A time-frequency synchronization method and system for wireless network
Through distributed deep learning network and adaptive feature fusion technology, the problem of reduced time-frequency synchronization accuracy and high computational complexity in wireless networks is solved, and the accuracy and robustness of time-frequency synchronization are improved. It is suitable for complex signal environments and Large-scale network.
Patent Information
- Application Number
- CN202311497480.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-11
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2043-11-11
AI Technical Summary
The time-frequency synchronization of existing wireless networks is not stable enough in complex signal propagation environments, resulting in a reduced accuracy of time-frequency synchronization. The number of parameters is large under large-scale networks and complex topological structures, and the calculation complexity is high, making it difficult to utilize feature information.
A distributed deep learning network is used for signal processing and feature extraction, a data set is established and adaptive feature fusion is carried out, a spatial attention module is introduced for feature refinement, reducing the computational complexity, and improving the robustness and practicality of the model.
It improves the accuracy and robustness of time-frequency synchronization in complex environments, reduces the number of model parameters, reduces the computational complexity, and improves the accuracy and automation of signal processing.
Smart Images

Figure CN117528756B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a time-frequency synchronization method and system for wireless networks. Background Art
[0002] Time and frequency synchronization of wireless networks is a method of synchronization using the time and frequency information of physical layer signals, which aims to ensure the consistency of time and frequency between different devices and ensure the correct reception and transmission of data. However, the time and frequency synchronization of existing wireless networks has technical problems such as the instability of traditional time and frequency synchronization algorithms in complex signal propagation environments, resulting in reduced time and frequency synchronization accuracy; and the large number of parameters and high computational complexity make it difficult to use feature information for large-scale networks and wireless networks with complex topologies. Summary of the invention
[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a time-frequency synchronization method and system for wireless networks. In view of the technical problem that the traditional time-frequency synchronization algorithm is not stable enough in a complex signal propagation environment, resulting in reduced time-frequency synchronization accuracy, a data set is established and a distributed deep learning network is used for signal processing and feature extraction to improve the accuracy, automation and real-time performance of signal processing, learn more complex signal patterns, and improve the accuracy of time-frequency synchronization in complex environments; for large-scale networks and wireless networks with complex topology structures, the number of parameters is large, the calculation complexity is high, and it is difficult to utilize feature information. Adaptive feature fusion is used to better utilize the feature information of different nodes, improve the feature expression capability, effectively reduce the number of model parameters, reduce the calculation complexity, and improve the robustness and practicality of the model.
[0004] The technical solution adopted by the present invention is as follows: The present invention provides a time-frequency synchronization method for a wireless network, the method comprising the following steps:
[0005] Step S1: data collection;
[0006] Step S2: data preprocessing;
[0007] Step S3: constructing a time-frequency synchronization model, establishing a data set of received signals, quantifying signal classification criteria and generating a label set, constructing a distributed deep learning network, preprocessing and converting received signals, performing discrete Fourier transform and discrete short-time Fourier transform on baseband IQ signals, and performing feature extraction;
[0008] Step S4: Time-frequency synchronization model training, distributed nodes jointly train the deep learning network, calculate the global loss function and perform adaptive feature fusion, introduce the spatial attention module for adaptive feature refinement, calculate the spatial attention map, multiply it with the fusion feature map, obtain the final output feature map, and then use the fully connected layer to obtain the classification result;
[0009] Step S5: Time-frequency synchronization performance evaluation.
[0010] Furthermore, in step S1, the data collection includes the following steps:
[0011] Step S11: Node selection, determining the node that needs to collect time-frequency synchronization signal data, the node is a key node in the wireless network;
[0012] Step S12: Prepare data acquisition equipment, including equipment for collecting signal data, including acquisition cards and radio frequency receivers, and ensure that the performance and parameter settings of the equipment meet the requirements for collecting time-frequency synchronization signal data;
[0013] Step S13: signal acquisition, installing and configuring data acquisition equipment on the selected node, and starting to acquire time-frequency synchronization signal data within a specific time period, including the following steps:
[0014] Step S131: clock offset recording, recording the clock offset information of each node, and calculating the clock offset by comparing the difference between the local clock of the node and the reference clock;
[0015] Step S132: Propagation delay measurement, using the timestamp method to determine the sending and receiving time of the signal, and using the known synchronization signal sending time and receiving time to calculate the delay of the signal during the propagation process;
[0016] Step S133: Signal amplitude measurement, recording the amplitude information of the signal at the receiving node, and measuring the signal amplitude by the received signal power.
