Lightweight signal modulation category identification method, apparatus and device, and medium

Abstract

The embodiment of the invention provides a lightweight signal modulation category identification method and device, equipment and a medium, and is applied to the technical field of communication signal modulation identification. An IQ component of a modulation signal to be recognized is obtained and input into the trained lightweight gating fusion network model for analysis, a modulation recognition type is output, and the lightweight gating fusion network model comprises a time domain feature extraction module, a spatial feature extraction module, a spatial-temporal feature fusion module, a time sequence modeling module and a classification recognition module. According to the method, the model parameter quantity is reduced, the calculation complexity is reduced, and the recognition accuracy is improved.

Description

Technical Field

[0001] The present application relates to the technical field of communication signal modulation identification, and in particular to a lightweight signal modulation category identification method, apparatus, device and medium. Background Art

[0002] Automatic Modulation Recognition (AMR), a core research area in wireless communications, aims to automatically identify the modulation type of received signals. This technology has widespread applications in communications, military, and signal monitoring.

[0003] Traditional AMR algorithms fall into two main categories: likelihood-based and feature-based methods. Likelihood-based methods, based on statistical theory, classify modulation types by maximizing the likelihood function of the signal under specific assumptions. These methods have low error probabilities under ideal channel conditions and exhibit theoretical optimality. However, their complex statistical modeling and computational procedures lead to high computational costs in high-dimensional signals and complex scenarios, and their real-time and robustness are relatively limited.

[0004] In contrast, feature-based methods extract key signal features (such as amplitude, phase, instantaneous frequency, and waveform) to construct recognition models for classification. These methods have lower computational complexity and are more suitable for real-time processing. However, their performance is highly dependent on the quality of feature design and classifier selection, and may perform poorly in low signal-to-noise ratio or complex channel environments.

[0005] With the rapid development of deep learning, convolutional neural networks (CNNs) have made breakthroughs in image processing. In 2016, O'Shea et al. first introduced CNNs into automatic modulation recognition, providing new insights into radio signal modulation classification and demonstrating excellent performance in low signal-to-noise ratio environments.

[0006] A significant advantage of deep learning methods is their ability to integrate feature extraction and classification. Neural network models such as CNNs and RNNs are widely used in the AMR field. Researchers are constantly exploring model innovations: Wu et al. proposed a CLDNN network that combines CNNs and LSTMs to capture spatiotemporal correlations; Zhang et al. designed a multi-stream network that comprehensively utilizes different representations of I / Q signals; and Xu et al. developed a multi-stream architecture that focuses on solving the orthogonality problem of I / Q channels.

[0007] Although these methods have made progress in performance, key challenges still exist: large number of model parameters, complex calculations, limited accuracy, and some methods require preprocessing of the received I / Q components, which increases system latency. Summary of the Invention

[0008] The present application provides a lightweight signal modulation category identification method, apparatus, device and medium to solve the above-mentioned problems.

[0009] In a first aspect, the present application provides a method for identifying a modulation category of a lightweight signal, the method comprising:

[0010] Construct and train a lightweight gated fusion network model, obtain the IQ component of the modulation signal to be identified and input it into the trained lightweight gated fusion network model for analysis, and output the modulation identification type. The lightweight gated fusion network model includes a time domain feature extraction module, a spatial feature extraction module, a spatiotemporal feature fusion module, a time series modeling module, and a classification and recognition module;

[0011] The training process of the lightweight gated fusion network model includes the following steps:

[0012] S1: Obtain a training data set, where each set of training samples in the training data set includes IQ components of a modulated signal and corresponding category labels;

[0013] S2: Input the training sample into the spatial feature extraction module to extract the spatial features and obtain the spatial features of the modulated signal. Input the training sample into the time domain feature extraction module to extract the time domain features and obtain the time domain features of the modulated signal.

[0014] S3: Input the temporal features and spatial features into the spatiotemporal feature fusion module for feature fusion processing to obtain the first spatiotemporal fusion feature;

[0015] S4: Input the spatiotemporal fusion features into the time series modeling module to enhance the time series features and obtain the second spatiotemporal fusion features;

[0016] S5: inputting the second spatiotemporal fusion feature into the classification and recognition module for classification and recognition to obtain a modulation recognition type;

[0017] S6: Iteratively update the parameters of the lightweight gated fusion network model through the back propagation algorithm according to the preset loss function until the loss function converges or the maximum number of iterations is reached.

