dataset, input them into the initialized signal recognition model, train for 100 rounds according to the above parameter settings, with the loss function and training strategy being the same as the single-component recognition model, and save the optimal weights of the model after training is completed; Model fusion: The trained signal quantity detection module is connected in series with the single-component / multi-component recognition model. The feature scale of each module is unified through the batch normalization (BN) layer, so that the output of the signal quantity detection module can be directly used as the routing basis of the recognition model, forming an end-to-end multi-component communication signal recognition model, and saving the overall model weight file.

[0044] Model reasoning process The trained recognition model can adaptively recognize actual received communication signals, fully following the logic of data preprocessing → signal quantity detection → targeted recognition. The specific reasoning steps are as follows: Signal reception: Communication signals y(t) in complex electromagnetic environments are acquired using signal acquisition equipment such as radio frequency receivers. The mathematical model of the signal is: ,in For the first i The amplitude of each component signal For the first i The phase function of each component signal. It is Gaussian white noise. k The number of signal components; Data preprocessing: processing the received communication signals The Choi-Williams distribution (CWD) was used for time-frequency transformation to generate a 256×256 two-dimensional time-frequency image. The image pixel values ​​were normalized to the [0,1] interval to complete the preprocessing. Signal component count detection: The preprocessed time-frequency image is input into the signal component count detection module. The module uses three CNN sub-models for weighted fusion and outputs the signal component count. k (k =0 indicates noise. k =1 indicates a single component. k =2 indicates a two-component structure. k =3 represents three components. k =4 represents four components); Targeted identification: Based on the results of component quantity detection k The time-frequency images are routed to the corresponding recognition models: like k =0, directly outputting the recognition result as noise, without needing to proceed to the subsequent recognition process; like k =1, the time-frequency image is input into the single-component recognition model, and after the feature enhancement block denoises and fuses, and the inverse difference block amplifies the category difference, the modulation type of the single-component communication signal is output by the fully connected classification layer; like k ≥2, time frequency image The input multi-component recognition model is processed by feature enhancement blocks and inverse difference blocks, and the fully connected classification layer outputs the modulation type combination of each component in the multi-component communication signal. Recognition result output: The model outputs the final result in the form of labels. Single-component signals output single modulation type labels, and multi-component signals output combined modulation type labels, thus completing the entire recognition process.

[0045] Model performance verification The performance of this method was tested on a constructed multi-component communication signal dataset and compared with existing mainstream models (DCNN, SE-InceptionNet). The test metric was the average recognition accuracy. The verification results are as follows: Adaptability to different signal-to-noise ratios: This method maintains stable high accuracy in the signal-to-noise ratio range of -14dB to 10dB; while the DCNN model and the SE-InceptionNet model have relatively low accuracy. The method in this embodiment has significantly better adaptability to low signal-to-noise ratios. Performance of different component counts: At a signal-to-noise ratio of -10dB, this method significantly improves the accuracy of noise recognition. The accuracy of single-component, dual-component, three-component, and four-component methods is also improved compared to existing technologies. In related technologies, the accuracy of models for three / four-component signals drops significantly. The accuracy of four-component recognition in the DCNN model is less than 65%, and that in the SE-InceptionNet model is only 70%. Parameter robustness: When signal parameters (carrier frequency, symbol rate, amplitude ratio) that are outside the training range are substituted into the test, the average recognition accuracy of this method for signals with parameter deviations is still greater than 0% at a signal-to-noise ratio of 0dB, with only a slight decrease. The method in this embodiment is more adaptable to dynamic changes in signal parameters. Inference efficiency: On the NVIDIA GeForce RTX 3090 GPU, the inference time for a single time-frequency image is ≤12ms (4ms for the detection module and 8ms for the recognition module); on the Jetson AGX Orin embedded hardware, after 30% pruning optimization, the inference time for a single time-frequency image is ≤20ms, meeting the low latency requirements of edge computing.

[0046] In one embodiment of the present invention, a multi-component communication signal identification device is provided, such as... Figure 6 As shown, the device used to perform all the steps of the above-mentioned multi-component communication signal identification method is a combination of logical functional modules that can be deployed on the above-mentioned hardware computing platform. It is implemented based on the Python 3.9+ and PyTorch 2.1+ frameworks and can achieve the same technical effect as the above-mentioned method. It is suitable for complex electromagnetic environment scenarios such as electronic reconnaissance, spectrum monitoring, and electronic countermeasures.