[0017] Furthermore, in step S2, the data preprocessing is specifically processing the collected time-frequency synchronization signal data, including data cleaning, denoising and filtering, to ensure data quality and accuracy.
[0018] Furthermore, in step S3, the time-frequency synchronization model is constructed, including the following steps:
[0019] Step S31: Establishing a data set of received signals, generating random symbol sequences for a variety of data formats and modulation schemes that are not randomly distributed, calculating the minimum symbol power difference and quantifying the signal classification criteria, and generating a label set, including the following steps:
[0020] Step S311: Calculate the minimum symbol power difference, using the following formula:
[0021] ;
[0022] ;
[0023] Where △p is the calculated minimum symbol power difference per frame, N1 is the total symbol length of the data frame, and n a and n b is the length of the corresponding symbol, a represents the ath symbol, and b represents the bth symbol;
[0024] Step S312: quantizing signal classification standards and generating label sets, formulating quantized signal classification standards and generating label sets according to the relationship between the minimum symbol power difference of each frame and the total symbol length of the data frame;
[0025] Step S32: constructing a distributed deep learning network, using a distributed deep learning framework to construct a distributed deep learning network according to the generated data set and label set;
[0026] Step S33: signal preprocessing module, including the following steps:
[0027] Step S331: Processing the received signal through a down-conversion and down-sampling module to reduce the frequency of the signal and the sampling rate, which is helpful for the efficiency of subsequent signal processing and reduces the demand for storage space;
[0028] Step S332: Obtain the down-converted and down-sampled signal, and extract the baseband IQ signal, which contains the real part and the imaginary part of the signal, so as to facilitate spectrum analysis and time domain analysis;
[0029] Step S333: Perform discrete Fourier transform on the baseband IQ signal, convert the time domain signal into a frequency domain signal, obtain a one-dimensional spectrum of the signal, calculate the discrete Fourier transform of the discrete signal, and obtain the frequency response of the signal at different frequencies. The formula used is as follows:
[0030] ;
[0031] In the formula, X[k 1 ] indicates that the signal is at frequency k 1 The frequency response at, x[n] is the signal in the discrete time domain, n is the sample point in time, N is the discrete frequency domain signal X[k 1 ] length, k 1 is the sample point in frequency, j is the imaginary unit, is the frequency k 1 The corresponding complex forms of sine and cosine waves;
[0032] Step S334: Perform discrete short-time Fourier transform on the baseband IQ signal, localize the time domain signal in time and frequency, obtain a two-dimensional time spectrum of the signal, obtain the time-frequency characteristics of the signal, and calculate the discrete short-time Fourier transform of the discrete signal. The formula used is as follows:
[0033] ;
[0034] Where m is the time window, w[mn] is the window function, X[m, k 1 ] describes the change of the frequency response of the signal over the time window m;
[0035] Step S34: The signal feature extraction module converts the signal into preprocessed images in the time domain, frequency domain and time-frequency domain respectively, selects the CNN structure based on Lenet5 to extract the signal features, obtains the output feature map on each distributed branch, extracts more detailed feature changes, and further reduces the number of parameters that need to be calculated.
[0036] Furthermore, in step S4, the time-frequency synchronization model training includes the following steps:
[0037] Step S41: The distributed nodes are trained together. Each distributed node broadcasts a global model of a parameter vector to all other nodes and calculates the loss function of the i-th node. The formula used is as follows:
[0038] ;
[0039] In the formula, is a sample (x k ,y k ), is the input vector, is the output vector, k represents the dataset index, is the size of each node training sample, w is the parameter vector, L i (w) is the loss function of the i-th node;
[0040] Step S42: Calculate the global loss function using the following formula:
[0041] ;
[0042] In the formula, , D s is the sum of the sizes of node training samples, I is the total number of nodes, and L(w) is the global loss function;
[0043] Step S43: Adaptive feature fusion, adaptively fusing the feature map of each node, including the following steps:
[0044] Step S431: The output feature map of the signal feature extraction module on each distributed branch is used as the input of the co-attention module, and the co-attention module is used to fuse the feature map, and the fused feature map input includes the feature map and the attention weight distributed node obtained therefrom;
[0045] Step S432: Fusing feature maps, performing weighted fusion of feature maps and attention weights through a maximum pooling layer, an average pooling layer, and a multi-layer perceptron structure to obtain fused features. The attention weights can be adaptively calculated based on different distributed nodes to highlight important feature information and suppress noise.
