Signal modulation identification method and system based on multi-scale convolution and multi-dimensional attention

Abstract

The invention belongs to the technical field of signal modulation recognition, and discloses a signal modulation recognition method and system based on multi-scale convolution and multi-dimensional attention, and the method comprises the steps: obtaining to-be-classified signal data, and carrying out the preprocessing of the to-be-classified signal data, and obtaining a first feature matrix graph; performing feature extraction of different scales on the first feature matrix graph, and splicing and fusing the extracted features along a channel dimension to obtain a second feature matrix graph; extracting global information of the second feature matrix graph in different directions and local information of the second feature matrix graph under different scales, and fusing the global information and the local information to obtain a third feature matrix graph; performing context sequential relation modeling on the third feature matrix graph to realize classification identification; and defining a loss function, optimizing model parameters, and obtaining a trained signal modulation identification model. According to the method, a residual multi-scale feature extraction module and a multi-dimensional attention module are designed, the feature extraction and fusion capability of the model is enhanced, and the accuracy of a classification task is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of signal modulation recognition, and in particular to a signal modulation recognition method and system based on multi-scale convolution and multi-dimensional attention. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Modulation recognition is one of the basic tasks of communication systems. It refers to determining the modulation method used by a signal under the premise of unknown modulation information content and parameters. With the rapid development of wireless communication technology, the communication environment has become more and more complex, and the number of signal types and modulation methods has increased, which has brought great challenges to signal modulation recognition. Traditional modulation recognition methods rely on prior knowledge and specific signal environments, and are difficult to adapt to complex and changeable communication environments. They have poor robustness and scalability.

[0004] With the development of artificial intelligence technology, deep learning technology has been widely used in many fields, such as natural language processing, computer vision, and sentiment analysis. Compared with traditional data analysis and processing methods, deep learning methods can automatically learn discriminative features without manually selecting features, so they are gradually applied to the research of modulation recognition in the communication field. For example, there are studies based on 2D convolutional neural networks (CNN) to identify modulation modes. Convolutional neural networks can automatically learn discriminative features from input signals for classification. However, although convolutional neural networks have strong local feature extraction capabilities, they are limited by the size of their convolution kernels and the limited scope of operations. They are deficient in extracting global features and long-term dependencies. In order to extract features on a time scale, there are currently methods for modulation recognition using 3D CNN. However, 3D CNN models have the problems of large model size and large amount of computation, making them difficult to deploy in real-time detection applications. Moreover, we used IQ (in-phase and quadrature) technology when actually transmitting signals. The information contained in the two IQ channels can also reflect the unique characteristics of the transmitted signals to a certain extent. Simply passing the input signal through the convolutional neural network ignores the actual signal transmission situation.

[0005] On the other hand, the electromagnetic environment is complex and changeable, and various modulated signals are inevitably interfered by various noises during transmission, resulting in differences between the actual received signal and the ideal signal. At low signal-to-noise ratios, the recognition accuracy of modulation classification is very low. When the noisy signal is sent to the deep neural network classifier to extract features, the features extracted for each modulation method are likely to be affected by the randomness of the noise. This randomness reduces the stability of the features extracted for each modulation method on the one hand, and causes confusion between the features extracted for different modulation methods on the other hand, ultimately leading to incorrect classification results.

[0006] In summary, the existing automatic modulation recognition technology still has many problems: the modulation model feature expression ability is limited, it is difficult to extract long-term signal features, and it is difficult to improve the recognition accuracy when the noise cannot be ignored, that is, when the signal-to-noise ratio is low. Summary of the invention

[0007] In order to solve the above problems, the present invention proposes a signal modulation recognition method and system based on multi-scale convolution and multi-dimensional attention, designs an end-to-end model for modulation recognition tasks, adds RNN architecture to CNN architecture, introduces multi-dimensional attention module, and enhances the model's ability to extract global and local features. At the same time, a residual multi-scale feature extraction module is designed to extract and fuse multi-scale features to improve the accuracy of classification tasks.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides a signal modulation recognition method based on multi-scale convolution and multi-dimensional attention, comprising the following steps:

