A modulation recognition method and system based on higher-order moments and their spectrum

Abstract

The application discloses a modulation recognition method and system based on high-order moments and spectrum, and relates to the technical field of wireless communication.The method comprises the following steps: establishing the interaction between a dynamic modulation terminal and a communication channel, introducing a target signal into a sparse encoder at the terminal edge, performing sparse compression projection, and determining a sparse signal; determining high-order moments for the sparse signal, triggering a tensor decomposition layer to perform tensor decomposition, and determining low-rank moments; triggering a modulation recognition layer to perform multi-stage classification modulation decision for the low-rank moments, determining a modulation recognition result, and performing signal modulation processing on the target signal.The application solves the technical problems of low precision, high computational complexity and lack of dynamic adaptability of the existing modulation recognition method in the case of complex channel environment, low signal-to-noise ratio and coexistence of multiple modulation modes, and achieves the technical effects of improving modulation recognition precision, reducing computational complexity and realizing high adaptability to complex channels through high-order moment analysis and multi-stage decision.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and specifically to a modulation identification method and system based on higher-order moments and their spectra. Background Technology

[0002] In fields such as communications, radar, and electronic warfare, modulation identification technology is used to determine the modulation scheme from a signal, which is crucial for the efficient operation of communication systems. Traditional methods are mainly based on the analysis of signal characteristics in the time domain, frequency domain, or time-frequency domain. However, in complex and variable channel environments, especially in scenarios with low signal-to-noise ratio (SNR) and multiple modulation schemes, these methods lack sufficient accuracy and robustness. In recent years, modulation identification technology based on higher-order moments and their spectra has attracted attention. Higher-order moments can capture the non-Gaussian properties of signals, but they have high computational complexity, and the accuracy of feature extraction under noise and interference needs improvement. In addition, existing technologies are difficult to adapt to dynamic communication requirements and lack the ability to adjust the modulation scheme in real time. Summary of the Invention

[0003] This application provides a modulation identification method and system based on higher-order moments and their spectra, which solves the technical problems of low accuracy, computational complexity and lack of dynamic adaptability of existing modulation identification methods in complex channel environments, low signal-to-noise ratios and the coexistence of multiple modulation methods.

[0004] The first aspect of this application provides a modulation recognition method based on higher-order moments and their spectra. The method includes: establishing an interaction between a dynamic modulation end and a communication channel; triggering a dynamic modulation command during communication transmission; importing a target signal into a sparse encoder at the end; performing sparse compression projection to determine a sparse signal, wherein the sparse encoder follows an equidistant principle; determining higher-order moments for the sparse signal and triggering a tensor decomposition layer to perform tensor decomposition to determine low-rank moments, wherein the tensor decomposition includes a first tensor decomposition mode and a second tensor decomposition mode; triggering a modulation recognition layer to perform multi-level classification modulation decisions for the low-rank moments; determining the modulation recognition result; performing signal modulation processing on the target signal; and returning the result to the communication channel for communication transmission management. The cascaded sparse encoder, tensor decomposition layer, and modulation recognition layer serve as built-in plugins for the dynamic modulation end.

[0005] A second aspect of this application provides a modulation recognition system based on higher-order moments and their spectra. The system includes: a sparse coding module for establishing interaction between a dynamic modulation end and a communication channel, triggering dynamic modulation commands during communication transmission, and importing the target signal into a sparse encoder at the end, performing sparse compression projection, and determining a sparse signal, wherein the sparse encoder follows an equidistant principle; a tensor decomposition module for determining higher-order moments for the sparse signal and triggering a tensor decomposition layer to perform tensor decomposition, determining low-rank moments, wherein the tensor decomposition includes a first tensor decomposition mode and a second tensor decomposition mode; and a multi-level classification modulation decision module for triggering a modulation recognition layer to perform multi-level classification modulation decisions for the low-rank moments, determining the modulation recognition result, performing signal modulation processing on the target signal, and returning to the communication channel for communication transmission management; wherein the cascaded sparse encoder, tensor decomposition layer, and modulation recognition layer serve as built-in plug-ins for the dynamic modulation end.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] This application provides a modulation recognition method and system based on higher-order moments and their spectra, relating to the field of wireless communication technology. Through the interaction between a dynamic modulation end and the communication channel, a sparse encoder is used for signal compression, tensor decomposition is employed to extract low-rank moments, and modulation schemes are identified under multi-level classification decisions. Finally, signal modulation processing is completed and the signal is returned to the communication channel. This solves the technical problems of existing modulation recognition methods, such as low accuracy, computational complexity, and lack of dynamic adaptability, in complex channel environments, low signal-to-noise ratios, and the coexistence of multiple modulation schemes. It achieves the technical effect of improving modulation recognition accuracy and reducing computational complexity through higher-order moment analysis and multi-level decision-making, thus realizing high adaptability to complex channels. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A schematic flowchart of a modulation recognition method based on higher-order moments and their spectra provided in this application embodiment;

[0010] Figure 2 This is a schematic diagram of a modulation recognition system based on higher-order moments and their spectra, provided as an embodiment of this application.