[0046] Step S433: introducing a spatial attention module for adaptive feature refinement, the spatial attention module receives the fused feature map as input, sends the feature map to the maximum pooling layer and the average pooling layer respectively, obtains two feature maps after dimensionality reduction, and superimposes the two feature maps after dimensionality reduction into a fused feature map;
[0047] Step S434: applying a convolution layer and an activation function layer with the stacking layer as input to calculate the spatial attention map, and calculating the spatial attention map by applying the convolution layer and the activation function layer;
[0048] Step S435: multiply the spatial attention map and the fusion feature map element by element, adjust the response in the feature map according to the attention degree of different spatial positions, and obtain the final output feature map;
[0049] Step S44: Fully connected layer. There are three fully connected layers at the end of the network. The neurons in the fully connected layer are fully connected to each activation function of the previous layer, converting the feature map learned by the neural network into the final classification result.
[0050] Furthermore, in step S5, the time-frequency synchronization performance evaluation includes the following steps:
[0051] Step S51: time-frequency offset, using root mean square error and mean absolute error as metrics to evaluate the difference between the actual value and the estimated value;
[0052] Step S52: Robustness evaluation, evaluating whether the time-frequency synchronization model can maintain stable performance for different data distributions, different transmission channel conditions and noise environments.
[0053] A time-frequency synchronization system for a wireless network provided by the present invention comprises a data collection module, a data preprocessing module, a time-frequency synchronization model construction module, a time-frequency synchronization model training module and a time-frequency synchronization performance evaluation module;
[0054] The data collection module is specifically used to perform node selection and equipment preparation, and collect time-frequency synchronization signal data;
[0055] The data preprocessing module specifically processes the collected time-frequency synchronization signal data, including data cleaning, denoising and filtering, to ensure data quality and accuracy;
[0056] The time-frequency synchronization model building module specifically establishes a data set of received signals, quantifies signal classification standards and generates a label set, preprocesses and converts the received signals, performs discrete Fourier transform and discrete short-time Fourier transform on the baseband IQ signal, and performs feature extraction;
[0057] The time-frequency synchronization model training module is specifically a distributed node jointly training a deep learning network, calculating a global loss function and performing adaptive feature fusion, introducing a spatial attention module for adaptive feature refinement, calculating a spatial attention map, multiplying it with the fusion feature map, obtaining a final output feature map, and then using a fully connected layer to obtain a classification result;
[0058] The time-frequency synchronization performance evaluation module specifically performs time-frequency offset and robustness evaluation on the time-frequency synchronization performance.
[0059] The beneficial results achieved by the present invention using the above scheme are as follows:
[0060] (1) In order to solve the technical problem that the traditional time-frequency synchronization algorithm is not stable enough in a complex signal propagation environment, resulting in reduced time-frequency synchronization accuracy, a data set is established and a distributed deep learning network is used for signal processing and feature extraction to improve the accuracy, automation and real-time performance of signal processing, learn more complex signal patterns, and improve the accuracy of time-frequency synchronization in complex environments;
[0061] (2) In order to address the technical problems of large-scale wireless networks and complex topology structures, such as large number of parameters, high computational complexity, and difficulty in utilizing feature information, adaptive feature fusion is used to better utilize the feature information of different nodes, improve feature expression capabilities, effectively reduce the number of model parameters, reduce computational complexity, and improve the robustness and practicality of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 A schematic diagram of a flow chart of a time-frequency synchronization method for a wireless network provided by the present invention;
[0063] Figure 2 A schematic diagram of a time-frequency synchronization system for a wireless network provided by the present invention;
[0064] Figure 3 is a schematic flow chart of step S3;
[0065] Figure 4 It is a schematic diagram of the process of step S4.
[0066] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION
[0067] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0068] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.
[0069] Example 1, see Figure 1 The present invention provides a time-frequency synchronization method for a wireless network, the method comprising the following steps:
[0070] Step S1: data collection;
[0071] Step S2: data preprocessing;
[0072] Step S3: constructing a time-frequency synchronization model, establishing a data set of received signals, quantifying signal classification criteria and generating a label set, constructing a distributed deep learning network, preprocessing and converting received signals, performing discrete Fourier transform and discrete short-time Fourier transform on baseband IQ signals, and performing feature extraction;
[0073] Step S4: Time-frequency synchronization model training, distributed nodes jointly train the deep learning network, calculate the global loss function and perform adaptive feature fusion, introduce the spatial attention module for adaptive feature refinement, calculate the spatial attention map, multiply it with the fusion feature map, obtain the final output feature map, and then use the fully connected layer to obtain the classification result;
[0074] Step S5: Time-frequency synchronization performance evaluation.