[0010] Acquire the signal data to be classified and perform preprocessing to obtain a first feature matrix diagram;

[0011] Extract features of different scales from the first feature matrix, splice and fuse the extracted features along the channel dimension, and obtain a second feature matrix;

[0012] Extracting the global information of the second feature matrix in different directions and the local information at different scales, fusing the global information and the local information to obtain a third feature matrix;

[0013] Model the contextual temporal relationship of the third feature matrix diagram to achieve classification and recognition;

[0014] Define the loss function, optimize the model parameters, and obtain the trained signal modulation recognition model.

[0015] As an optional implementation, the signal data to be classified is signal data under different signal-to-noise ratios modulated by different modulation methods. Preliminary features of the signal data to be classified are extracted by 1×3 convolution and the number of channels is increased to obtain a first feature matrix diagram.

[0016] As an optional implementation, the training of the model is supervised by the joint loss function of Center loss and double-layer Softmax loss, which is expressed as follows:

[0017] L total =L S +λ1L R +λ2L C

[0018] Among them, L S represents Softmax loss, L C represents Center loss, L R represents the loss of classifying modulation types from a coarse granularity, and λ1 and λ2 are adjustable parameters.

[0019] In a second aspect, the present invention provides a signal modulation recognition system based on multi-scale convolution and multi-dimensional attention, comprising:

[0020] The data acquisition and preprocessing module is configured to: acquire the signal data to be classified and perform preprocessing to obtain a first feature matrix diagram;

[0021] The residual multi-scale feature extraction module is configured to: extract features of different scales from the first feature matrix image, splice and fuse the extracted features along the channel dimension, and obtain a second feature matrix image;

[0022] The multi-dimensional attention module is configured to: extract the global information of the second feature matrix in different directions and the local information at different scales, and fuse the global information with the local information to obtain a third feature matrix;

[0023] The classification recognition module is configured to: perform context temporal relationship modeling on the third feature matrix diagram to achieve classification recognition;

[0024] The model training module is configured to: define a loss function, optimize model parameters, and obtain a trained signal modulation recognition model.

[0025] As an optional implementation, the residual multi-scale feature extraction modules are provided in plurality, the plurality of residual multi-scale feature extraction modules are connected in series, and feature extraction is performed sequentially through the plurality of residual multi-scale feature extraction modules in series.

[0026] As an optional implementation, the multi-dimensional attention module adopts group learning and multi-branch parallel methods to capture feature information in different dimensions and scales.

[0027] As an optional implementation, the multidimensional attention module includes four branches, wherein the first branch and the second branch perform horizontal pooling and vertical pooling respectively to extract global information in different directions, and the third branch and the fourth branch perform ordinary convolution and dilated convolution respectively to extract different local information.

[0028] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in the first aspect is performed.

[0029] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the method described in the first aspect is performed.

[0030] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention proposes a signal modulation recognition method and system based on multi-scale convolution and multi-dimensional attention, takes the sampling values ​​of IQ dual channels as input, pays attention to multi-scale features, and designs a modulation recognition model using multi-scale features. The residual multi-scale feature extraction module proposed in the present invention enhances the robustness of the network through multi-scale feature expression, optimizes features from the element-level granularity, improves the network's ability to extract discriminative features, and thus effectively improves the recognition accuracy. The present invention adopts a plurality of residual multi-scale feature extraction modules to be connected in series, and sequentially extracts features through a plurality of stacked residual multi-scale feature extraction modules, and a pooling layer is added after each feature extraction module, thereby realizing the lightweight of the model; the multi-dimensional attention module proposed in the present invention captures and fuses global information by performing pooling in different dimensions and performing 1×1 convolution, and at the same time introduces 3×3 convolution and hole convolution to capture local spatial features, thereby enhancing the model's information capture capability at the global and local levels, and further improving the accuracy of modulation recognition; the present invention adopts a loss function combining two-layer cross entropy loss and center loss to supervise the training of the modulation pattern recognition network in the classification output stage, classifies the recognition tasks from coarse and fine granularity and improves the feature discrimination capability, and by introducing the concept of feature center, makes the features of the same category more compact and the features of different categories more separated.