[0011] Figure labeling: sparse coding module 11, tensor decomposition module 12, multi-level classification modulation decision module 13. Detailed Implementation

[0012] This application provides a modulation identification method and system based on higher-order moments and their spectra, which solves the technical problems of low accuracy, computational complexity and lack of dynamic adaptability of existing modulation identification methods in complex channel environments, low signal-to-noise ratios and the coexistence of multiple modulation methods.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, this application provides a modulation recognition method based on higher-order moments and their spectra, the method comprising:

[0016] P10: Establish interaction between the dynamic modulation end and the communication channel, trigger dynamic modulation commands during communication transmission, and import the target signal into the sparse encoder at the end, perform sparse compression projection, and determine the sparse signal. The sparse encoder follows the equidistant principle. A cascaded sparse encoder, tensor decomposition layer, and modulation recognition layer serve as built-in plugins for the dynamic modulation end.

[0017] Specifically, establishing an interaction mechanism between the dynamic modulation end and the communication channel is fundamental to the entire modulation identification process. As a key component of the communication system, the dynamic modulation end's main function is to trigger dynamic modulation commands in real time according to the needs of communication transmission. The communication channel's state parameters, such as signal-to-noise ratio (SNR), signal strength, and transmission delay, are monitored in real time by the dynamic modulation end. Based on these parameters, the dynamic modulation end can generate corresponding modulation commands, thereby dynamically adjusting the modulation scheme to optimize the quality and efficiency of signal transmission.

[0018] Once the dynamic modulation command is triggered, the target signal is immediately fed into the edge-side sparse encoder. Edge computing is a distributed computing architecture whose core idea is to offload computing tasks from the cloud or central server to devices at the network edge to reduce data transmission latency and bandwidth consumption. In this application, the edge-side sparse encoder undertakes the important task of sparsely compressing and projecting the target signal. Based on sparse representation theory, the sparse encoder's main function is to compress high-dimensional signals into a low-dimensional space while preserving the signal's key features. The core assumption of sparse representation theory is that a signal is sparsity in a certain transform domain, meaning that most of the signal's energy is concentrated in a few coefficients. Through the processing of the sparse encoder, the target signal is converted into a sparse signal, which can significantly reduce the signal's dimensionality and computational complexity.

[0019] During sparse compression projection, the sparse encoder strictly adheres to the isometric principle. The isometric principle is a crucial concept in signal processing, essentially ensuring that the signal's geometry and feature distribution are preserved during compression. Specifically, the isometric principle requires that the distance relationships between signals remain unchanged when mapping from a high-dimensional space to a low-dimensional space. By adhering to the isometric principle, the sparse encoder ensures that key signal features are preserved during compression, thus providing an accurate signal foundation for subsequent high-order moment analysis and modulation recognition.

[0020] After sparse compression projection, the sparse signal is passed to subsequent processing modules: the tensor decomposition layer and the modulation recognition layer. These two modules are cascaded with the sparse encoder, forming a built-in plug-in for the dynamic modulation end. This cascaded structure enables the entire modulation recognition system to process signals efficiently, from sparse signal representation to higher-order moment extraction, and finally to modulation recognition, forming a complete signal processing flow. This structure not only effectively reduces computational complexity but also enables high-precision modulation recognition in complex and ever-changing communication environments.

[0021] Furthermore, before importing the target signal into the sparse encoder at the end, in order to construct the sparse encoder, step P10 of this embodiment further includes:

[0022] P11a: Construct a sparse coding network, where each grid corresponds to a type of signal point; P12a: Traverse the sparse coding network, perform coding relationship mining and training grid by grid until N coding grids are determined, where N is the number of signal point types; P13a: Concatenate the N coding grids to determine the sparse encoder.

[0023] Optionally, the construction process of the sparse encoder can be described in further detail. First, a sparse coding network is constructed. The core design of this network is to map the features of the signal to different types of signal points in a grid. Each grid represents a class of signal points, and each class of signal points carries different types of signal features. The network effectively maps the multidimensional features of the signal, with each grid designed to independently process a specific category of signal features. This provides a structured framework for efficient signal compression and dimensionality reduction.

[0024] Next, the sparse coding network is traversed, and coding relationships are mined and trained grid by grid. This process involves analyzing and learning the signal points within each grid to uncover the coding relationships between them, and then optimizing these relationships through training. The goal of training is to enable the sparse encoder to accurately map the input signal to the corresponding coding grid. This process needs to be repeated until N coding grids are determined, where N is the number of signal point types. These N coding grids correspond to different signal point types, thus providing the sparse encoder with comprehensive coding capabilities.

[0025] Finally, these N coding grids are concatenated to determine the sparse encoder. The concatenation process integrates the various coding grids into a unified encoder structure, enabling it to process the input signal as a whole. The sparse encoder constructed in this way can effectively encode different types of signal points and, in the subsequent sparse compression projection process, adheres to the equidistant principle, ensuring the preservation of key signal features. Ultimately, the constructed sparse encoder provides strong support for subsequent signal processing, ensuring both signal feature preservation and efficient compression.

