Method, device and equipment for classifying modulation signals

Abstract

The present invention provides a method, apparatus, and device for classifying modulated signals. The method comprises: obtaining a modulated signal to be classified; inputting the modulated signal to be classified into a pre-trained modulated signal classification model to obtain a modulated signal classification result; a pre-trained Mamba model in the pre-trained modulated signal classification model can effectively extract and process features of the modulated signal to be classified by combining features extracted by a selective state space (SSM) and a Mamba convolution block, thereby improving the accuracy of the pre-trained modulated signal classification model in recognizing the modulated signal to be classified; in addition, by designing a specific loss function, the initial MCLDNN model and the initial Mamba model can learn from each other during normal training without introducing additional perturbations to the data, thereby improving the performance of the pre-trained modulated signal classification model. Therefore, the method of the present invention improves the utilization rate of computing resources and the generalization ability of the pre-trained modulated signal classification model.

Description

Technical Field

[0001] The present invention relates to the technical field of modulation signal processing, and in particular to a method, device and equipment for classifying modulation signals. Background Art

[0002] With the rapid development of information technology, wireless communication systems are placing increasing demands on the utilization and management of spectrum resources. Full-wireless spectrum sensing, an advanced technology based on software-defined radio, can monitor the wireless spectrum in real time and assess channel quality and signal strength, providing strong support for the efficient operation of wireless communications. However, modulation signal classification in wireless communications, a critical step in ensuring transmission quality and stability, faces challenges in complex and changing channel environments and noise interference. Traditional modulation signal recognition methods rely primarily on manually selected features and decision boundaries. This approach is complex and inefficient, and its effectiveness is significantly reduced in harsh channel environments. Therefore, exploring an efficient, intelligent, and reliable automatic modulation recognition technology to address the complex challenges of wireless communications has become an urgent research need.

[0003] To address these issues, researchers have proposed a variety of technical solutions. Among them, models based on deep neural networks (DNNs) have achieved some success in classifying modulated signals by leveraging their ability to extract complex features in high-dimensional space. Furthermore, long-short-term memory networks (LSTMs) have also been applied to automatic modulation recognition due to their ability to capture signal time series characteristics. To combine the advantages of CNNs in local feature extraction with the LSTMs' ability to model long-term dependencies in sequence data, researchers have also constructed a CLDNN architecture, achieving a balance between performance and complexity, further improving modulation recognition performance.

[0004] While existing technologies have improved the accuracy of modulation recognition to a certain extent, they still have many shortcomings. While DNN-based models perform well under specific modulation schemes, their performance degrades significantly in low signal-to-noise ratio environments. Their dependence on expert features also limits their widespread application. While LSTM-based models can capture time series features, their ability to mine features of complex modulated signals is limited. Their model structure is simplistic, and their adaptability to different input forms is limited. Their performance is also limited in low signal-to-noise ratio environments. While the CLDNN model combines the advantages of CNN and LSTM, it has limitations in its model fusion strategy, with significant issues in computational resource utilization and training efficiency. It lacks feature mining for signals with special structures or high-dimensional modulation, and its fusion approach is inflexible. Especially on constrained devices, computational resource utilization is inefficient, training time is long, and the model is significantly affected by data. These issues hinder the further development of existing technologies in the field of automatic modulation recognition. Therefore, a more efficient, intelligent, and adaptable automatic modulation recognition technology is urgently needed to address the complex challenges in wireless communications. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a method, device and equipment for classifying modulated signals.

[0006] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] In a first aspect, the present invention provides a method for classifying a modulated signal, comprising:

[0008] Obtaining a modulated signal to be classified;

[0009] The modulated signal to be classified is input into the pre-trained modulated signal classification model to obtain the modulated signal classification result; the pre-trained modulated signal classification model is set up with a parallel connected pre-trained Mamba model and a pre-trained MCLDNN model; the pre-trained Mamba model is constructed based on the selective state space model, and the Mamba timing layer is introduced on the basis of the Mamba convolution block. The Mamba timing layer combines the selective SSM mechanism and the MLP structure to form a homogeneous stacking module; the pre-trained MCLDNN model includes multiple MCLDNN convolution blocks and timing layers. The MCLDNN convolution block is used to extract local features, and the timing layer is used to capture long sequence dependencies; the pre-trained modulated signal classification model adopts modulated signal samples and a preset loss function, and is constructed based on the mutual learning strategy between the initial Mamba model and the initial MCLDNN model.