[0033] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0035] Figure 1 A data processing flow chart of a signal modulation recognition method based on multi-scale convolution and multi-dimensional attention provided in Example 1 of the present invention;

[0036] Figure 2 A framework diagram of a signal modulation recognition method based on multi-scale convolution and multi-dimensional attention provided in Example 1 of the present invention;

[0037] Figure 3 Schematic diagram of the structure of the residual multi-scale feature extraction module of the present invention;

[0038] Figure 4 Schematic diagram of the structure of the multi-dimensional attention module of the present invention;

[0039] Figure 5 It is a confusion matrix diagram of the method of the present invention on the test set of the RML2016.10a data set;

[0040] Figure 6 Figure 3 is a confusion matrix diagram of the method of the present invention under different signal-to-noise ratio conditions on the test set of the RML2016.10a dataset, wherein (a) is a confusion matrix diagram of the method of the present invention under high signal-to-noise ratio conditions on the test set of the RML2016.10a dataset, (b) is a confusion matrix diagram of the method of the present invention under the condition of a signal-to-noise ratio of -4 on the test set of the RML2016.10a dataset, and (c) is a confusion matrix diagram of the method of the present invention under the condition of a signal-to-noise ratio of 2 on the test set of the RML2016.10a dataset. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0042] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0043] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0044] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0045] Terminology explanation:

[0046] 1. Gated Recurrent Unit (GRU): It uses "update gate" and "reset gate" to transfer information between sequences, which solves the problem of gradient disappearance in standard recurrent neural networks and makes training faster.

[0047] 2. Centre Loss: It is a loss function used in deep learning, which aims to improve the discriminative ability of features. Centre loss introduces the concept of feature center, making features of the same category more compact and features of different categories more separated, thereby improving the classification performance of the model.

[0048] 3. RML2016.10a: Radio modulation signal dataset, currently the most widely used dataset in the field of modulation recognition. The sampling length of each frame signal in the dataset is 128, and it contains two dimensions, namely the real part and the imaginary part of the IQ sample, which are represented as a 2×128 matrix. The signal-to-noise ratio is distributed in the range of -18dB to 20dB, with a step size of 2dB, and the samples are evenly distributed at each signal-to-noise ratio. This public dataset contains 11 different types of modulation signals, namely 8PSK, AM-DSB, AM-SSB, WBFM, BPSK, CPFSK, QPSK, GFSK, 4-PAM, 16-QAM, 64-QAM, and the total number of samples is 220,000.

[0049] Example 1

[0050] like Figure 1-2 As shown, this embodiment provides a signal modulation recognition method based on multi-scale convolution and multi-dimensional attention, comprising the following steps:

[0051] S1, obtaining the signal data to be classified and preprocessing it to obtain a first feature matrix diagram;

[0052] S2, extracting features of different scales from the first feature matrix, splicing and fusing the extracted features along the channel dimension, and obtaining a second feature matrix;

[0053] S3, extracting the global information of the second feature matrix in different directions and the local information at different scales, fusing the global information and the local information to obtain a third feature matrix;

[0054] S4, modeling the contextual temporal relationship of the third feature matrix diagram to achieve classification and recognition;

[0055] S5. Define the loss function, optimize the model parameters, and obtain the trained signal modulation recognition model.

[0056] Since the IQ signal can reflect the characteristics of the signal in actual transmission, the original IQ signal is selected as the input to the model. The present invention also adopts a residual multi-scale feature extraction module, pays attention to the multi-scale characteristics of the modulation recognition mode, adaptively learns the multi-scale features at the element-level granularity, and improves the discrimination ability of the network. In addition, the present invention adopts a multi-dimensional attention module to optimize and improve the features learned from the residual multi-scale feature extraction module. Among them, the multi-dimensional attention module adopts group learning and multi-branch parallel methods to capture feature information on different dimensions and scales. In addition, in the classification stage, two layers of Softmax loss and Centre loss are used to jointly supervise the training of the iterative model. The classification performance of the model is further improved by optimizing and increasing the inter-class distance and reducing the intra-class distance at both coarse and fine granularity.