[0026] Furthermore, in this embodiment, step P112a further includes performing encoding relationship mining and training grid by grid:

[0027] P12-1a: Identify the first grid, and perform dimensionality reduction and compression projection relationship mining on the first signal point type based on the equidistant value of the signal points to determine the first encoding relationship, wherein the first grid is any grid in the sparse coding network; P12-2a: Write the first encoding relationship into the first grid, and perform grid autoencoder supervised training until convergence to determine the first encoding grid.

[0028] In one possible embodiment of this application, in order to further improve the construction process of the sparse encoder, the encoding relationship can be mined and trained grid by grid to ensure that the role and function of each grid in the sparse encoder are optimized and effectively confirmed.

[0029] First, the first grid needs to be identified and selected. This grid is any grid in the sparse coding network. Each grid corresponds to a class of signal points, and processing is performed based on the characteristics of these signal points. In this step, the signal points within the grid are first arranged at equal intervals based on their value. Equal intervals mean that the weight or importance of each signal point is evenly distributed across the grid according to its characteristics, ensuring that each signal point has a reasonable representation within the grid. After determining the first grid and its signal points, the relationship mining for dimensionality reduction and compression projection is performed. The process of dimensionality reduction and compression projection maps high-dimensional signal data to a low-dimensional space through a mathematical transformation (such as principal component analysis, random projection, etc.), thereby reducing computational complexity and extracting the most representative signal features. In this process, the first coding relationship can be determined by mining the mapping relationship between signal points, i.e., the transformation method and projection rule of the signal during dimensionality reduction. This coding relationship ensures that the signal features are effectively represented in the dimensionality-reduced low-dimensional space, providing a theoretical basis for subsequent coding training and compression processing.

[0030] After determining the first encoding relationship, this relationship is written into the first grid, enabling the first grid to process signals according to the defined encoding relationship. Next, supervised training of the grid autoencoder is performed, which involves learning and optimizing the signal through an autoencoder network. An autoencoder network is an unsupervised learning method that learns the intrinsic features of a signal by encoding the input signal into a low-dimensional representation and then decoding it to recover the original signal. During this training process, the system optimizes the network using supervised learning based on known signal features and label information until the network error converges. The goal of the training is to enable the first encoding grid to accurately reconstruct the input signal through autoencoding, and to adjust the encoding method within the grid to better adapt to the feature requirements of different signal points. Finally, the first encoding grid, i.e., the trained and optimized grid, is determined to effectively encode the first signal point type.

[0031] P20: For the sparse signal, determine the higher-order moments and trigger the tensor decomposition layer to perform tensor decomposition, and determine the lower-rank moments, wherein the tensor decomposition includes a first tensor decomposition mode and a second tensor decomposition mode.

[0032] Furthermore, step P20 in this embodiment of the application also includes:

[0033] P21: Traverse the sparse signal to locate the first stationary signal segment and the second non-stationary signal segment; P22: For the first stationary signal segment, determine the first tensor decomposition mode, which uses the rank tensor as the decomposition criterion; P23: For the second non-stationary signal segment, determine the second tensor decomposition mode, which uses the product of the core tensor and the multidimensional factor tensor as the decomposition criterion; P24: Based on the first and second tensor decomposition modes, perform multi-threaded tensor decomposition processing on the higher-order moments to determine the lower-rank moments. The tensor decomposition layer is based on programmable gating deployment, and the first and second tensor decomposition modes can be extended through gating programming.

[0034] It should be understood that high-order moment analysis is performed on sparse signals and tensor decomposition layers are triggered for processing to extract the low-rank moments of the signals.

[0035] First, the sparse signal is traversed to locate the first stationary signal segment and the second non-stationary signal segment. This process involves analyzing the characteristics of the sparse signal to identify its stationary and non-stationary properties. A stationary signal segment refers to a region where the signal's statistical properties remain constant or change only slightly over time, while a non-stationary signal segment refers to a region where the signal's statistical properties change significantly over time. This localization provides a basis for selecting an appropriate mode for subsequent tensor decomposition.

[0036] Next, for the first stationary signal segment, the first tensor decomposition mode is determined. The first tensor decomposition mode uses the rank-one tensor as the decomposition standard. Rank-one tensor decomposition is a method that decomposes a high-order tensor into a linear combination of several rank-one components. A rank-one component is a tensor that can be represented as the outer product of multiple vectors. This decomposition method can effectively extract the main features of the signal while reducing computational complexity. For stationary signal segments, due to their relatively stable statistical properties, rank-one tensor decomposition can capture their features well, thus providing effective information for modulation identification.