[0075] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the data collection includes the following steps:
[0076] Step S11: Node selection, determining the node that needs to collect time-frequency synchronization signal data, the node is a key node in the wireless network;
[0077] Step S12: Prepare data acquisition equipment, including equipment for collecting signal data, including acquisition cards and radio frequency receivers, and ensure that the performance and parameter settings of the equipment meet the requirements for collecting time-frequency synchronization signal data;
[0078] Step S13: signal acquisition, installing and configuring data acquisition equipment on the selected node, and starting to acquire time-frequency synchronization signal data within a specific time period, including the following steps:
[0079] Step S131: clock offset recording, recording the clock offset information of each node, and calculating the clock offset by comparing the difference between the local clock of the node and the reference clock;
[0080] Step S132: Propagation delay measurement, using the timestamp method to determine the sending and receiving time of the signal, and using the known synchronization signal sending time and receiving time to calculate the delay of the signal during the propagation process;
[0081] Step S133: Signal amplitude measurement, recording the amplitude information of the signal at the receiving node, and measuring the signal amplitude by the received signal power.
[0082] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, the data preprocessing is specifically processing the collected time-frequency synchronization signal data, including data cleaning, denoising and filtering to ensure data quality and accuracy.
[0083] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S3, the time-frequency synchronization model is constructed, including the following steps:
[0084] Step S31: Establishing a data set of received signals, generating random symbol sequences for a variety of data formats and modulation schemes that are not randomly distributed, calculating the minimum symbol power difference and quantifying the signal classification criteria, and generating a label set, including the following steps:
[0085] Step S311: Calculate the minimum symbol power difference, using the following formula:
[0086] ;
[0087] ;
[0088] Where △p is the calculated minimum symbol power difference per frame, N1 is the total symbol length of the data frame, and n a and nb is the length of the corresponding symbol, a represents the ath symbol, and b represents the bth symbol;
[0089] Step S312: quantizing the signal classification standard and generating a label set, formulating the quantized signal classification standard according to the relationship between the minimum symbol power difference of each frame and the total symbol length of the data frame;
[0090] Step S32: constructing a distributed deep learning network, using a distributed deep learning framework to construct a distributed deep learning network according to the generated data set and label set;
[0091] Step S33: signal preprocessing module, including the following steps:
[0092] Step S331: Processing the received signal through a down-conversion and down-sampling module to reduce the frequency of the signal and the sampling rate, which is helpful for the efficiency of subsequent signal processing and reduces the demand for storage space;
[0093] Step S332: Obtain the down-converted and down-sampled signal, and extract the baseband IQ signal, which contains the real part and the imaginary part of the signal, so as to facilitate spectrum analysis and time domain analysis;
[0094] Step S333: Perform discrete Fourier transform on the baseband IQ signal, convert the time domain signal into a frequency domain signal, obtain a one-dimensional spectrum of the signal, calculate the discrete Fourier transform of the discrete signal, and obtain the frequency response of the signal at different frequencies. The formula used is as follows:
[0095] ;
[0096] In the formula, X[k 1 ] indicates that the signal is at frequency k 1 The frequency response at, x[n] is the signal in the discrete time domain, n is the sample point in time, N is the discrete frequency domain signal X[k 1 ] length, k 1 is the sample point in frequency, j is the imaginary unit, is the frequency k 1 The corresponding complex forms of sine and cosine waves;
[0097] Step S334: Perform discrete short-time Fourier transform on the baseband IQ signal, localize the time domain signal in time and frequency, obtain a two-dimensional time spectrum of the signal, obtain the time-frequency characteristics of the signal, and calculate the discrete short-time Fourier transform of the discrete signal. The formula used is as follows:
[0098] ;
[0099] Where m is the time window, w[mn] is the window function, X[m, k1 ] describes the change of the frequency response of the signal over the time window m;
[0100] Step S34: The signal feature extraction module converts the signal into preprocessed images in the time domain, frequency domain and time-frequency domain respectively, selects the CNN structure based on Lenet5 to extract the signal features, obtains the output feature map on each distributed branch, extracts more detailed feature changes, and further reduces the number of parameters that need to be calculated.