[0057] First, a signal data set under different signal-to-noise ratios modulated by different modulation modes is obtained and preprocessed, and the preprocessed signal data set (first feature matrix diagram) is divided into a training set, a test set, and a validation set.

[0058] Preprocess the signal to be classified, specifically:

[0059] The acquired signal data set to be classified was normalized and a dimension was added to make it meet the input format requirements of the model. The preliminary features of the signal data to be classified were extracted through 1×3 convolution and the number of channels was increased to obtain the first feature matrix diagram.

[0060] The overall network structure of the present invention is as follows Figure 2As shown in the figure. First, the original IQ data is passed through a 1×3 convolutional layer to extract preliminary features and increase the number of channels, and then sent to the subsequent stacked residual multi-scale feature extraction module. Among them, the stacked residual multi-scale feature extraction modules are connected in series. Next, a multi-dimensional attention module is used to optimize the features. Then it is sent to a recurrent neural network with 128 GRU units for temporal context modeling. Finally, classification is achieved using a fully connected layer and a Softmax function. In the model training stage, the Center loss and double-layer Softmax loss joint loss function are used to supervise the training of the model.

[0061] (1) The residual multi-scale feature extraction module is used to extract the multi-scale features of the input feature matrix.

[0062] The residual multi-scale feature extraction module extracts features of different scales from the input features through three parallel branches. The extracted features are activated by the activation function and concatenated and aggregated along the channel dimension. Feature extraction is performed sequentially by stacking multiple residual multi-scale feature extraction modules.

[0063] The residual multi-scale feature extraction module of the present invention is shown in Figure 3 The specific process of data processing in this module is as follows:

[0064] a) Through three branches consisting of convolutional layers with convolution kernel sizes of 3×3, 3×5, and 3×7 and batch normalization layers, features of different scales are extracted in parallel from the input features.

[0065] b) After each convolution block is batch normalized, PReLU is used as the activation function to activate the convolution layer and adaptively optimize the multi-scale features.

[0066] It should be noted that the residual multi-scale feature extraction module contains a Max-pooling branch for multi-scale information extraction, which is connected after each residual multi-scale feature module.

[0067] In a specific implementation, PReLU is an adaptive element-level attention module that can improve the model's fitting ability and reduce the risk of overfitting without adding any additional parameters. PReLU can reduce the "dead neuron" problem by learning α, and its mathematical representation is as follows:

[0068]

[0069] Here, x is an element in the input feature map.

[0070] PReLU also needs to update its own weight W in the Backward phase. Assuming that the loss function defined by the model training is E, to update w cFor example, the gradient descent algorithm will make the following transformations:

[0071] c=0,1,2…,C i -1; where η represents the learning rate, is the required gradient, which is calculated according to the chain rule as follows: C=0,1,2…,C i -1; n represents the total number of layers in the model, t represents the current layer, as long as the calculation The weights of the PReLU layer will get updated.

[0072] c) The features extracted from different branches are padded with zeros to the same size, and then the optimized features extracted in parallel are concatenated along the channel dimension and fused through 1×1 convolution as the output of the residual multi-scale feature extraction module.

[0073] The output of the last residual multi-scale feature extraction module (the second feature matrix diagram) is fed into the multi-dimensional attention module.

[0074] (2) The multi-dimensional attention mechanism is used to capture the part of the optimized feature map that can better represent the discriminative features of the modulation type.

[0075] The present invention optimizes the local features and global features of features in parallel through a multi-dimensional attention module. The multi-dimensional attention module includes four branches, two of which extract global information in different directions of the input feature map by performing pooling operations on different dimensions. The other two branches perform ordinary convolution and dilated convolution respectively to extract local information of different scales. The module also groups channels and enhances the diversity of features through grouped convolution.