[0037] Next, for the second non-stationary signal segment, a second tensor decomposition mode is determined. This second tensor decomposition mode uses the product of the core tensor and the multidimensional factor tensor as the decomposition criterion. This decomposition method typically employs Tucker decomposition, which preserves the multidimensional interaction features of the tensor and approximates the original tensor through the product of the core tensor and the multidimensional factor tensor. For non-stationary signal segments, because their statistical properties change significantly over time, Tucker decomposition can better adapt to this change, extracting the multidimensional interaction features of the signal and thus improving the accuracy of modulation recognition.

[0038] Finally, based on the first and second tensor decomposition patterns, high-order moments are processed using multi-threaded tensor decomposition to determine low-rank moments. Multi-threading is a parallel computing technique that significantly improves computational efficiency by distributing computational tasks across multiple threads for simultaneous execution. The tensor decomposition layer is based on programmable gating deployment, meaning that tensor decomposition patterns can be flexibly configured and extended through gating programming. Gating programming allows for dynamic adjustment of tensor decomposition parameters and patterns according to signal characteristics, thereby improving the system's adaptability and flexibility.

[0039] The first and second tensor decomposition modes can be extended through gating programming. This flexibility allows the system to adaptively adjust to different signal characteristics, improving the overall system's processing power and robustness.

[0040] Furthermore, step P24 in this embodiment of the application also includes:

[0041] P24-1: Calculate higher-order moments for the sparse signal; P24-2: Based on the higher-order moments corresponding to the first stationary signal segment, perform moment decomposition under the first tensor decomposition mode to determine a class of low-rank moments; P24-3: Based on the higher-order moments corresponding to the second non-stationary signal segment, perform moment decomposition under the second tensor decomposition mode to determine a class of low-rank moments; P24-4: Based on the class of low-rank moments and the class of low-rank moments, determine the low-rank moments, and establish the mutual mapping between the low-rank moments, higher-order moments, sparse signal and target signal.

[0042] Specifically, the process of tensor decomposition of the higher-order moments of sparse signals can be further refined to better understand how each step works together, ultimately determining the low-rank moments and the inter-mapping between signals.

[0043] First, higher-order moments are calculated for the sparse signal. Higher-order moments are a process of in-depth analysis of the statistical properties of a signal, capable of capturing its non-Gaussian and transient variation characteristics. By calculating higher-order moments, the complexity and structural features of the signal can be better understood, providing a foundation for subsequent tensor decomposition processing and helping the system identify various features of the signal.

[0044] Next, based on the higher-order moments corresponding to the first stationary signal segment, moment decomposition based on the first tensor decomposition mode is performed. During this process, the rank-tensor decomposition mode can be used to process the stationary signal segment. The rank-tensor decomposition mode is suitable for processing stationary signals because such signals have relatively stable statistical characteristics in the time domain, and complex time variations do not need to be considered during decomposition. Through moment decomposition, higher-order moments are transformed into lower-rank moments. These lower-rank moments can effectively describe the main features of the signal and remove redundant parts, resulting in a class of lower-rank moments that can effectively extract the main features of the stationary signal segment and reduce computational complexity.

[0045] Next, for the higher-order moments corresponding to the second non-stationary signal segment, moment decomposition based on the second tensor decomposition mode is performed. Unlike stationary signals, the statistical properties of non-stationary signals change over time, thus requiring a more complex tensor decomposition mode. For example, the product of the core tensor and the multidimensional factor tensor is used as the decomposition criterion, thereby more effectively modeling the multidimensional characteristics of the signal. By capturing the changing patterns of the signal at different time points and determining two types of low-rank moments through moment decomposition, these moments can accurately describe the dynamic characteristics of the non-stationary signal segment.

[0046] Finally, based on the first and second types of low-rank moments, the final low-rank moment is determined. This low-rank moment is obtained by combining the low-rank moments of stationary and non-stationary signal segments. The process of determining the low-rank moment fuses the features of stationary and non-stationary signal segments to obtain a unified low-rank moment representation of the overall signal characteristics. Simultaneously, a cross-mapping relationship is established between low-rank moments, higher-order moments, sparse signals, and the target signal. This cross-mapping relationship is achieved by analyzing the correspondence between low-rank moments and higher-order moments, as well as the correlation between low-rank moments and sparse signals and the target signal, to ensure the integrity and consistency of signal features. Establishing this cross-mapping relationship helps to more accurately trace and understand the original features and modulation scheme of the signal in the subsequent modulation identification process.

[0047] Through the above steps, not only can stationary and non-stationary signals be processed efficiently, but key features in the signal can also be extracted through tensor decomposition. Finally, these features are accurately mapped to the target signal, providing high-quality signal features for subsequent modulation recognition. This process ensures the accuracy and efficiency of signal processing, especially in complex signal environments.

[0048] P30: For the low-rank moment, trigger the modulation recognition layer to perform multi-level classification modulation decision, determine the modulation recognition result, perform signal modulation processing on the target signal, and return to the communication channel for communication transmission management.