[0047] Overall structure of the device: The device includes a data preprocessing module, a signal quantity detection module, and a signal recognition module. The signal recognition module integrates a feature enhancement submodule and a reverse difference submodule. The signal recognition module also includes independent single-component recognition units and multi-component recognition units. The modules / units are electrically connected to each other. Data is transmitted unidirectionally in the order of "data preprocessing module → signal quantity detection module → signal recognition module" to achieve end-to-end signal recognition.

[0048] Functions and operating methods of each module / unit The data preprocessing module performs time-frequency transformation and normalization on externally acquired communication signals, providing standard feature inputs for subsequent modules. Specifically, it receives communication signals acquired by an RF receiver. The Choi-Williams distribution (kernel function σ=1) is used to perform time-frequency transformation on the image to generate a 256×256 two-dimensional time-frequency image. The image pixel values ​​are normalized to the [0,1] interval to eliminate scale differences. The processed CWD time-frequency image features are then transmitted to the signal quantity detection module.

[0049] The signal quantity detection module is used to detect the number of signal components in standard time-frequency image features and outputs the signal component quantity results, providing a routing basis for the signal recognition module. This module includes three CNN sub-model units and an ensemble learning unit, and its specific operation is as follows: All three CNN sub-model units adopt the structure of "Conv(3×3)+BN+ReLU→MaxPool→Conv(3×3)+BN+ReLU→fully connected layer", with focus loss as the loss function. =[0.3, 0.25, 0.2, 0.15, 0.1], γ=2), process the time-frequency image features respectively, and output the prediction probability of the number of each component; Ensemble learning units by sub-model weights =[0.4, 0.35, 0.25] is used to perform a weighted average fusion of the predicted probabilities of the three CNN sub-model units. Determine the number of signal components k (0 / 1 / 2 / 3 / 4), and the time-frequency image features and component counts. k The signal is transmitted synchronously to the signal recognition module.

[0050] The feature enhancement submodule is the core submodule of the signal recognition module. It is used to denoise and refine time-frequency image features and perform multi-scale fusion to output enhanced signal features. This submodule includes a denoising unit and a multi-scale attention fusion unit, and its specific operation is as follows: The denoising unit is based on an autoencoder framework and dynamically learns the noise threshold through global average pooling (4×4 window). τ = GAP ( X rec ),according to The screening rules retain effective features and filter out noisy features, achieving a balance between noise suppression and feature preservation. The multi-scale attention fusion unit extracts signal features at different scales through the MS-CAM module (1×1, 3×3, and 5×5 scale convolutional kernels), and after adaptive fusion by the AFF module, outputs enhanced features through residual connections. The enhanced features are then transmitted to the reverse difference submodule.

[0051] The inverse difference submodule is the core submodule of the signal recognition module. It is used to amplify the category differences of the enhanced signal features and output highly discriminative features. Specifically, it receives the enhanced features transmitted from the feature enhancement submodule, performs forward training based on the original label y, and then performs inverse training based on the inverse label. Perform reverse training and calculate the forward cross-entropy loss separately. and inverse cross-entropy loss ;pass Calculate the forward and reverse feature differences, amplify the category feature distance of signals with different modulation types, and transmit the feature differences to a single-component recognition unit or a multi-component recognition unit.

[0052] The single-component recognition unit / multi-component recognition unit is the execution unit of the signal recognition module, used to classify modulation types of high-discrimination features and output the final recognition result; the two units have completely identical structures and parameters, only adapting to signals with different numbers of components; the specific working mode is as follows: If the signal quantity detection module outputs k=1, then the single-component recognition unit is activated and the multi-component recognition unit is dormant; the single-component recognition unit receives the feature difference value from the inverse difference submodule, and outputs the modulation type label of the single-component communication signal through the fully connected classification layer (2048 input neurons, 6 output neurons) and the Softmax activation function; If the signal quantity detection module outputs k If the value is ≥2, the multi-component recognition unit is activated and the single-component recognition unit is in sleep mode. The multi-component recognition unit receives the feature difference value from the inverse difference submodule, and outputs the modulation type combination label of the multi-component communication signal through the same fully connected classification layer and Softmax activation function. If the signal quantity detection module outputs k If the value is 0, all units of the signal recognition module will go into sleep mode and directly output the "noise" recognition result.

[0053] In one embodiment of the present invention, an electronic device is provided as the hardware carrier of the above-mentioned multi-component communication signal identification method and device. It can realize adaptive identification of multi-component communication signals in complex electromagnetic environments and is applicable to products such as electronic reconnaissance equipment, spectrum monitoring terminals, and edge computing nodes. The hardware configuration of the device meets the operating requirements of the above-mentioned method / device, supports GPU accelerated inference, has low latency and high accuracy, and can be directly implemented in actual engineering scenarios.