[0076] The specific process of data processing by the multi-dimensional attention module of the present invention is as follows: Figure 4 Shown are:

[0077] a) The features after the residual multi-scale feature extraction module are passed through the four branches of the module in parallel to obtain an optimized feature matrix. In a specific implementation, we assume that the input feature Where B is the batch size, C is the number of channels, and the size of the feature map is H×W, which are the height and width of the feature map. First, the input feature map is grouped by channel, and each group is processed independently. Here we choose the number of groups to be 8, and merge the batch dimension and the group dimension so that subsequent operations can process all batches and all groups at the same time. At this time, the shape of the feature map is: Where G=8.

[0078] b) First, the first branch and the second branch perform horizontal pooling and vertical pooling respectively. At the same time, we adjust the dimension order of the feature map of the second branch to adapt to the subsequent operations. The shapes of the feature maps obtained are as well as Then the two feature maps are concatenated along the second dimension and fused with a 1×1 convolution. The shape of the feature map is Through such a fusion operation, the dependency between height and width features can be captured and the expressiveness of the features can be enhanced. We then separate the fused features into height and width features again, but the features at this time already contain the dependency between height and width. This design allows the model to better utilize spatial information in subsequent operations. The separated height and width features are multiplied by the sigmoid function respectively, and then multiplied with the original feature map to obtain a weighted new feature map, which is named F1.

[0079] c) The last two branches perform ordinary convolution and dilated convolution respectively to extract different local information. The height and width of the feature map are kept consistent by padding, that is, the shape of the feature map is: After splicing along the channel dimension, they are fused through 1×1 convolution and the number of channels is adjusted again to remain unchanged, that is, C / / G. The feature map obtained at this time is named F2.

[0080] d) The global information extracted in b) and the local information extracted in c) are fused by matrix multiplication to obtain a third feature matrix diagram. The specific operation is as follows:

[0081] Reshape F1 and F2 into a shape suitable for matrix multiplication, that is, flatten the spatial dimensions of the feature map. The shape of the feature map obtained at this time is Represents all the information of the feature map at each channel and spatial position. The new tensor obtained after this operation is named F 11 and F 21 .

[0082] Global average pooling is performed on both F1 and F2, that is, the spatial dimension (H, W) is compressed to (1, 1). The result is two tensors with the shape of (B*G, C / / G, 1, 1). The shape of the tensor is then adjusted to (B*G, 1, C / / G) by reshaping and transposing to ensure that the subsequent matrix multiplication is performed correctly. The new tensor obtained after this operation is named F 12 and F 22 . Apply the softmax activation function to the two newly obtained tensors to obtain a standardized probability distribution. This step ensures that when each channel is weighted, the sum of the weights between different channels is 1, resulting in a normalized coefficient. Then, matrix multiplication is used to calculate F 11and F with softmax activation function applied 22 、F 21 and F with softmax activation function applied 12 Perform matrix multiplication and get the final weighting coefficients.

[0083] (3) The GRU recurrent neural network is used to model the temporal relationship of the context.

[0084] Finally, a recurrent neural network is added to obtain a signal modulation recognition network. The recurrent neural network is composed of gated recurrent units and is used to model the contextual temporal relationship of the input features (the third feature matrix diagram). In a specific embodiment, the signal data is a time series, and the contextual temporal relationship of the input features is modeled by a recurrent neural network, so that the model can fully understand the discriminative features of each modulation mode in combination with the contextual information.

[0085] The features optimized by the attention module are sent to the recurrent neural network composed of GRU units for temporal relationship modeling, as shown below:

[0086] where x t is the input at time t, is the output of the hidden layer at the previous moment. Then we get the output H = [j1,…,h T ], and finally the classification task is realized by the Softmax activation function through the fully connected layer.

[0087] (4) The modulation recognition network is trained using the training set to obtain a trained modulation recognition network model.