[0049] Furthermore, before triggering the modulation recognition layer to perform multi-level classification modulation decisions, in order to construct the modulation recognition layer, step P30 of this application embodiment also includes:

[0050] P31a: Using support vector data description as the judgment processing method, and aiming at coarse classification based on modulation type, a first-layer classification deployment is carried out; P32a: Using deep residual features as the judgment processing method, and aiming at fine-grained decision-making based on modulation guidance, a second-layer classification deployment is carried out; P33a: Based on the cascading of the first-layer classification and the second-layer classification, a modulation recognition layer is determined, wherein the layer classification is scalable.

[0051] It should be understood that modulation identification is based on low-rank moments, and the final modulation identification result is determined through multi-level classification modulation decisions. Before triggering the modulation identification layer to perform multi-level classification modulation decisions, the modulation identification layer must first be constructed to ensure that different types of modulation signals can be accurately identified during processing.

[0052] First, Support Vector Data Description (SVDD) is used to determine the processing method, and the first layer of the modulation recognition layer is deployed for classification. SVDD is an unsupervised learning method based on Support Vector Machines (SVMs) that can effectively classify data and determine its boundaries based on the data distribution. In this step, the features of the target signal are mapped to a high-dimensional space, and the SVDD algorithm is used to perform a coarse classification of the signal, that is, to initially distinguish it based on the modulation type. The classification result at this stage is a coarse classification, and the possible modulation types can be effectively filtered out using support vector machine technology, laying the foundation for subsequent fine-grained recognition.

[0053] Next, using deep residual features as the decision-making method, the second layer of the modulation recognition layer is deployed for classification. In this classification layer, signal features are extracted using a deep residual network (ResNet). Deep residual networks have strong feature extraction capabilities, especially when dealing with complex and nonlinear signals. Residual connections can help avoid information loss and improve the efficiency of feature learning. The goal of the second classification layer is to perform modulation-guided fine-grained decision-making through more refined feature extraction and learning. Through deep residual feature extraction, signals can be classified more accurately, signal type identification can be refined, and higher accuracy can be provided for the final modulation recognition.

[0054] Finally, the modulation recognition layer is determined based on the cascading of the first and second-level classifications. This cascaded structure allows the modulation recognition layer to combine the advantages of coarse classification and fine-grained decision-making, achieving high-precision modulation recognition. The first-level classification provides initial modulation type screening, while the second-level classification performs more detailed recognition based on this. Through this cascaded approach, the modulation recognition layer can effectively handle complex modulation signals, improving the accuracy and robustness of recognition. Furthermore, the system's modulation recognition layer is scalable, meaning that more levels of classification can be flexibly added according to different needs to handle more complex signal recognition tasks.

[0055] Furthermore, to perform multi-level classification modulation decision-making and determine the modulation identification result, step P30 in this embodiment of the application also includes:

[0056] P34: Perform a multidimensional Fourier transform on the low-rank moments to obtain the low-rank spectrum; P35: For the low-rank moments and the low-rank spectrum, trigger the modulation recognition layer, perform a coarse decision on the modulation type under the first-level classification to determine the first-level modulation recognition result, and perform a fine decision on the modulation features under the second-level classification according to the inter-layer cascading to determine the second-level modulation recognition result; P36: Integrate the first-level modulation recognition result and the second-level modulation recognition result to determine the modulation recognition result.

[0057] Optionally, the process of multi-level classification modulation decision-making can be further refined to ensure that the system can accurately identify the modulation mode of the signal.

[0058] First, a multidimensional Fourier transform is performed on the low-rank moments to obtain the low-rank spectrum. The multidimensional Fourier transform is a commonly used signal analysis method that can transform a signal from the time domain to the frequency domain, thereby revealing the signal's frequency characteristics. By performing a Fourier transform on the low-rank moments, the spectral information of the signal can be obtained, thus revealing the signal's frequency domain characteristics. The generation of the low-rank spectrum is based on the key information in the low-rank moments, and through the Fourier transform, the frequency composition of the signal can be analyzed more comprehensively, further improving the accuracy of modulation identification.

[0059] Next, based on the low-rank moments and low-rank spectra, the modulation identification layer is triggered to perform multi-level classification modulation decisions. Specifically, firstly, a coarse decision on the modulation type is made at one classification level to determine the modulation identification result. In this classification level, the modulation type of the signal can be preliminarily judged based on the characteristics of the low-rank moments and low-rank spectra. For example, relatively simple classification algorithms, such as support vector machines (SVM) or decision trees, can be used to broadly filter out possible modulation methods. The coarse decision helps the system quickly narrow down the range of modulation types, providing effective guidance for the next step of detailed classification.

[0060] Subsequently, based on the results of the first-level classification, a finer decision is made on the modulation features under the second-level classification to determine the second-level modulation recognition result. The second-level classification further refines the results based on the first-level classification, mainly by using more complex algorithms (such as deep neural networks, convolutional neural networks, etc.) to perform refined feature extraction and classification of the signal. The second-level classification considers the finer-grained features of the signal, enabling more accurate judgments based on the initial screening, thereby ensuring the accuracy of the recognition process.