[0010] Optionally, the pre-trained MCLDNN model includes: a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a fifth convolution block, a first sequential layer, a second sequential layer, a first fully connected layer, and a second fully connected layer;

[0011] The first convolution block, the second convolution block and the third convolution block are connected in parallel, the outputs of the second convolution block and the third convolution block are connected in series with the fourth convolution block, the output of the first convolution block is connected in parallel with the output of the fourth convolution block, the first convolution block and the fourth convolution block are connected in series with the fifth convolution block, and the fifth convolution block is connected in series with the first sequential layer, the second sequential layer, the first fully connected layer and the second fully connected layer.

[0012] Optionally, the pre-trained Mamba model includes: a sixth convolution block, a seventh convolution block, an eighth convolution block, a ninth convolution block, a tenth convolution block, a first Mamba timing layer, a second Mamba timing layer, a third fully connected layer, and a fourth fully connected layer;

[0013] The sixth convolution block, the seventh convolution block, and the eighth convolution block are connected in parallel, the outputs of the seventh and eighth convolution blocks are connected in series with the ninth convolution block, the output of the sixth convolution block is connected in parallel with the output of the ninth convolution block, the sixth and ninth convolution blocks are connected in series with the tenth convolution block, and the tenth convolution block is sequentially connected in series with the first Mamba timing layer, the second Mamba timing layer, the third fully connected layer, and the fourth fully connected layer.

[0014] Optionally, the output dimensions of the first fully connected layer and the third fully connected layer are both 128, and the output dimensions of the second fully connected layer and the fourth fully connected layer are both 11; and the optimizers of the MCLDNN model and the Mamba model both use Adam.

[0015] Optionally, the preset loss function is expressed as:

[0016]

[0017] Among them, L1 represents the preset loss function, CrossEntropyLoss represents the cross entropy loss function, Represents the modulated signal sample X label Input the prediction score of the initial MCLDNN model, Represents the modulated signal sample X label Enter the prediction score of the initial Mamba model, Represents the modulated signal sample X label The true label, Represents the one-hot encoding converted based on the prediction score of the initial MCLDNN model, Represents the one-hot encoding transformed according to the prediction score of the initial Mamba model.

[0018] Optionally, the training process of the pre-trained modulation signal classification model includes:

[0019] Obtaining a modulated signal sample;

[0020] Input the modulated signal samples into the initial modulated signal classification model for iterative training;

[0021] The initial modulation signal classification model that meets the preset iteration stop condition is used as the pre-trained modulation signal classification model;

[0022] The preset iteration stopping conditions include: the number of iterations is greater than the preset iteration threshold or the preset loss function continues to converge.

[0023] Optionally, the pre-trained Mamba model is used to process the modulated signal to be classified in a low signal-to-noise ratio environment, and the nonlinear expression ability and dynamic information filtering effect of the pre-trained modulated signal classification model are enhanced through a selective SSM mechanism.

[0024] In a second aspect, the present invention provides a modulation signal classification device, comprising:

[0025] The acquisition unit is used to: acquire the modulated signal to be classified;

[0026] The classification unit is used to: input the modulated signal to be classified into a pre-trained modulated signal classification model to obtain the modulation signal classification result; the pre-trained modulated signal classification model is provided with a parallel connected pre-trained Mamba model and a pre-trained MCLDNN model; the pre-trained Mamba model is constructed based on the selective state space model, and the Mamba timing layer is introduced on the basis of the Mamba convolution block. The Mamba timing layer combines the selective SSM mechanism and the MLP structure to form a homogeneous stacking module; the pre-trained MCLDNN model includes multiple MCLDNN convolution blocks and timing layers. The MCLDNN convolution block is used to extract local features, and the timing layer is used to capture long sequence dependencies; the pre-trained modulated signal classification model adopts the modulated signal samples and the preset loss function, and is constructed based on the mutual learning strategy between the initial Mamba model and the initial MCLDNN model.