[0054] The electronic device includes at least one processor, at least one memory, and a radio frequency communication interface. It may also include auxiliary modules such as a display module and a storage expansion interface. The memory and the processor are electrically connected via a system bus. The radio frequency communication interface is electrically connected to both the processor and the memory. The auxiliary modules are electrically connected to the processor. The system bus can be divided into an address bus, a data bus, and a control bus to realize information transmission between various hardware components.

[0055] Functions and operating methods of each hardware component The radio frequency (RF) communication interface is used to acquire external communication signals and convert analog signals into digital signals to provide raw signal input to the processor. This interface supports a 10MHz signal bandwidth, a 10MHz sampling frequency, and a 16-bit sampling precision. Specifically, it receives communication signals in complex electromagnetic environments through an RF antenna, converts the analog signals into digital time-domain signals through analog-to-digital conversion, and transmits the digital signals to the processor. It also supports parallel reception of multiple channels (up to 8 channels), making it suitable for large-scale spectrum monitoring scenarios.

[0056] The memory is a non-volatile computer-readable memory, including media such as ROM, RAM, solid-state drives, and USB flash drives that can store computer program code; it is used to store computer-executable instructions for implementing the above-mentioned multi-component communication signal recognition method, the trained model weight file (including the weights of 3 CNN sub-models, feature enhancement blocks, and inverse difference blocks), and temporary data generated during processor operation (such as acquired digital signals, time-frequency image features, recognition results, etc.); the storage capacity of the memory is not less than 512GB to ensure the stable storage of the dataset and model file.

[0057] The processor is a central processing unit (CPU, clock speed ≥ 2.40GHz) or a graphics processing unit (GPU, such as NVIDIA GeForce RTX 3090, Jetson AGX Orin), or a CPU+GPU heterogeneous computing architecture; it is used to call computer-executable instructions and model weight files in memory to execute all the steps of the above-mentioned multi-component communication signal recognition method; the specific working mode is as follows: The system receives digital time-domain signals transmitted via the radio frequency communication interface, performs Choi-Williams distributed time-frequency transformation and normalization processing, and generates 256×256 CWD time-frequency image features. The model weights of the signal quantity detection module are invoked to perform component quantity detection on the time-frequency image features, and the output is... k value; according to k The value calls the corresponding single-component / multi-component recognition model weights, sequentially performs feature enhancement, inverse difference calculation, and modulation type classification, and outputs the recognition result; the processor supports batch processing and parallel inference, the inference time of a single time-frequency image is ≤12ms, and the single-channel delay remains unchanged when multi-channel signals are in parallel inference.

[0058] The auxiliary module display module is an LCD screen or touch screen, used to visually display the recognition results (such as the number of signal components, modulation type, recognition accuracy, etc.); the storage expansion interface is USB3.0, Type-C, SD card interface, etc., used to expand storage capacity or export recognition results and raw data.

[0059] Equipment workflow: The electronic device is started, the radio frequency communication interface, memory and processor are initialized, the processor calls the computer-executable instructions and model weight file in memory, and enters the identification state; The radio frequency communication interface collects external communication signals through an antenna, and transmits the digital signals to the processor after analog-to-digital conversion; The processor performs data preprocessing on the digital signal to generate standard CWD time-frequency image features; The processor performs signal component count detection and outputs... kvalue; processor according to k The value is routed to the corresponding recognition model, where feature enhancement, reverse difference calculation, and classification recognition are performed to generate the recognition result. The recognition results are visualized through the display module and stored in the memory, and can be exported through the storage expansion interface; Repeat the above steps to achieve continuous, real-time identification of multi-component communication signals.

[0060] In one embodiment of the present invention, a computer-readable storage medium is provided. This storage medium is a non-volatile storage medium storing a computer program. When the computer program is executed by a processor, it implements all the steps of the above-described multi-component communication signal identification method, achieving the same technical effect as the above-described method, apparatus, and electronic device. This storage medium can be ported and used between different computing devices. As long as the processor of the computing device meets the requirements of a main frequency ≥ 2.40 GHz and supports Python 3.9+ and PyTorch 2.1+ frameworks, adaptive identification of multi-component communication signals in complex electromagnetic environments can be realized, exhibiting good portability, versatility, and compatibility.

[0061] Storage media type The computer-readable storage medium includes, but is not limited to, any non-volatile medium capable of storing computer program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, solid-state drives (SSDs), and embedded storage chips.