[0088] The training of the model is supervised by the joint loss function of Center loss and double-layer Softmax loss, which is expressed as follows:

[0089] L total =L S +λ1L R +λ2L C

[0090] In this formula, L S stands for Softmax Loss, which is the initial loss function for classification and recognition tasks. R Represents the classification of the modulation type from a coarse granularity, that is, whether the modulation type belongs to analog modulation or digital modulation, L C represents the center loss, λ1 and λ2 are adjustable parameters, L S The representation is as follows:

[0091] In this formula, represents the i-th feature map, d is its dimension, is the weight matrix of the fully connected layer, is the weight bias of the fully connected layer, y i ∈{1,2,...,n} represents the predicted category, n refers to the number of categories, and m refers to the batch size. Refers to the j-th column of the weight matrix W.

[0092] L C The representation is as follows:

[0093] In this formula, represents the modulation type at the center of the discriminant hyperplane, c yi Initialized randomly and updated as the model training progresses.

[0094] The present invention builds a modulation recognition network based on multi-scale convolution and multi-dimensional attention. The network takes the original IQ data as input. In order to achieve effective multi-scale feature extraction, the present invention proposes a residual multi-scale feature extraction module, which can adaptively learn feature expression at multiple scales. At the same time, in order to fully capture local information and global information, the present invention proposes a multi-dimensional attention mechanism. In addition, the training of the model is supervised by a loss function combining two-layer Softmax loss and Centre loss to enhance the model's discrimination ability.

[0095] In order to verify the performance of the modulation recognition method of the present invention, the present invention conducts experiments on the RML2016.10a dataset based on the Pytorch framework. The confusion matrix on this dataset is as follows: Figure 5 As shown in the figure, the overall accuracy of the model on this dataset reached 63.64%, which is mainly limited by the fact that the signal is submerged by the noise at low signal-to-noise ratio. The confusion matrix of this model on this dataset at high signal-to-noise ratio (14dB in the figure) and when the signal-to-noise ratio is equivalent (-4dB and 2dB in the figure) is as follows Figure 6 (a) Figure 6 (b) Figure 6 (c) as shown.

[0096] Experiments show that the method of the present invention has excellent performance, and the accuracy is further improved by extracting features of different scales and grasping local features and global features.

[0097] Example 2

[0098] This embodiment provides a signal modulation recognition system based on multi-scale convolution and multi-dimensional attention, including:

[0099] The data acquisition and preprocessing module is configured to: acquire the signal data to be classified and perform preprocessing to obtain a first feature matrix diagram;

[0100] The residual multi-scale feature extraction module is configured to: extract features of different scales from the first feature matrix image, splice and fuse the extracted features along the channel dimension, and obtain a second feature matrix image;

[0101] The multi-dimensional attention module is configured to: extract the global information of the second feature matrix in different directions and the local information at different scales, and fuse the global information with the local information to obtain a third feature matrix;

[0102] The classification recognition module is configured to: perform context temporal relationship modeling on the third feature matrix diagram to achieve classification recognition;

[0103] The model training module is configured to: define a loss function, optimize model parameters, and obtain a trained signal modulation recognition model.

[0104] As an optional implementation, the residual multi-scale feature extraction modules are provided in plurality, the plurality of residual multi-scale feature extraction modules are connected in series, and feature extraction is performed sequentially through the plurality of residual multi-scale feature extraction modules in series.

[0105] As an optional implementation, the multi-dimensional attention module adopts group learning and multi-branch parallel methods to capture feature information in different dimensions and scales.

[0106] As an optional implementation, the multidimensional attention module includes four branches, wherein the first branch and the second branch perform horizontal pooling and vertical pooling respectively to extract global information in different directions, and the third branch and the fourth branch perform ordinary convolution and dilated convolution respectively to extract different local information.

[0107] It should be noted that the above modules correspond to the steps described in Example 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above Example 1. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer executable instructions.

[0108] In further embodiments, there is also provided:

[0109] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, it will not be described in detail here.

[0110] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0111] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0112] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method described in Example 1 is completed.

[0113] The method in Example 1 can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it is not described in detail here.

[0114] A computer program product includes a computer program, and when the computer program is executed by a processor, the method described in embodiment 1 is implemented.