[0018] Optionally, the time domain feature extraction module includes: two temporal convolution layer modules, two dimensionality reshaping layers, and a first feature splicing layer. The training sample is input into the time domain feature extraction module for time domain feature extraction to obtain the time domain features of the modulated signal, including:

[0019] The I component and Q component of the modulated signal are respectively input into two parallel temporal convolutional layer modules for time domain feature extraction to obtain the time domain features of the I component and the Q component;

[0020] The time domain features of the I component and the time domain features of the Q component are respectively input into two parallel dimension reshaping layers for tensor dimension reconstruction;

[0021] The reconstructed time domain features of the I component and the Q component are fused along the height dimension using the first feature concatenation layer to obtain fused time domain features.

[0022] Optionally, the temporal convolution layer module includes: a first causal convolution layer, a batch normalization layer, two parallel second causal convolution layers and a Swish activation function.

[0023] Optionally, the spatial feature extraction module includes: a first depthwise separable convolution layer and a first pointwise convolution layer, and inputting the training sample into the spatial feature extraction module to extract the spatial features to obtain the spatial features of the modulated signal, including:

[0024] The I component and Q component of the modulated signal are simultaneously input into the first depth-wise separable convolutional layer to extract the spatial features of the modulated signal through a 2×8 convolution operation to obtain the first spatial feature of the modulated signal;

[0025] The first point-by-point convolution layer performs channel expansion on the first spatial feature through a 1×1 convolution operation to obtain the second spatial feature of the modulated signal.

[0026] Optionally, inputting the time domain features and the spatial features into a spatiotemporal feature fusion module for feature fusion processing to obtain a first spatiotemporal fusion feature includes:

[0027] The fused time domain features are input into the second depth-wise separable convolutional layer through a 1×8 convolution operation to obtain the first time domain features;

[0028] Input the first time domain feature into the second point-by-point convolution layer through a 1×1 convolution operation to obtain the second time domain feature;

[0029] The second time domain feature and the second space feature are fused through the second feature splicing layer to obtain the first time-space fusion feature.

[0030] Optionally, the timing modeling module includes: two temporal convolutional network residual blocks with the same structure, the temporal convolutional network residual block includes: a third depth-separable convolutional layer, a third point-by-point convolutional layer, a Swish activation function, a fourth depth-separable convolutional layer, a fourth point-by-point convolutional layer and a ReLU activation function, and the temporal convolutional network residual block adopts a residual connection.

[0031] Optionally, the classification and recognition module includes: a global average pooling layer, a fully connected layer, and an output layer. The second spatiotemporal fusion feature is input into the classification and recognition module for classification and recognition to obtain a modulation recognition type, including:

[0032] The second spatiotemporal fusion feature is input into the global average pooling layer for dimensionality reduction;

[0033] The second spatiotemporal fusion feature after dimensionality reduction is input into the fully connected layer to extract high-level semantic features to obtain high-dimensional features;

[0034] The output layer uses the Softmax function to process the high-dimensional features to obtain the modulation probability, and determines the modulation recognition type based on the modulation probability.

[0035] In a second aspect, the present application provides a lightweight signal modulation category identification device, comprising:

[0036] Training module, used to train lightweight gated fusion network models;

[0037] An acquisition module, configured to acquire the IQ components of the modulation signal to be identified;

[0038] The processing module is used to input the IQ components of the modulated signal into a trained lightweight gated fusion network model for analysis and output the modulation identification type. The lightweight gated fusion network model includes a time domain feature extraction module, a spatial feature extraction module, a spatiotemporal feature fusion module, a time series modeling module and a classification and recognition module.

[0039] In a third aspect, the present application provides a lightweight signal modulation category identification device, comprising:

[0040] Memory;

[0041] processor;

[0042] wherein the memory stores computer-executable instructions;

[0043] The processor executes the computer-executable instructions stored in the memory to implement the lightweight signal modulation category identification method as described in the first aspect and various possible implementations of the first aspect.

[0044] In a fourth aspect, the present application provides a computer storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the lightweight signal modulation category identification method as described in the first aspect and various possible implementations of the first aspect.