[0027] In a third aspect, the present invention provides a device for classifying modulated signals, comprising: a processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor. When the device for classifying modulated signals is running, the processor and the storage medium communicate through the bus, and the processor executes the machine-readable instructions to perform the steps of the method for classifying modulated signals as described in the first aspect above.

[0028] The present invention provides a method, device, and apparatus for classifying modulated signals. A method for classifying modulated signals includes: obtaining a modulated signal to be classified; inputting the modulated signal to be classified into a pre-trained modulated signal classification model to obtain a modulation signal classification result; the pre-trained modulated signal classification model is provided with a pre-trained Mamba model and a pre-trained MCLDNN model connected in parallel; the pre-trained Mamba model is constructed based on a selective state space model, and a Mamba timing layer is introduced on the basis of a Mamba convolution block, the Mamba timing layer combines a selective SSM mechanism and an MLP structure to form a homogeneous stacking module; the pre-trained MCLDNN model includes multiple MCLDNN convolution blocks and a timing layer, the MCLDNN convolution blocks are used to extract local features, and the timing layer is used to capture long sequence dependencies; the pre-trained modulated signal classification model uses modulated signal samples and a preset loss function, and is constructed based on a mutual learning strategy between an initial Mamba model and an initial MCLDNN model. In the present invention, the pre-trained Mamba model in the pre-trained modulated signal classification model can effectively extract and process the features of the modulated signal to be classified by combining the features extracted by the selective state space (SSM) and the Mamba convolution block, thereby improving the accuracy of the pre-trained modulated signal classification model in recognizing the modulated signal to be classified; in addition, by designing a specific loss function, the initial Mamba model and the initial MCLDNN model can learn from each other during normal training without introducing additional disturbances to the data, thereby improving the performance of the pre-trained modulated signal classification model. Since the method of the present invention no longer relies on expert experience and improves the generalization ability and computing resource utilization of the pre-trained modulated signal classification model by using a more flexible fusion strategy, efficient training and real-time processing can be achieved.

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A schematic diagram of a flow chart of a method for classifying modulation signals provided by an embodiment of the present invention;

[0031] Figure 2 The structural diagram of the pre-trained MCLDNN model is shown as an example;

[0032] Figure 3 The structural diagram of the pre-trained Mamba model is shown as an example;

[0033] Figure 4 The schematic diagram of the structure of the Mamba timing layer is shown as an example;

[0034] Figure 5 The overall architecture diagram of the pre-trained modulation signal classification model is shown exemplarily;

[0035] Figure 6 A schematic structural diagram of a modulation signal classification device provided by an embodiment of the present invention;

[0036] Figure 7 A schematic structural diagram of a modulation signal classification device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0038] In order to improve the generalization ability and computing resource utilization of a pre-trained modulation signal classification model, an embodiment of the present invention provides a modulation signal classification method. Figure 1 Schematic diagram of a flow chart of a modulation signal classification method provided by an embodiment of the present invention. Figure 1 Shown, including:

[0039] S101: Obtain a modulated signal to be classified.

[0040] It should be noted that the specific form of the modulated signal depends on the modulation technology used. In this embodiment, the modulated signal to be classified can be obtained based on the following digital modulation methods: 8-phase shift keying (8PSK), binary phase shift keying (BPSK), continuous phase frequency shift keying (CPFSK), Gaussian filtered frequency shift keying (GFSK), four-level pulse amplitude modulation (PAM4), 16-quadrature amplitude modulation (16QAM), 64-quadrature amplitude modulation (64QAM), and four-phase phase shift keying (QPSK). In addition, three analog modulation methods can also be used: double-sideband amplitude modulation (AM-DSB), single-sideband amplitude modulation (AM-SSB), and wideband frequency modulation (WBFM).