[0061] Finally, the modulation identification results from the first and second layers are integrated to determine the final modulation identification result. By integrating the results of the first and second layer classifications, the system combines the advantages of coarse and fine classification to obtain the most accurate modulation identification result. This integration process can use methods such as weighted voting and Bayesian inference to ensure that the output of the multi-level classification fully reflects the contribution of each layer, thereby obtaining an optimal identification result. The final determined modulation identification result accurately reflects the modulation type of the signal, providing reliable technical support for subsequent signal modulation processing and communication transmission management.

[0062] Furthermore, the target signal undergoes signal modulation processing. In this embodiment, step P30 further includes:

[0063] P37: Based on the modulation identification result, the target signal is modulated to determine the modulation target signal; P38: Based on the interaction between the dynamic modulation end and the communication channel, the modulation target signal is returned to the communication channel to perform communication transmission management, wherein when a dynamic modulation command is triggered during the communication transmission process, signal modulation identification processing is performed.

[0064] Specifically, the process of modulating the target signal can be further refined, and the modulated signal can be returned to the communication channel for transmission management to ensure the flexibility and efficiency of the signal in a dynamic modulation environment.

[0065] First, based on the modulation identification results, the target signal is modulated to determine the target signal for modulation. This process is based on the modulation identification results derived from the previous multi-level classification modulation decisions, and the target signal is modulated accordingly. The purpose of modulation processing is to convert the signal into a form suitable for transmission through the communication channel, while ensuring signal integrity and accuracy. The choice of modulation method depends on the modulation identification results. For example, if the identification results indicate that the signal uses amplitude modulation (AM), the target signal is modulated accordingly; if the identification results indicate that the signal uses frequency modulation (FM), the target signal is modulated accordingly. This process ensures that the signal maintains its modulation characteristics during transmission, thereby improving the reliability and efficiency of the communication system.

[0066] Next, based on the interaction between the dynamic modulation end and the communication channel, the modulated target signal is returned to the communication channel, and communication transmission management is executed. Through this process, the modulated target signal is transmitted to the communication channel to continue subsequent data transmission. Furthermore, when a dynamic modulation command is triggered during communication transmission, the system will perform signal modulation identification processing again to ensure that the signal in the communication channel can be adjusted in real time according to changes in the environment. This mechanism enables the system to flexibly respond to changes in channel conditions, improving the stability and reliability of communication.

[0067] During this process, the close interaction between the dynamic modulation end and the communication channel ensures real-time adjustment and optimization of the signal. When channel conditions change or the system detects a problem in transmission, the dynamic modulation command will be triggered, and the system will perform modulation identification processing again to re-evaluate the signal modulation method and modulate the signal according to the new identification result, ensuring that the signal can always adapt to the current channel environment and maximize the performance of the communication system.

[0068] In summary, the embodiments of this application have at least the following technical effects:

[0069] This application significantly improves modulation recognition accuracy through high-order moment analysis and tensor decomposition techniques, especially in complex environments with low signal-to-noise ratios and multiple modulation schemes. Simultaneously, the use of a sparse encoder for signal compression reduces computational complexity and improves processing efficiency. Through a dynamic modulation endpoint-communication channel interaction mechanism, the modulation scheme can be adjusted in real-time according to changes in the channel environment, enhancing the system's robustness and adaptability. Furthermore, combining multi-level classification modulation decision-making with signal modulation processing optimizes signal modulation management, ensuring the optimal modulation scheme under different transmission conditions, further improving the system's transmission efficiency and stability.

[0070] It achieves the technical effect of improving modulation recognition accuracy, reducing computational complexity, and realizing high adaptability to complex channels through high-order moment analysis and multi-level decision-making.

[0071] Example 2 is based on the same inventive concept as the modulation recognition method based on higher-order moments and their spectra in the foregoing examples, such as... Figure 2 As shown, this application provides a modulation recognition system based on higher-order moments and their spectra. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0072] The sparse coding module 11 is used to establish the interaction between the dynamic modulation end and the communication channel, trigger the dynamic modulation command with the communication transmission, and import the target signal into the sparse encoder at the end, perform sparse compression projection, and determine the sparse signal. The sparse encoder follows the equidistant principle.

[0073] Tensor decomposition module 12 is used to determine higher-order moments and trigger the tensor decomposition layer to perform tensor decomposition for the sparse signal, and to determine lower-rank moments. The tensor decomposition includes a first tensor decomposition mode and a second tensor decomposition mode.

[0074] The multi-level classification modulation decision module 13 is used to trigger the modulation recognition layer to perform multi-level classification modulation decision for the low-rank moment, determine the modulation recognition result, perform signal modulation processing on the target signal, and return to the communication channel for communication transmission management; wherein, the cascaded sparse encoder, tensor decomposition layer and modulation recognition layer are built-in plug-ins of the dynamic modulation end.

[0075] Furthermore, the sparse coding module 11 is also used to perform the following steps:

[0076] A sparse coding network is constructed, wherein each grid corresponds to a type of signal point; the sparse coding network is traversed, and coding relationship mining and training are performed grid by grid until N coding grids are determined, where N is the number of signal point types; the N coding grids are concatenated to determine the sparse encoder.