[0101] By performing the above operations, a data set is established and a distributed deep learning network is used for signal processing and feature extraction, thereby improving the accuracy, automation and real-time performance of signal processing, learning more complex signal patterns, and improving the accuracy of time-frequency synchronization in complex environments. This solves the technical problem that the traditional time-frequency synchronization algorithm is not stable enough in complex signal propagation environments, resulting in reduced time-frequency synchronization accuracy.
[0102] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S4, the time-frequency synchronization model training includes the following steps:
[0103] Step S41: The distributed nodes are trained together. Each distributed node broadcasts a global model of a parameter vector to all other nodes and calculates the loss function of the i-th node. The formula used is as follows:
[0104] ;
[0105] In the formula, is a sample (x k ,y k ), is the input vector, is the output vector, k represents the dataset index, is the size of each node training sample, w is the parameter vector, L i (w) is the loss function of the i-th node;
[0106] Step S42: Calculate the global loss function using the following formula:
[0107] ;
[0108] In the formula, , D s is the sum of the sizes of node training samples, I is the total number of nodes, and L(w) is the global loss function;
[0109] Step S43: Adaptive feature fusion, adaptively fusing the feature map of each node, including the following steps:
[0110] Step S431: The output feature map of the signal feature extraction module on each distributed branch is used as the input of the co-attention module, and the co-attention module is used to fuse the feature map, and the fused feature map input includes the feature map and the attention weight distributed node obtained therefrom;
[0111] Step S432: Fusing feature maps, performing weighted fusion of feature maps and attention weights through a maximum pooling layer, an average pooling layer, and a multi-layer perceptron structure to obtain fused features. The attention weights can be adaptively calculated based on different distributed nodes to highlight important feature information and suppress noise.
[0112] Step S433: introducing a spatial attention module for adaptive feature refinement, the spatial attention module receives the fused feature map as input, sends the feature map to the maximum pooling layer and the average pooling layer respectively, obtains two feature maps after dimensionality reduction, and superimposes the two feature maps after dimensionality reduction into a fused feature map;
[0113] Step S434: applying a convolution layer and an activation function layer with the stacking layer as input to calculate the spatial attention map, and calculating the spatial attention map by applying the convolution layer and the activation function layer;
[0114] Step S435: multiply the spatial attention map and the fusion feature map element by element, adjust the response in the feature map according to the attention degree of different spatial positions, and obtain the final output feature map;
[0115] Step S44: Fully connected layer. There are three fully connected layers at the end of the network. The neurons in the fully connected layer are fully connected to each activation function of the previous layer, converting the feature map learned by the neural network into the final classification result.
[0116] By performing the above operations and using adaptive feature fusion, we can better utilize the feature information of different nodes, improve the feature expression ability, effectively reduce the number of model parameters, reduce the computational complexity, improve the robustness and practicality of the model, and solve the technical problems of large number of parameters, high computational complexity and difficulty in utilizing feature information for large-scale networks and wireless networks with complex topology structures.
[0117] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the time-frequency synchronization performance evaluation includes the following steps:
[0118] Step S51: time-frequency offset, using root mean square error and mean absolute error as metrics to evaluate the difference between the actual value and the estimated value;
[0119] Step S52: Robustness evaluation, evaluating whether the time-frequency synchronization model can maintain stable performance for different data distributions, different transmission channel conditions and noise environments.
[0120] Embodiment 7, see Figure 2 , this embodiment is based on the above embodiment, and a time-frequency synchronization system for a wireless network provided by the present invention includes a data collection module, a data preprocessing module, a time-frequency synchronization model construction module, a time-frequency synchronization model training module and a time-frequency synchronization performance evaluation module;
[0121] The data collection module is specifically used to perform node selection and equipment preparation, and collect time-frequency synchronization signal data;
[0122] The data preprocessing module specifically processes the collected time-frequency synchronization signal data, including data cleaning, denoising and filtering, to ensure data quality and accuracy;
[0123] The time-frequency synchronization model building module specifically establishes a data set of received signals, quantifies signal classification standards and generates a label set, preprocesses and converts the received signals, performs discrete Fourier transform and discrete short-time Fourier transform on the baseband IQ signal, and performs feature extraction;
[0124] The time-frequency synchronization model training module is specifically a distributed node jointly training a deep learning network, calculating a global loss function and performing adaptive feature fusion, introducing a spatial attention module for adaptive feature refinement, calculating a spatial attention map, multiplying it with the fusion feature map, obtaining a final output feature map, and then using a fully connected layer to obtain a classification result;
[0125] The time-frequency synchronization performance evaluation module specifically performs time-frequency offset and robustness evaluation on the time-frequency synchronization performance.