[0041] S102 : Input the modulation signal to be classified into a pre-trained modulation signal classification model to obtain a modulation signal classification result.

[0042] The pre-trained modulated signal classification model is equipped with a parallel connected pre-trained Mamba model and a pre-trained MCLDNN model; the pre-trained Mamba model is constructed based on the selective state-space model, and the Mamba timing layer is introduced on the basis of the Mamba convolution block. The Mamba timing layer combines the selective SSM mechanism and the MLP structure to form a homogeneous stacked module; the pre-trained MCLDNN model includes multiple MCLDNN convolution blocks and timing layers. The MCLDNN convolution blocks are used to extract local features, and the timing layers are used to capture long sequence dependencies; the pre-trained modulated signal classification model uses modulated signal samples and a preset loss function, and is constructed based on the mutual learning strategy between the initial Mamba model and the initial MCLDNN model.

[0043] The embodiment of the present invention provides a method, device and equipment for classifying modulated signals. Among them, a method for classifying modulated signals includes: obtaining a modulated signal to be classified; inputting the modulated signal to be classified into a pre-trained modulated signal classification model to obtain a modulated signal classification result; the pre-trained modulated signal classification model is provided with a pre-trained Mamba model and a pre-trained MCLDNN model connected in parallel; the pre-trained Mamba model is constructed based on a selective state space model, and a Mamba timing layer is introduced on the basis of the Mamba convolution block, and the Mamba timing layer combines the selective SSM mechanism and the MLP structure to form a homogeneous stacking module; the pre-trained MCLDNN model includes multiple MCLDNN convolution blocks and timing layers, the MCLDNN convolution blocks are used to extract local features, and the timing layers are used to capture long sequence dependencies; the pre-trained modulated signal classification model adopts modulated signal samples and a preset loss function, and is constructed based on the mutual learning strategy between the initial Mamba model and the initial MCLDNN model. In an embodiment of the present invention, the pre-trained Mamba model in the pre-trained modulated signal classification model can effectively extract and process the features of the modulated signal to be classified by combining the features extracted by the selective state space (SSM) and the Mamba convolution block, thereby improving the accuracy of the pre-trained modulated signal classification model in recognizing the modulated signal to be classified; in addition, by designing a specific loss function, the initial Mamba model and the initial MCLDNN model can learn from each other during normal training without introducing additional disturbances to the data, thereby improving the performance of the pre-trained modulated signal classification model. Since the method of the present invention no longer relies on expert experience and improves the generalization ability and computing resource utilization of the pre-trained modulated signal classification model by using a more flexible fusion strategy, efficient training and real-time processing can be achieved.

[0044] To illustrate the model structure of the pre-trained MCLDNN model, Figure 2 The schematic diagram of the structure of the pre-trained MCLDNN model is shown as an example. Figure 2 As shown, the pre-trained MCLDNN model includes: a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a fifth convolution block, a first sequential layer, a second sequential layer, a first fully connected layer, and a second fully connected layer;

[0045] The first convolution block, the second convolution block and the third convolution block are connected in parallel, the outputs of the second convolution block and the third convolution block are connected in series with the fourth convolution block, the output of the first convolution block is connected in parallel with the output of the fourth convolution block, the first convolution block and the fourth convolution block are connected in series with the fifth convolution block, and the fifth convolution block is connected in series with the first sequential layer, the second sequential layer, the first fully connected layer and the second fully connected layer.

[0046] The specific structure of each convolution block in the pre-trained MCLDNN model is: convolution layer, batch normalization layer, ReLU activation layer. Figure 2 As shown in the figure, the number of convolution kernels of the first convolution block, the second convolution block, the third convolution block, the fourth convolution block, and the fifth convolution block are 50, 50, 50, 50, and 100 respectively, and the sizes of the convolution kernels are 2*8, 1*8, 1*8, 1*8, and 2*5 respectively. The stride is 1 and the padding is 0. The hidden layer dimension of the first LSTM (time series) layer is 128, the hidden layer dimension of the second LSTM (time series) layer is 128, the output dimension of the first fully connected layer is 128, and the output dimension of the second fully connected layer is 11. The optimizer uses Adam.