[0115] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer executable instructions, such as instructions included in a program module, which are executed in a device on a real or virtual processor of the target to perform the process / method as described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functions of program modules can be combined or divided between program modules as needed. Machine executable instructions for program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0116] The computer program code for implementing the method of the present invention can be written in one or more programming languages. These computer program codes can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the computer or other programmable data processing device, causes the function / operation specified in the flow chart and / or block diagram to be implemented. The program code can be executed completely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer or completely on a remote computer or server.

[0117] In the context of the present invention, computer program codes or related data may be carried by any appropriate carrier to enable a device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, and the like. Examples of signals may include electrical, optical, radio, acoustic or other forms of propagation signals, such as carrier waves, infrared signals, and the like.

[0118] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0119] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A signal modulation recognition method based on multi-scale convolution and multi-dimensional attention, characterized in that: The following steps are involved: Acquire the signal data to be classified and perform preprocessing to obtain a first feature matrix diagram; Extract features of different scales from the first feature matrix, splice and fuse the extracted features along the channel dimension, and obtain a second feature matrix; Extracting the global information of the second feature matrix in different directions and the local information at different scales, fusing the global information and the local information to obtain a third feature matrix; Model the contextual temporal relationship of the third feature matrix diagram to achieve classification and recognition; Define the loss function, optimize the model parameters, and obtain the trained signal modulation recognition model.

2. The signal modulation recognition method based on multi-scale convolution and multi-dimensional attention as claimed in claim 1, characterized in that: The signal data to be classified is signal data under different signal-to-noise ratios modulated by different modulation modes. Preliminary features of the signal data to be classified are extracted through 1×3 convolution and the number of channels is increased to obtain a first feature matrix diagram.

3. The signal modulation recognition method based on multi-scale convolution and multi-dimensional attention as claimed in claim 1, characterized in that: The training of the model is supervised by the joint loss function of Center loss and double-layer Softmax loss, which is expressed as follows: L total =L S +λ1L R +λ2L C Among them, L S represents Softmax loss, L C represents Center loss, L R represents the loss of classifying modulation types from a coarse granularity, and λ1 and λ2 are adjustable parameters.

4. Signal modulation recognition system based on multi-scale convolution and multi-dimensional attention, characterized by: include: The data acquisition and preprocessing module is configured to: acquire the signal data to be classified and perform preprocessing to obtain a first feature matrix diagram; The residual multi-scale feature extraction module is configured to: extract features of different scales from the first feature matrix image, splice and fuse the extracted features along the channel dimension, and obtain a second feature matrix image; The multi-dimensional attention module is configured to: extract the global information of the second feature matrix in different directions and the local information at different scales, and fuse the global information with the local information to obtain a third feature matrix; The classification recognition module is configured to: perform context temporal relationship modeling on the third feature matrix diagram to achieve classification recognition; The model training module is configured to: define a loss function, optimize model parameters, and obtain a trained signal modulation recognition model.

5. The signal modulation recognition system based on multi-scale convolution and multi-dimensional attention as claimed in claim 4, characterized in that: There are multiple residual multi-scale feature extraction modules, and the multiple residual multi-scale feature extraction modules are connected in series, and feature extraction is performed sequentially through the multiple serial residual multi-scale feature extraction modules.

6. The signal modulation recognition system based on multi-scale convolution and multi-dimensional attention as claimed in claim 4, characterized in that: The multi-dimensional attention module adopts group learning and multi-branch parallel methods to capture feature information in different dimensions and scales.

7. The signal modulation recognition system based on multi-scale convolution and multi-dimensional attention as claimed in claim 6, characterized in that: The multidimensional attention module includes four branches, wherein the first branch and the second branch perform horizontal pooling and vertical pooling respectively to extract global information in different directions, and the third branch and the fourth branch perform ordinary convolution and dilated convolution respectively to extract different local information.

8. An electronic device, characterized in that: The invention comprises a memory and a processor and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 3 is completed.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1 to 3.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 3.

Images