[0062] Computer program execution methods The computer program is the complete program code for implementing a multi-component communication signal recognition method, including all program logic for data preprocessing, signal quantity detection, feature enhancement, reverse difference calculation, and targeted recognition. It also includes code for calling the trained model weight file. When the processor reads and executes this computer program, the specific execution flow is as follows: The processor receives externally acquired communication signals through the communication interface. Execute the data preprocessing procedure: use the Choi-Williams distribution transformation on the communication signal to generate a 256×256 time-frequency image, and normalize the image pixel values ​​to the [0,1] interval; The processor executes a signal quantity detection program: it calls the model weight file, inputs the time-frequency image features into three CNN sub-models, and each sub-model uses a focus loss ( =[0.3, 0.25, 0.2, 0.15, 0.1], γ =2) is the loss function output predicted probability, which is then weighted and fused ( =[0.4, 0.35, 0.25]) and then passed Determine the number of signal components k ; The processor executes a targeted identification program: like k =0, the processor directly outputs the recognition result as noise; like k =1, the processor calls the single-component recognition model weights, executes the feature enhancement program (dynamic denoising + multi-scale attention fusion) and the reverse difference program (forward and reverse training + feature difference calculation), and outputs the modulation type of the single-component communication signal after processing the time-frequency image features; like k ≥2, the processor calls the multi-component recognition model weights, executes the same feature enhancement and inverse difference program, and outputs the modulation type combination of the multi-component communication signal; The processor transmits the recognition result to the display module or storage module to complete one signal recognition; if it is a continuous recognition, the processor repeats the above steps 1-4.

[0063] The storage capacity of the storage medium shall not be less than 10GB to ensure the complete storage of the computer program code and model weight files; The read speed of the storage medium is no less than 100MB / s to ensure the efficiency of the processor in calling programs and models and to avoid excessive inference latency; The connection interface between the storage medium and the processor is a universal interface (such as USB, SATA, PCIe) to ensure the stability and compatibility of data transmission.

[0064] This invention discloses a method, apparatus, electronic device, and computer-readable storage medium for identifying multi-component communication signals. This technical solution solves the core problems in the prior art, such as poor component quantity adaptability, poor low signal-to-noise ratio feature extraction effect, and severe multi-component feature interference, through an innovative framework of "signal quantity detection + targeted identification" combined with dynamic threshold denoising, multi-scale attention fusion, and forward and reverse dual training mechanisms. It achieves adaptive and accurate identification of noisy, single-component, and dual / triple / quad-component communication signals.

[0065] Experimental and verification results show that the technical solution maintains a stable high recognition accuracy within a signal-to-noise ratio range of -14dB to 10dB, exhibits good robustness to dynamic changes in signal parameters, has high model inference efficiency and flexible deployment, and can be adapted to various computing platforms such as CPU / GPU and embedded hardware. It can be directly applied to engineering scenarios in complex electromagnetic environments such as electronic reconnaissance, spectrum monitoring, and electronic countermeasures.

[0066] The technical solution of this embodiment is fully disclosed, and all steps, structures and parameters are clearly defined. Those skilled in the art can implement the entire process without creative effort based on the content of this embodiment. They only need to fine-tune details such as signal type, dataset parameters, and model hyperparameters according to actual application needs to adapt to the multi-component communication signal recognition needs in different complex electromagnetic environments.

[0067] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0068] The above description is only a specific embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for identifying multi-component communication signals, characterized in that, Includes the following steps: Step 1: Perform time-frequency transformation on the received communication signal to generate a time-frequency image of a preset size, and normalize the time-frequency image to obtain time-frequency image features; Step 2: Input the time-frequency image features into a pre-trained signal quantity detection module and output the signal component quantity result; wherein, the signal component quantity result includes noise, single-component signal and multi-component signal types; The signal quantity detection module consists of multiple convolutional neural network sub-models and an ensemble learning unit. The multiple convolutional neural network sub-models are trained using focus loss as the loss function. The ensemble learning unit performs weighted fusion on the prediction results of the multiple convolutional neural network sub-models and determines the number of signal components based on the weighted fusion result. Step 3: Perform targeted processing on the time-frequency image features based on the signal component count result. If the signal component count result is noise, output the noise identification result directly. If the signal component count result is a single-component signal, input the time-frequency image features into a single-component identification model, and after feature enhancement processing and category difference amplification processing, output the modulation type identification result of the single-component communication signal. If the signal component count result is a multi-component signal, input the time-frequency image features into a multi-component identification model, and after feature enhancement processing and category difference amplification processing, output the modulation type identification result of the multi-component communication signal. The single-component recognition model and the multi-component recognition model are equipped with a feature enhancement module and an inverse difference module. The feature enhancement module performs denoising and purification operations and multi-scale feature fusion operations on the time-frequency image features. The inverse difference module is used to amplify the category feature differences of communication signals with different modulation types.