[0047] Figure 3 The schematic diagram of the structure of the pre-trained Mamba model is shown as an example. Figure 3 As shown, the pre-trained Mamba model includes: the sixth convolution block, the seventh convolution block, the eighth convolution block, the ninth convolution block, the tenth convolution block, the first Mamba timing layer, the second Mamba timing layer, the third fully connected layer, and the fourth fully connected layer;

[0048] The sixth convolution block, the seventh convolution block, and the eighth convolution block are connected in parallel, the outputs of the seventh and eighth convolution blocks are connected in series with the ninth convolution block, the output of the sixth convolution block is connected in parallel with the output of the ninth convolution block, the sixth and ninth convolution blocks are connected in series with the tenth convolution block, and the tenth convolution block is sequentially connected in series with the first Mamba timing layer, the second Mamba timing layer, the third fully connected layer, and the fourth fully connected layer.

[0049] In the pre-trained Mamba model, each convolutional block consists of a convolutional layer, a batch normalization layer, and a ReLU activation layer. The number of convolution kernels in the sixth, seventh, eighth, ninth, and tenth convolutional blocks is 50, 50, 50, 50, and 100, respectively, and the kernel sizes are 2*8, 1*8, 1*8, 1*8, and 2*5, respectively. The stride is 1, and the padding is 0. The hidden dimension of the first and second Mamba temporal layers is 128, the output dimension of the third and fourth fully connected layers is 128, and the output dimension of the fourth fully connected layer is 11. The Adam optimizer is used.

[0050] Optionally, the output dimensions of the first fully connected layer and the third fully connected layer are both 128, and the output dimensions of the second fully connected layer and the fourth fully connected layer are both 11; and the optimizers of the MCLDNN model and the Mamba model both use Adam.

[0051] Figure 4The schematic diagram of the Mamba timing layer is shown as an example. The Mamba timing layer cleverly integrates technologies such as linear projection, state-space model, deep convolution, and activation function (SiLU). Through linear projection, the pre-trained Mamba model can efficiently extract key features of the input sequence; the state-space model provides a mechanism for dynamically selecting and updating state information, ensuring that the pre-trained Mamba model remains efficient and accurate when processing long sequences; the deep convolution layer further enhances the pre-trained Mamba model's ability to learn features of sequence data, capturing complex temporal dependencies through multi-layer convolution operations; and the introduction of activation functions gives the pre-trained Mamba model nonlinear expression capabilities, enabling it to fit more complex temporal patterns. The organic combination of these features and technologies enables the Mamba timing layer to demonstrate excellent performance and flexibility when processing time series data.

[0052] It should be noted that the activation function used in the Mamba timing layer is SiLU. SiLU (Sigmoid Linear Unit) is an activation function that combines the characteristics of the sigmoid function and the linear function. It has the advantages of nonlinearity, smoothness, and self-gating. It is suitable for a variety of deep learning tasks and can improve model performance and training efficiency.

[0053] Optionally, the preset loss function is expressed as:

[0054]

[0055] Among them, L1 represents the preset loss function, CrossEntropyLoss represents the cross entropy loss function, Represents the modulated signal sample X label Input the prediction score of the initial MCLDNN model, Represents the modulated signal sample X label Enter the prediction score of the initial Mamba model, Represents the modulated signal sample X label The true label, Represents the one-hot encoding converted based on the prediction score of the initial MCLDNN model, Represents the one-hot encoding transformed according to the prediction score of the initial Mamba model.

[0056] Optionally, the training process of the pre-trained modulation signal classification model includes:

[0057] Obtaining a modulated signal sample;

[0058] Input the modulated signal samples into the initial modulated signal classification model for iterative training;

[0059] The initial modulation signal classification model that meets the preset iteration stop condition is used as the pre-trained modulation signal classification model;

[0060] The preset iteration stopping conditions include: the number of iterations is greater than the preset iteration threshold or the preset loss function continues to converge.

[0061] Specifically, in this embodiment, the number of iterations can be set to 200 rounds, and the learning rate can be set to 0.0003.