[0045] The present application provides a lightweight signal modulation category recognition method, device, equipment and medium. The method constructs and trains a lightweight gated fusion network model, obtains the IQ component of the modulation signal to be identified, inputs the trained lightweight gated fusion network model for analysis, and outputs the modulation recognition type. The lightweight gated fusion network model includes a time domain feature extraction module, a spatial feature extraction module, a spatiotemporal feature fusion module, a timing modeling module and a classification recognition module. The model greatly reduces the model complexity through a dual-branch asynchronous feature extraction architecture, adopts deep separable convolution and an efficient timing module to reduce computational time; strengthens feature interaction through channel reshaping and attention mechanism, and combines deep fusion of spatial-temporal features to improve recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0047] Figure 1 A schematic diagram of the structure of a lightweight gated fusion network model provided in an embodiment of the present application;

[0048] Figure 2 A schematic diagram of the structure of the temporal convolution layer module provided in an embodiment of the present application;

[0049] Figure 3 A schematic diagram of the structure of the temporal convolutional network residual block provided in an embodiment of the present application;

[0050] Figure 4 A schematic diagram of the structure of a lightweight signal modulation type identification device provided in an embodiment of the present application;

[0051] Figure 5 A schematic diagram of the structure of a lightweight signal modulation category identification device provided in an embodiment of the present application.

[0052] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of this application more clear, the technical solutions in this application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0054] The terms "first," "second," "third," "fourth," and so forth (if any) in the description and claims of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequential sequence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in sequences other than those illustrated or described herein.

[0055] In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described in this application as "exemplary" or "for example" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0056] The following specific embodiments describe in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0057] An embodiment of the present application provides a lightweight signal modulation category identification method, including: constructing and training a lightweight gated fusion network model, obtaining the IQ component of the modulation signal to be identified and inputting it into the trained lightweight gated fusion network model for analysis, and outputting the modulation identification type. Figure 1 A schematic diagram of the structure of a lightweight gated fusion network model provided in an embodiment of the present application is shown in FIG. Figure 1 As shown in the figure, the lightweight gated fusion network model includes a time domain feature extraction module, a spatial feature extraction module, a spatiotemporal feature fusion module, a time series modeling module and a classification and recognition module.

[0058] The training process of the lightweight gated fusion network model includes the following steps:

[0059] S1: Obtain a training data set, where each set of training samples in the training data set includes IQ components of a modulated signal and corresponding category labels.

[0060] It can be understood that in the wireless communication modulation recognition task, the core of the training data is a large-scale labeled dataset containing the IQ components of the modulation signal and their corresponding categories.

[0061] This paper uses the RML2016.10a public dataset, which is a benchmark dataset in the field of modulation recognition. It covers 11 typical digital modulation modes (including BPSK, QPSK, 8PSK, 16QAM, 64QAM, QAM16 / 64, BFSK, CPFSK, GFSK, PAM4, and WBFM). It contains a total of 1,000,000 samples, each of which is a complex signal r(n) of length 128, consisting of IQ components (real part I(n) and imaginary part Q(n)), that is, r(n) = I(n) + jQ(n).

[0062] First, the signal is normalized (mean = 0, variance = 1) to eliminate gain differences between different devices. Second, channel equalization is performed on the IQ components to compensate for channel distortion. Finally, the training set, validation set, and test set are randomly divided in a ratio of 3:1:1 to ensure that the distribution of each modulation category in each subset is consistent, avoiding performance degradation caused by data skew.

[0063] S2: Input the training sample into the spatial feature extraction module for spatial feature extraction to obtain the spatial features of the modulated signal, and input the training sample into the time domain feature extraction module for time domain feature extraction to obtain the time domain features of the modulated signal.

[0064] Among them, the spatial feature extraction module includes: the first depth-separable convolution layer and the first point-by-point convolution layer; the time domain feature extraction module includes: two temporal convolution layer modules, two dimensionality reshaping layers, and the first feature splicing layer.

[0065] The training samples are input into the spatial feature extraction module to extract spatial features and obtain the spatial features of the modulated signal, including:

[0066] Step 1: Input the I component and Q component of the modulated signal into the first depth-wise separable convolutional layer at the same time, perform spatial feature extraction on the modulated signal through a 2×8 convolution operation, and obtain the first spatial feature of the modulated signal.