2. The multi-component communication signal identification method according to claim 1, characterized in that, The formula for calculating the focus loss is: Among them, the category weight coefficient =[0.3, 0.25, 0.2, 0.15, 0.1], Focusing Parameter r =2, Let be the predicted probability of each convolutional neural network sub-model for the number of the c-th class component. K The maximum number of categories in terms of component quantity; The weighted fusion formula for the integrated learning unit is: ； in, Predict the probability of the number of the c-th component after fusion. Let be the weight of the m-th convolutional neural network sub-model, and the sum of the weights of all convolutional neural network sub-models is 1. Let be the predicted probability of the m-th convolutional neural network sub-model for the number of c-th class components; The integrated learning unit, through Determine the final number of signal components.

3. The multi-component communication signal identification method according to claim 1 or 2, characterized in that, The feature enhancement module includes a denoising sub-block and an attention sub-block, and the processing procedure is as follows: The denoising submodule is built on an autoencoder framework and dynamically learns the noise threshold through global average pooling. τ = GAP ( X rec ), X rec The effective features in the time-frequency image features are selected as the reconstruction features of the autoencoder using preset rules; The attention submodule employs a multi-scale channel attention module and an adaptive feature fusion strategy to extract multi-scale features of the time-frequency image features and complete adaptive fusion. The fused features are added to the denoised features via residual connections to output enhanced features. ,in, X As a feature of fusion, X These are the features after denoising.

4. The multi-component communication signal identification method according to claim 1, characterized in that, The preset rule is: pixel values ​​in the time-frequency image features that are greater than a noise threshold are considered. τ The features are retained, and the remaining features are set to 0. The formula is: in, These are the feature pixel values ​​after denoising. The pixel values ​​of the original time-frequency image features. The coordinates of the feature pixel.

5. The multi-component communication signal identification method according to claim 1, characterized in that, The calculation formula for the adaptive feature fusion is as follows: in, A For adaptive learning fusion weights, X m , X n Features of different scales extracted by the multi-scale channel attention module.

6. The multi-component communication signal identification method according to claim 1, characterized in that, The processing procedure of the reverse difference module is as follows: Forward training: Based on the original labels of communication signals y The feature extraction network is trained using cross-entropy loss to obtain positive feature output. The cross-entropy loss formula is as follows: ,in c The number of modulation types for communication signals. One-hot encoding for the original label, Predict the probability for the positive label; Reverse training: based on reverse labels The feature extraction network is trained using cross-entropy loss to obtain inverse feature output. The cross-entropy loss formula is as follows: ,in Predict probabilities for inverse labels; Feature difference calculation: Calculate the feature difference between forward training and backward training, input the feature difference into the fully connected classification layer, and output the modulation type prediction result after processing by the activation function.

7. The multi-component communication signal identification method according to claim 6, characterized in that, The formula for calculating the feature difference is: in, The output of the feature extraction network is used for positive training. The output of the reverse training feature extraction network, z For characteristic differences, x Enhanced features for the input.

8. A multi-component communication signal identification device, characterized in that, It includes a data preprocessing module, a signal quantity detection module, and a signal recognition module; The data preprocessing module is used to perform time-frequency transformation on the received communication signal, generate a time-frequency image of a preset size, complete normalization processing, and output time-frequency image features. The signal quantity detection module is used to receive the time-frequency image features and output the signal component quantity result. The signal quantity detection module consists of multiple convolutional neural network sub-models and an ensemble learning unit. The multiple convolutional neural network sub-models are trained with focus loss as the loss function. The ensemble learning unit is used to weight and fuse the prediction results of the multiple convolutional neural network sub-models and determine the signal component quantity. The signal recognition module integrates a feature enhancement submodule and an inverse difference submodule, and also includes independent single-component recognition units and multi-component recognition units. The feature enhancement submodule is used to perform denoising and purification operations and multi-scale feature fusion operations on the time-frequency image features, and outputs enhanced features. The reverse difference submodule is used to perform a category difference amplification operation on the enhanced features and output high-discrimination features; the single-component recognition unit and the multi-component recognition unit are selected to start according to the number of signal components, and the high-discrimination features are classified by modulation type and the recognition result is output.

9. An electronic device, characterized in that, include: The method comprises at least one processor, at least one memory, and computer-executable instructions stored in the memory; the processor is electrically connected to the memory, and when the computer-executable instructions are executed by the processor, the processor causes the processor to perform the steps of the multi-component communication signal identification method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, cause the processor to perform the steps of the multi-component communication signal identification method as described in any one of claims 1-7.

Images