[0062] Figure 5 The overall architecture diagram of the pre-trained modulation signal classification model is shown as an example. Figure 5 As shown in Figure 2, the initial Mamba model and the initial MCLDNN model are each trained based on the cross entropy loss and learn from each other through a preset loss function.

[0063] Optionally, the pre-trained Mamba model is used to process the modulated signal to be classified in a low signal-to-noise ratio environment, and the nonlinear expression ability and dynamic information filtering effect of the pre-trained modulated signal classification model are enhanced through a selective SSM mechanism.

[0064] It should be noted that the core of the pre-trained Mamba model lies in structured state space models (SSMs). SSMs are a mechanism that allows a model to dynamically adjust its state based on input. This selective mechanism enables the pre-trained Mamba model to effectively filter out important information when processing sequence data, while ignoring or forgetting irrelevant parts. Furthermore, the pre-trained Mamba model simplifies the traditional deep sequence model architecture by combining state space models (SSMs) with multi-layer perceptron (MLP) blocks in the Transformer to form a simple and unified structure, namely a homogeneous stacked module. This design not only improves the model's flexibility but also reduces its complexity, enabling the pre-trained Mamba model to efficiently process long sequence data during both training and inference.

[0065] In summary, the initial Mamba model designed in this paper, suitable for signal modulation, combines features extracted by the Selective State Space Matrix (SSM) and convolutional blocks, and applies a nonlinear transformation to the output using a nonlinear activation function. Each temporal layer utilizes sequence transformations to capture temporal dependencies, and combined with a linear projection layer to preserve the global temporal information of the input, effectively extracts and processes the features of the modulated signal to be classified, helping to improve the accuracy of automatic modulation recognition.

[0066] Furthermore, the present invention constructs a pre-trained modulation signal classification model based on two randomly initialized network models, the initial Mamba model and the initial MCLDNN model. By designing a specific loss function, the two models can learn from each other during normal training. This interactive learning method can make the data predictions different without introducing additional perturbations to the data, and then train the initial modulation signal classification model through consistency constraints, ultimately improving the performance of the pre-trained modulation signal classification model.

[0067] The method provided in the embodiment of the present invention can be applied to electronic devices. Specifically, the electronic devices can be desktop computers, portable computers, smart mobile terminals, servers, etc., which are not limited in the embodiment of the present invention.

[0068] Based on the same inventive concept, an embodiment of the present invention further provides a device for classifying modulated signals. Figure 6 Schematic diagram of a modulation signal classification device provided by an embodiment of the present invention. Figure 6 As shown, it includes: an acquisition unit 601 and a classification unit 602;

[0069] The acquisition unit 601 is used to: acquire a modulated signal to be classified;

[0070] The classification unit 602 is used to: input the modulated signal to be classified into the pre-trained modulated signal classification model to obtain the modulation signal classification result; the pre-trained modulated signal classification model is provided with a parallel connected pre-trained Mamba model and a pre-trained MCLDNN model; the pre-trained Mamba model is constructed based on the selective state space model, and the Mamba timing layer is introduced on the basis of the Mamba convolution block, and the Mamba timing layer combines the selective SSM mechanism and the MLP structure to form a homogeneous stacking module; the pre-trained MCLDNN model includes multiple MCLDNN convolution blocks and timing layers, the MCLDNN convolution blocks are used to extract local features, and the timing layers are used to capture long sequence dependencies; the pre-trained modulated signal classification model adopts the modulation signal samples and the preset loss function, and is constructed based on the mutual learning strategy between the initial Mamba model and the initial MCLDNN model.

[0071] Figure 7 A schematic structural diagram of a modulated signal classification device provided in an embodiment of the present invention includes: a processor 710, a storage medium 720, and a bus 730. The storage medium 720 stores machine-readable instructions executable by the processor 710. When the modulated signal classification device is in operation, the processor 710 and the storage medium 720 communicate via the bus 730, and the processor 710 executes the machine-readable instructions to perform the steps of the above-described method embodiment. The specific implementation methods and technical effects are similar and will not be repeated here.