[0067] It can be understood that the (2,8) depth-wise separable convolution kernel is used for spatial feature extraction. Compared with the traditional two-dimensional convolution, it decouples the convolution operation into two independent steps: spatial convolution and channel combination. The structure only requires 16 parameters ((2×8)×1 output channel) and can effectively capture multi-dimensional spatial patterns (such as time-frequency feature distribution). The rectangular kernel shape is suitable for processing communication features with direction sensitivity, focusing on spatial features in the vertical dimension and time series features in the horizontal dimension, which can achieve more fine-grained signal representation modeling.

[0068] Step 2: The first point-by-point convolution layer performs channel expansion on the first spatial feature through a 1×1 convolution operation to obtain the second spatial feature of the modulated signal.

[0069] It can be understood that the 1×1 convolution performs channel expansion and nonlinear mapping while keeping the spatial dimension unchanged. This layer projects the single-channel features output by the depth convolution into a 16-dimensional feature space, providing a rich feature representation basis for subsequent processing.

[0070] The training samples are input into the time domain feature extraction module to extract the time domain features and obtain the time domain features of the modulated signal, including:

[0071] Step 1: Input the I component and Q component of the modulated signal into two parallel temporal convolutional layer modules respectively for time domain feature extraction to obtain the time domain features of the I component and the time domain features of the Q component.

[0072] Figure 2 A schematic diagram of the structure of the temporal convolution layer module provided in the embodiment of the present application is shown in FIG. Figure 2 As shown in Figure 1, the temporal convolutional layer module includes: a first causal convolutional layer, a batch normalization layer, two parallel second causal convolutional layers and a Swish activation function.

[0073] The first causal convolution layer is a one-dimensional causal convolution with a kernel size of 8 and 16 filter channels. This causal convolution uses left padding to ensure that the output at each time point depends only on the current and historical input values, strictly preventing future information leakage. This makes it suitable for streaming signal processing scenarios and meets the timing dependency requirements of real-time communication systems. This layer contains 128 trainable parameters (8 × 1 input channels × 16 output channels), and the bias term is disabled to simplify the model.

[0074] A batch normalization layer is added after the first causal convolutional layer. The core function of the batch normalization layer is to alleviate the internal covariate shift problem during training by standardizing the distribution of activation values. This layer maintains four parameters for each feature channel: a scaling factor γ (16), an offset β (16), and a mean (16) and variance (16) that are dynamically updated during training. A total of 64 parameters can be folded and optimized during inference. This method enables the model to tolerate larger learning rates, significantly accelerating convergence while providing a slight regularization effect.

[0075] The temporal convolutional layer module consists of two parallel three-point causal convolution paths. The main path performs a linear transformation without an activation function; the gated path uses a Swish activation function to output a gating coefficient in the range [0, 1]. The two paths are fused through a multiplication operation: the main path features are element-wise multiplied by the gating coefficient. This design implements a dynamic feature selection mechanism: the gated path learns to suppress noise-dominated temporal segments while enhancing feature regions containing key patterns.

[0076] Step 2: Input the time domain features of the I component and the time domain features of the Q component into two parallel dimension reshaping layers respectively to reconstruct the tensor dimension.

[0077] Specifically, a tensor reshape operation (Reshape((1,128,16))) is used to reshape the tensor dimensions, converting the one-dimensional time series feature vector into a pseudo-two-dimensional structure. This transformation preserves all original information (a zero-parameter operation) and is essentially a virtual extension that adds a "height" dimension. Its core purpose is to convert the time series signal into an image-like structure, establishing the data foundation for the subsequent application of two-dimensional convolution operations in the joint time-space space, and opening up the possibility of feature interaction across time points.

[0078] Step 3: Perform dual-channel feature fusion along the height dimension on the reconstructed time domain features of the I component and the Q component through the first feature splicing layer to obtain fused time domain features.

[0079] Specifically, a concatenate operation is used to perform dual-channel feature fusion along the height dimension (axis = 1), concatenating the feature tensors of the two branches into a composite feature cube. This process does not introduce any parameters and maximizes the preservation of the differentiated information of the original branches. This design enables subsequent layers to explicitly learn the complementary correlations between the two-path signals (such as the phase constraints between the I / Q orthogonal components), capturing the joint statistical properties of the real and imaginary parts of the communication signal.

[0080] S3: Input the time domain features and spatial features into the spatiotemporal feature fusion module for feature fusion processing to obtain the first spatiotemporal fusion features.