[0077] Furthermore, the sparse coding module 11 is also used to perform the following steps:

[0078] Identify the first grid, and perform dimensionality reduction and compression projection relationship mining on the first signal point type based on the equidistant value of the signal points to determine the first encoding relationship. Here, the first grid is any grid in the sparse coding network. Write the first encoding relationship into the first grid, and perform grid autoencoder supervised training until convergence to determine the first encoding grid.

[0079] Furthermore, the tensor decomposition module 12 is also used to perform the following steps:

[0080] The sparse signal is traversed to locate a first stationary signal segment and a second non-stationary signal segment. For the first stationary signal segment, a first tensor decomposition mode is determined, using the rank tensor as the decomposition criterion. For the second non-stationary signal segment, a second tensor decomposition mode is determined, using the product of the core tensor and the multidimensional factor tensor as the decomposition criterion. Based on the first and second tensor decomposition modes, the higher-order moments are subjected to multi-threaded tensor decomposition processing to determine the lower-rank moments. The tensor decomposition layer is deployed based on programmable gating, and the first and second tensor decomposition modes can be extended through gating programming.

[0081] Furthermore, the tensor decomposition module 12 is also used to perform the following steps:

[0082] For the sparse signal, calculate higher-order moments; based on the higher-order moments corresponding to the first stationary signal segment, perform moment decomposition under the first tensor decomposition mode to determine a class of low-rank moments; based on the higher-order moments corresponding to the second non-stationary signal segment, perform moment decomposition under the second tensor decomposition mode to determine a class of low-rank moments; based on the class of low-rank moments and the class of low-rank moments, determine the low-rank moments, and establish a mutual mapping between the low-rank moments, higher-order moments, sparse signal, and target signal.

[0083] Furthermore, the multi-level classification modulation decision module 13 is also used to perform the following steps:

[0084] The system uses support vector data description as the judgment processing method, and performs a first-layer classification deployment with the goal of coarse classification based on modulation type; it uses deep residual features as the judgment processing method, and performs a second-layer classification deployment with the goal of fine-grained decision-making based on modulation guidance; the modulation recognition layer is determined according to the cascade of the first-layer classification and the second-layer classification, wherein the layer classification is scalable.

[0085] Furthermore, the multi-level classification modulation decision module 13 is also used to perform the following steps:

[0086] A multidimensional Fourier transform is performed on the low-rank moments to obtain the low-rank spectrum. For the low-rank moments and the low-rank spectrum, the modulation recognition layer is triggered to perform a coarse decision on the modulation type under the first-level classification to determine the first-level modulation recognition result. Based on the inter-layer cascading, a fine decision on the modulation features under the second-level classification is performed to determine the second-level modulation recognition result. The first-level modulation recognition result and the second-level modulation recognition result are integrated to determine the modulation recognition result.

[0087] Furthermore, the multi-level classification modulation decision module 13 is also used to perform the following steps:

[0088] Based on the modulation identification result, the target signal is modulated to determine the modulation target signal; based on the interaction between the dynamic modulation end and the communication channel, the modulation target signal is returned to the communication channel to perform communication transmission management, wherein when a dynamic modulation command is triggered during the communication transmission process, signal modulation identification processing is performed.

[0089] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0090] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0091] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A modulation recognition method based on higher-order moments and their spectra, characterized in that, The method includes: An interaction is established between the dynamic modulation end and the communication channel. The dynamic modulation command is triggered during communication transmission, and the target signal is imported into the sparse encoder at the end. Sparse compression projection is performed to determine the sparse signal. The sparse encoder follows the equidistant principle. For the sparse signal, higher-order moments are determined and a tensor decomposition layer is triggered to perform tensor decomposition, and lower-rank moments are determined. The tensor decomposition includes a first tensor decomposition mode and a second tensor decomposition mode. For the low-rank moment, the modulation recognition layer is triggered to perform multi-level classification modulation decision, determine the modulation recognition result, perform signal modulation processing on the target signal, and return to the communication channel for communication transmission management; Among them, a cascaded sparse encoder, tensor decomposition layer and modulation recognition layer are used as built-in plug-ins for the dynamic modulation end; Determine higher-order moments and trigger the tensor decomposition layer for tensor decomposition, determining lower-rank moments, including: Traverse the sparse signal to locate the first stationary signal segment and the second non-stationary signal segment; For the first stable signal segment, a first tensor decomposition mode is determined, and the first tensor decomposition mode uses the rank-one tensor as the decomposition standard. For the second non-stationary signal segment, a second tensor decomposition mode is determined, wherein the second tensor decomposition mode uses the product of the core tensor and the multidimensional factor tensor as the decomposition criterion. Based on the first tensor decomposition mode and the second tensor decomposition mode, the higher-order moments are subjected to multi-threaded tensor decomposition processing to determine the lower-rank moments.