[0072] The storage medium may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage. Alternatively, the storage medium may be at least one storage device located away from the processor.

[0073] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0074] The present invention also provides a computer-readable storage medium in which a computer program is stored, and when the computer program is executed by a processor, the steps of any of the above-mentioned modulation signal classification methods are implemented.

[0075] It should be noted that the terms "first," "second," and the like are used to distinguish similar objects and are not necessarily used to describe a particular order or precedence. 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. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention.

[0076] In the description of this specification, the reference terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" mean that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification.

[0077] Although the present invention is described herein in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art can understand and implement other variations of the above-mentioned disclosed embodiments by viewing the drawings and the disclosed content. In the description of the present invention, the word "comprising" does not exclude other components or steps, "one" or "an" does not exclude multiple situations, and the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, certain measures are recorded in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0078] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention cannot be considered to be limited to these descriptions. For those skilled in the art of the present invention, several simple deductions or substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for classifying modulated signals, characterized in that: include: Obtaining a modulated signal to be classified; Inputting the modulation signal to be classified into a pre-trained modulation signal classification model to obtain a modulation signal classification result; The pre-trained modulated signal classification model is provided with a pre-trained Mamba model and a pre-trained MCLDNN model connected in parallel; the pre-trained Mamba model is constructed based on a selective state space model, and a Mamba timing layer is introduced on the basis of the Mamba convolution block, and the Mamba timing layer combines the selective SSM mechanism and the MLP structure to form a homogeneous stacking module; the pre-trained MCLDNN model includes multiple MCLDNN convolution blocks and a timing layer, the MCLDNN convolution block is used to extract local features, and the timing layer is used to capture long sequence dependencies; the pre-trained modulated signal classification model adopts modulated signal samples and a preset loss function, and is constructed based on a mutual learning strategy between an initial Mamba model and an initial MCLDNN model; The pre-trained MCLDNN model includes: a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a fifth convolution block, a first sequential layer, a second sequential layer, a first fully connected layer, and a second fully connected layer; The first convolution block, the second convolution block, and the third convolution block are connected in parallel, the outputs of the second convolution block and the third convolution block are both connected in series with the fourth convolution block, the output of the first convolution block is connected in parallel with the output of the fourth convolution block, the first convolution block and the fourth convolution block are both connected in series with the fifth convolution block, and the fifth convolution block is sequentially connected in series with the first sequential layer, the second sequential layer, the first fully connected layer, and the second fully connected layer; The pre-trained Mamba model includes: a sixth convolution block, a seventh convolution block, an eighth convolution block, a ninth convolution block, a tenth convolution block, a first Mamba timing layer, a second Mamba timing layer, a third fully connected layer, and a fourth fully connected layer; The sixth convolution block, the seventh convolution block, and the eighth convolution block are connected in parallel, the outputs of the seventh convolution block and the eighth convolution block are both connected in series with the ninth convolution block, the output of the sixth convolution block is connected in parallel with the output of the ninth convolution block, the sixth convolution block and the ninth convolution block are both connected in series with the tenth convolution block, and the tenth convolution block is sequentially connected in series with the first Mamba timing layer, the second Mamba timing layer, the third fully connected layer, and the fourth fully connected layer; The first Mamba timing layer and the second Mamba timing layer implement sequence feature extraction based on linear projection, state space model, depth convolution and SiLU activation function; The preset loss function is expressed as: ； in, represents the preset loss function, represents the cross entropy loss function, Represents the modulated signal sample Input the prediction score of the initial MCLDNN model, Represents the modulated signal sample Enter the prediction score of the initial Mamba model, Represents the modulated signal sample The true label, Represents the one-hot encoding converted based on the prediction score of the initial MCLDNN model, Represents the one-hot encoding transformed according to the prediction score of the initial Mamba model.

2. The method for classifying modulation signals according to claim 1, wherein: The output dimensions of the first fully connected layer and the third fully connected layer are both 128, and the output dimensions of the second fully connected layer and the fourth fully connected layer are both 11; the optimizers of the MCLDNN model and the Mamba model both use Adam.