[0081] Specifically, the time domain features and the spatial features are input into the spatiotemporal feature fusion module for feature fusion processing to obtain the first spatiotemporal fusion features, including the following steps:

[0082] S31: Input the fused time domain features into the second depth-wise separable convolutional layer through a 1×8 convolution operation to obtain the first time domain features.

[0083] It can be understood that the fused time domain features obtained in the above steps are input into the second depth-wise separable convolutional layer, and the time domain features capture the channel interaction in the time domain through the depth-wise separable convolutional layer.

[0084] S32: Input the first time domain feature into the second point-by-point convolution layer through a 1×1 convolution operation to obtain a second time domain feature.

[0085] It can be understood that the processed time domain features are obtained by fusing channel information through the point-by-point convolution layer in order to perform spatiotemporal feature fusion.

[0086] S33: Fusing the second temporal feature and the second spatial feature through a second feature concatenation layer to obtain a first spatiotemporal fusion feature.

[0087] Specifically, the spatial features and temporal features are merged into a 40-channel feature map through the feature splicing layer, and the spatial-temporal joint pattern is extracted by Conv2D. Its small convolution kernel (2×5) adapts to the spatial (2-way) and temporal (128-point) dimensions of the IQ components to avoid invalid calculations; the output is Reshaped to [batch_size, 124, 32] to retain key timing information.

[0088] S4: Input the spatiotemporal fusion features into the time series modeling module to enhance the time series features and obtain the second spatiotemporal fusion features.

[0089] Among them, the temporal modeling module includes: two temporal convolutional network residual blocks with the same structure. Figure 3 A schematic diagram of the structure of the temporal convolutional network residual block provided in the embodiment of the present application is shown in FIG. Figure 3 As shown in the figure, the temporal convolutional network residual block includes: a third depthwise separable convolutional layer, a third pointwise convolutional layer, a Swish activation function, a fourth depthwise separable convolutional layer, a fourth pointwise convolutional layer, and a ReLU activation function. The fourth depthwise separable convolutional layer introduces dilation = 2 to expand the receptive field (covering 5 time steps), the Swish function is used to alleviate gradient vanishing, and the ReLU function is used to enhance nonlinearity. The temporal convolutional network residual block uses residual connections.

[0090] Understandably, while spatiotemporal fusion features effectively integrate spatial dimension information with basic time series features, they still struggle to capture the long-range dynamic evolution patterns across longer time steps in complex communication scenarios. Therefore, the spatiotemporal feature fusion data obtained in the above steps is further subjected to time series modeling.

[0091] S5: Input the second spatiotemporal fusion feature into a classification and recognition module for classification and recognition to obtain a modulation recognition type.

[0092] Specifically, the classification and recognition module includes: a global average pooling layer, a fully connected layer, and an output layer. The second spatiotemporal fusion feature is input into the classification and recognition module for classification and recognition to obtain the modulation recognition type, including:

[0093] The second spatiotemporal fusion feature is input into the global average pooling layer for dimensionality reduction. The second spatiotemporal fusion feature after dimensionality reduction is input into the fully connected layer to extract high-level semantic features and obtain high-dimensional features. The output layer uses the Softmax function to process the high-dimensional features to obtain the modulation probabilities corresponding to different labels, and the label with the highest modulation probability is used as the final modulation recognition type.

[0094] S6: Iteratively update the parameters of the lightweight gated fusion network model through the back propagation algorithm according to the preset loss function until the loss function converges or the maximum number of iterations is reached.

[0095] In one embodiment, to verify the performance of the proposed method, the dataset of the present invention uses RML2016.10a as the training set, uses the Adam optimizer (initial learning rate lr = 0.001), and the loss function is cross entropy. The training strategy includes:

[0096] Learning rate decay: if the validation loss does not decrease for 10 consecutive rounds, then lr×0.8, and if it does not decrease for 60 consecutive rounds, then terminate early;

[0097] Batch training: The batch size is set to 128 to balance memory usage and gradient stability;

[0098] In order to verify the performance of the proposed algorithm, the lightweight gated fusion network model GateFusionNet designed in this paper is compared with PET-CGDNN, CLDNN2, LSTM and MCLDNN. The number of parameters, floating-point calculation amount and average recognition rate of the model are shown in Table 1.