2. The modulation recognition method based on higher-order moments and their spectra as described in claim 1, characterized in that, Before the target signal is fed into the sparse encoder at the end, the construction of the sparse encoder includes: Construct a sparse coding network, where each grid corresponds to a class of signal points; Traverse the sparse coding network, mine and train coding relationships grid by grid until N coding grids are determined, where N is the number of signal point types; The sparse encoder is determined by splicing the N coding grids.

3. The modulation recognition method based on higher-order moments and their spectra as described in claim 2, characterized in that, Encoding relationship mining and training are performed grid-by-grid, including: Identify the first grid, and perform dimensionality reduction and compression projection relationship mining on the first signal point type based on the equidistant value of the signal points to determine the first coding relationship, wherein the first grid is any grid in the sparse coding network; Write the first encoding relationship into the first grid, perform grid autoencoder supervised training until convergence, and determine the first encoding grid.

4. The modulation recognition method based on higher-order moments and their spectra as described in claim 1, characterized in that, Based on the first tensor decomposition mode and the second tensor decomposition mode, multi-threaded tensor decomposition processing is performed on the higher-order moments, including: For the sparse signal, calculate the higher-order moments; Based on the higher-order moments corresponding to the first stationary signal segment, moment decomposition based on the first tensor decomposition mode is performed to determine a class of low-rank moments. Based on the higher-order moments corresponding to the second non-stationary signal segment, moment decomposition based on the second tensor decomposition mode is performed to determine two types of low-rank moments; Based on the first type of low-rank moments and the second type of low-rank moments, the low-rank moments are determined, and the mutual mapping between the low-rank moments, higher-order moments, sparse signals and target signals is established.

5. The modulation recognition method based on higher-order moments and their spectra as described in claim 4, characterized in that, The tensor decomposition layer is deployed based on programmable gating, and the first tensor decomposition mode and the second tensor decomposition mode can be extended through gating programming.

6. The modulation recognition method based on higher-order moments and their spectra as described in claim 1, characterized in that, Before triggering the modulation recognition layer to perform multi-level classification modulation decisions, the construction of the modulation recognition layer includes: The processing method is determined by the support vector data description, and a single-layer classification is deployed with the goal of coarse classification based on modulation type. Using deep residual features as the judgment processing method and modulation-guided fine-grained decision-making as the goal, a two-layer classification deployment is carried out; Based on the cascading of the first-level classification and the second-level classification, a modulation recognition layer is determined, wherein the layer classification is expandable.

7. The modulation recognition method based on higher-order moments and their spectra as described in claim 6, characterized in that, Perform multi-level classification modulation decision-making to determine the modulation identification result, including: The low-rank moments are subjected to a multidimensional Fourier transform to obtain the low-rank spectrum; For the low-rank moments and low-rank spectra, the modulation recognition layer is triggered to perform a coarse decision on the modulation type under the first-level classification to determine the first-level modulation recognition result. Based on the inter-layer cascading, a fine decision on the modulation features under the second-level classification is performed to determine the second-level modulation recognition result. The modulation identification result is determined by integrating the first-layer modulation identification result with the second-layer modulation identification result.

8. The modulation recognition method based on higher-order moments and their spectra as described in claim 1, characterized in that, Based on the modulation recognition result, the target signal is modulated to determine the modulated target signal; Based on the interaction between the dynamic modulation end and the communication channel, the modulated target signal is returned to the communication channel to perform communication transmission management. When a dynamic modulation command is triggered during the communication transmission process, signal modulation identification processing is performed.

9. A modulation recognition system based on higher-order moments and their spectra, characterized in that, The system includes: The sparse coding module is used to establish the interaction between the dynamic modulation end and the communication channel. It triggers the dynamic modulation command during communication transmission and imports the target signal into the sparse encoder at the end to perform sparse compression projection and determine the sparse signal. The sparse encoder follows the equidistant principle. The tensor decomposition module is used to determine higher-order moments and trigger the tensor decomposition layer to perform tensor decomposition for the sparse signal, and to determine lower-rank moments. The tensor decomposition includes a first tensor decomposition mode and a second tensor decomposition mode. A multi-level classification modulation decision module is used to trigger the modulation recognition layer to perform multi-level classification modulation decision for the low-rank moment, determine the modulation recognition result, perform signal modulation processing on the target signal, and return to the communication channel for communication transmission management; wherein, a cascaded sparse encoder, tensor decomposition layer and modulation recognition layer are used as built-in plug-ins for the dynamic modulation end. Furthermore, the tensor decomposition module is also used to perform the following steps: The sparse signal is traversed to locate a first stationary signal segment and a second non-stationary signal segment. For the first stationary signal segment, a first tensor decomposition mode is determined, which uses the rank tensor as the decomposition criterion. For the second non-stationary signal segment, a second tensor decomposition mode is determined, which uses the product of the core tensor and the multidimensional factor tensor as the decomposition criterion. Based on the first and second tensor decomposition modes, the higher-order moments are subjected to multi-threaded tensor decomposition processing to determine the lower-rank moments.

Images