3. The method for classifying modulation signals according to claim 1, wherein: The training process of the pre-trained modulation signal classification model includes: Obtaining a modulated signal sample; Inputting the modulated signal samples into an initial modulated signal classification model for iterative training; Using the initial modulation signal classification model that meets the preset iteration stop condition as the pre-trained modulation signal classification model; The preset iteration stopping condition includes: the number of iterations is greater than a preset iteration threshold or the preset loss function continues to converge.

4. The method for classifying modulation signals according to claim 1, wherein: The pre-trained Mamba model is used to process the modulated signal to be classified in a low signal-to-noise ratio environment, and enhances the nonlinear expression ability and dynamic information filtering effect of the pre-trained modulated signal classification model through a selective SSM mechanism.

5. A modulation signal classification device, characterized in that: include: The acquisition unit is used to: acquire the modulated signal to be classified; The classification unit is used to: input the modulated signal to be classified into a pre-trained modulated signal classification model to obtain a modulated signal classification result; the pre-trained modulated signal classification model is provided with a pre-trained Mamba model and a pre-trained MCLDNN model connected in parallel; the pre-trained Mamba model is constructed based on a selective state space model, and a Mamba timing layer is introduced on the basis of the Mamba convolution block, and the Mamba timing layer combines the selective SSM mechanism and the MLP structure to form a homogeneous stacking module; the pre-trained MCLDNN model includes multiple MCLDNN convolution blocks and a timing layer, the MCLDNN convolution block is used to extract local features, and the timing layer is used to capture long sequence dependencies; the pre-trained modulated signal classification model adopts modulated signal samples and a preset loss function, and is constructed based on a mutual learning strategy between an initial Mamba model and an initial MCLDNN model; The pre-trained MCLDNN model includes: a first convolution block, a second convolution block, a third convolution block, a fourth convolution block, a fifth convolution block, a first sequential layer, a second sequential layer, a first fully connected layer, and a second fully connected layer; The first convolution block, the second convolution block, and the third convolution block are connected in parallel, the outputs of the second convolution block and the third convolution block are both connected in series with the fourth convolution block, the output of the first convolution block is connected in parallel with the output of the fourth convolution block, the first convolution block and the fourth convolution block are both connected in series with the fifth convolution block, and the fifth convolution block is sequentially connected in series with the first sequential layer, the second sequential layer, the first fully connected layer, and the second fully connected layer; The pre-trained Mamba model includes: a sixth convolution block, a seventh convolution block, an eighth convolution block, a ninth convolution block, a tenth convolution block, a first Mamba timing layer, a second Mamba timing layer, a third fully connected layer, and a fourth fully connected layer; The sixth convolution block, the seventh convolution block, and the eighth convolution block are connected in parallel, the outputs of the seventh convolution block and the eighth convolution block are both connected in series with the ninth convolution block, the output of the sixth convolution block is connected in parallel with the output of the ninth convolution block, the sixth convolution block and the ninth convolution block are both connected in series with the tenth convolution block, and the tenth convolution block is sequentially connected in series with the first Mamba timing layer, the second Mamba timing layer, the third fully connected layer, and the fourth fully connected layer; The first Mamba timing layer and the second Mamba timing layer implement sequence feature extraction based on linear projection, state space model, depth convolution and SiLU activation function; The preset loss function is expressed as: ； in, represents the preset loss function, represents the cross entropy loss function, Represents the modulated signal sample Input the prediction score of the initial MCLDNN model, Represents the modulated signal sample Enter the prediction score of the initial Mamba model, Represents the modulated signal sample The true label, Represents the one-hot encoding converted based on the prediction score of the initial MCLDNN model, Represents the one-hot encoding transformed according to the prediction score of the initial Mamba model.

6. A modulation signal classification device, characterized in that include: A processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor. When the modulation signal classification device is running, the processor and the storage medium communicate via the bus, and the processor executes the machine-readable instructions to perform the steps of the modulation signal classification method according to any one of claims 1 to 4.

Images