[0099] Table 1 Comparison of model parameters and average recognition rates

[0100]

[0101] Experiments demonstrate that GateFusionNet offers significant performance advantages and potential for engineering applications. Compared to comparable models, this model achieves an average recognition rate of 61.9% with a significantly lower parameter count (23,795) and floating-point computational overhead (5.42M Flops). This not only surpasses traditional complex models (such as MCLDNN and CLDNN2, which have over 400,000 parameters) by two orders of magnitude in computational efficiency, but also surpasses these models in accuracy by approximately 1.6%-4.5%.

[0102] The embodiment of the present application provides a lightweight signal modulation category recognition method. The method constructs a lightweight gated fusion network model, which adopts a dual-branch asynchronous feature extraction architecture, in which the spatial feature branch adopts a deep separable convolution to capture the spatial distribution pattern of the signal; the time domain branch refines the dynamic features with a weight-sharing dual-path structure, and the extracted spatial features and time domain features are fused and compressed and input into the timing modeling module. The timing modeling module analyzes the long-range signal evolution law in an exponentially expanded receptive field through a deep separable dilated convolution residual block, so that the model can significantly reduce the hardware load while ensuring high-precision recognition, meeting the dual requirements of real-time and endurance in mobile scenarios; in systems that require distributed deployment, such as industrial Internet of Things and smart security, its lightweight characteristics can significantly reduce network transmission costs and storage overhead, and improve the overall deployment efficiency of the system.

[0103] Figure 4 This is a schematic diagram of the structure of the lightweight signal modulation category identification device provided in the embodiment of the present application. Figure 4 As shown, the lightweight signal modulation type identification device 400 provided in this embodiment includes:

[0104] A training module 401 is used to train a lightweight gated fusion network model;

[0105] An acquisition module 402 is configured to acquire an IQ component of a modulated signal to be identified;

[0106] The processing module 403 is used to input the IQ components of the modulated signal into the trained lightweight gated fusion network model for analysis and output the modulation identification type. The lightweight gated fusion network model includes a time domain feature extraction module, a spatial feature extraction module, a spatiotemporal feature fusion module, a time series modeling module and a classification and recognition module.

[0107] The lightweight signal modulation type identification device provided in this embodiment can execute the lightweight signal modulation type identification method provided in the above method embodiment. Its implementation principle and technical effects are similar and will not be described in detail in this embodiment.

[0108] Figure 5 This is a schematic diagram of the structure of a lightweight signal modulation category identification device provided in an embodiment of the present application. Figure 5 As shown, the lightweight signal modulation type identification device provided by the embodiment of the present application, the lightweight signal modulation type identification device 500 includes: a receiver 501, a transmitter 502, a processor 503 and a memory 504.

[0109] Receiver 501, for receiving instructions and data;

[0110] Transmitter 502, used to send instructions and data;

[0111] Memory 504, for storing computer-executable instructions;

[0112] The processor 503 is configured to execute the computer-executable instructions stored in the memory 504 to implement the various steps of the lightweight signal modulation type identification method in the above embodiment. For details, please refer to the relevant description in the above embodiment of the lightweight signal modulation type identification method.

[0113] Optionally, the memory 504 may be independent or integrated with the processor 503 .

[0114] When the memory 504 is independently provided, the electronic device further includes a bus for connecting the memory 504 and the processor 503 .

[0115] It will be appreciated by those skilled in the art that all or some of the steps, systems, and functional modules / units in the methods disclosed above may be implemented as software, firmware, hardware, and appropriate combinations thereof. In hardware implementations, the division between the functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed by several physical components in cooperation. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or temporary medium). As is well known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable, and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, it is well known to those skilled in the art that communication media generally embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0116] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of the present application and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0117] It should be understood that the present application is not limited to the exact structure described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A lightweight signal modulation category identification method, characterized in that: The method comprises: Construct and train a lightweight gated fusion network model, obtain the IQ component of the modulation signal to be identified and input it into the trained lightweight gated fusion network model for analysis, and output the modulation identification type. The lightweight gated fusion network model includes a time domain feature extraction module, a spatial feature extraction module, a spatiotemporal feature fusion module, a time series modeling module, and a classification and recognition module; The training process of the lightweight gated fusion network model includes the following steps: S1: Obtain a training data set, where each set of training samples in the training data set includes IQ components of a modulated signal and corresponding category labels; S2: Input the training sample into the spatial feature extraction module to extract the spatial features and obtain the spatial features of the modulated signal. Input the training sample into the time domain feature extraction module to extract the time domain features and obtain the time domain features of the modulated signal. S3: Input the temporal features and spatial features into the spatiotemporal feature fusion module for feature fusion processing to obtain the first spatiotemporal fusion feature; S4: Input the spatiotemporal fusion features into the time series modeling module to enhance the time series features and obtain the second spatiotemporal fusion features; S5: inputting the second spatiotemporal fusion feature into the classification and recognition module for classification and recognition to obtain a modulation recognition type; S6: Iteratively update the parameters of the lightweight gated fusion network model through the back propagation algorithm according to the preset loss function until the loss function converges or the maximum number of iterations is reached.

2. The method according to claim 1, characterized in that The time domain feature extraction module includes: two temporal convolution layer modules, two dimension reshaping layers, and a first feature splicing layer. The training sample is input into the time domain feature extraction module for time domain feature extraction to obtain the time domain features of the modulated signal, including: The I component and Q component of the modulated signal are respectively input into two parallel temporal convolutional layer modules for time domain feature extraction to obtain the time domain features of the I component and the Q component; The time domain features of the I component and the time domain features of the Q component are respectively input into two parallel dimension reshaping layers for tensor dimension reconstruction; The reconstructed time domain features of the I component and the Q component are fused along the height dimension using the first feature concatenation layer to obtain fused time domain features.

3. The method according to claim 2, characterized in that The temporal convolutional layer module includes: a first causal convolutional layer, a batch normalization layer, two parallel second causal convolutional layers and a Swish activation function.

4. The method according to claim 1, wherein The spatial feature extraction module includes: a first depth-wise separable convolution layer and a first point-by-point convolution layer. The training sample is input into the spatial feature extraction module to extract the spatial features and obtain the spatial features of the modulated signal, including: The I component and Q component of the modulated signal are simultaneously input into the first depth-wise separable convolutional layer to extract the spatial features of the modulated signal through a 2×8 convolution operation to obtain the first spatial feature of the modulated signal; The first point-by-point convolution layer performs channel expansion on the first spatial feature through a 1×1 convolution operation to obtain the second spatial feature of the modulated signal.

5. The method according to claim 2, characterized in that The step of inputting the time domain features and the spatial features into the spatiotemporal feature fusion module for feature fusion processing to obtain the first spatiotemporal fusion features includes: The fused time domain features are input into the second depth-wise separable convolutional layer through a 1×8 convolution operation to obtain the first time domain features; Input the first time domain feature into the second point-by-point convolution layer through a 1×1 convolution operation to obtain the second time domain feature; The second time domain feature and the second space feature are fused through the second feature splicing layer to obtain the first time-space fusion feature.

6. The method according to claim 1, characterized in that The temporal modeling module includes: two temporal convolutional network residual blocks with the same structure, the temporal convolutional network residual block includes: a third depth-separable convolutional layer, a third point-by-point convolutional layer, a Swish activation function, a fourth depth-separable convolutional layer, a fourth point-by-point convolutional layer and a ReLU activation function, and the temporal convolutional network residual block adopts a residual connection.

7. The method according to claim 1, characterized in that The classification and recognition module includes: a global average pooling layer, a fully connected layer and an output layer. The second spatiotemporal fusion feature is input into the classification and recognition module for classification and recognition to obtain a modulation recognition type, including: The second spatiotemporal fusion feature is input into the global average pooling layer for dimensionality reduction; The second spatiotemporal fusion feature after dimensionality reduction is input into the fully connected layer to extract high-level semantic features to obtain high-dimensional features; The output layer uses the Softmax function to process the high-dimensional features to obtain the modulation probability, and determines the modulation recognition type based on the modulation probability.

8. A lightweight signal modulation category identification device, characterized in that: The device comprises: Training module, used to train lightweight gated fusion network models; An acquisition module, configured to acquire the IQ components of the modulation signal to be identified; The processing module is used to input the IQ components of the modulated signal into a trained lightweight gated fusion network model for analysis and output the modulation identification type. The lightweight gated fusion network model includes a time domain feature extraction module, a spatial feature extraction module, a spatiotemporal feature fusion module, a time series modeling module and a classification and recognition module.

9. A lightweight signal modulation category identification device, characterized in that: The device comprises: Memory; processor; wherein the memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the lightweight signal modulation category identification method according to any one of claims 1 to 7.

10. A computer storage medium, characterized in that The computer storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the lightweight signal modulation category identification method according to any one of claims 1 to 7.

Images