[0126] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0127] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
[0128] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.
Claims
1. A time-frequency synchronization method for a wireless network, characterized in that: The method comprises the following steps: Step S1: data collection; Step S2: data preprocessing; Step S3: constructing a time-frequency synchronization model, establishing a data set of received signals, quantifying signal classification criteria and generating a label set, constructing a distributed deep learning network, preprocessing and converting received signals, performing discrete Fourier transform and discrete short-time Fourier transform on baseband IQ signals, and performing feature extraction; Step S4: Time-frequency synchronization model training, distributed nodes jointly train the distributed deep learning network, calculate the global loss function and perform adaptive feature fusion, introduce the spatial attention module for adaptive feature refinement, calculate the spatial attention map, multiply it with the fusion feature map, obtain the final output feature map, and then use the fully connected layer to obtain the classification result; Step S5: time-frequency synchronization performance evaluation; In step S3, the time-frequency synchronization model is constructed, including the following steps: Step S31: Establishing a data set of received signals, generating random symbol sequences for a variety of data formats and modulation schemes that are not randomly distributed, calculating the minimum symbol power difference and quantifying the signal classification criteria, and generating a label set, including the following steps: Step S311: Calculate the minimum symbol power difference, using the following formula: ; ; Where △p is the calculated minimum symbol power difference per frame, N1 is the total symbol length of the data frame, and n a and n b is the length of the corresponding symbol, a represents the ath symbol, and b represents the bth symbol; Step S312: quantizing signal classification standards and generating label sets, formulating quantized signal classification standards and generating label sets according to the relationship between the minimum symbol power difference of each frame and the total symbol length of the data frame; Step S32: constructing a distributed deep learning network, using a distributed deep learning framework to construct a distributed deep learning network according to the generated data set and label set; Step S33: signal preprocessing module, including the following steps: Step S331: Processing the received signal through a down-conversion and down-sampling module; Step S332: Obtain the down-converted and down-sampled signal, and extract the baseband IQ signal; Step S333: Perform discrete Fourier transform on the baseband IQ signal, convert the time domain signal into a frequency domain signal, obtain a one-dimensional spectrum of the signal, calculate the discrete Fourier transform of the discrete signal, and obtain the frequency response of the signal at different frequencies. The formula used is as follows: ; Where X[k1] represents the frequency response of the signal at frequency k1, x[n] is the signal in the discrete time domain, n is the sample point in time, N is the length of the discrete frequency domain signal X[k1], k1 is the sample point in frequency, j is the imaginary unit, It is the complex form of the sine and cosine waves corresponding to frequency k1; Step S334: Perform discrete short-time Fourier transform on the baseband IQ signal, localize the time domain signal in time and frequency, obtain a two-dimensional time spectrum of the signal, obtain the time-frequency characteristics of the signal, and calculate the discrete short-time Fourier transform of the discrete signal. The formula used is as follows: ; Where m is the time window, w[mn] is the window function, and X[m, k1] describes how the frequency response of the signal changes with the time window m; Step S34: a signal feature extraction module converts the signal into preprocessed images in the time domain, frequency domain and time-frequency domain respectively, selects a CNN structure based on Lenet5 to extract signal features, and obtains output feature maps on each distributed branch; In step S4, the time-frequency synchronization model training includes the following steps: Step S41: The distributed nodes are trained together. Each distributed node broadcasts a global model of a parameter vector to all other nodes and calculates the loss function of the i-th node. The formula used is as follows: ; In the formula, is a sample (x k ,y k ), is the input vector, is the output vector, k represents the dataset index, is the size of each node training sample, w is the parameter vector, L i (w) is the loss function of the i-th node; Step S42: Calculate the global loss function using the following formula: ; In the formula, , D s is the sum of the sizes of node training samples, I is the total number of nodes, and L(w) is the global loss function; Step S43: Adaptive feature fusion, adaptively fusing the feature map of each node, including the following steps: Step S431: using the output feature map of the signal feature extraction module on each distributed branch as the input of the co-attention module, the co-attention module is used to fuse the feature map, and the fused feature map input includes the feature map and the attention weight learned from the feature map; Step S432: Fusing feature maps, performing weighted fusion of feature maps and attention weights through a maximum pooling layer, an average pooling layer, and a multi-layer perceptron structure to obtain fused features, and the attention weights are adaptively calculated according to different distributed nodes; Step S433: introducing a spatial attention module for adaptive feature refinement, the spatial attention module receives the fused feature map as input, sends the feature map to the maximum pooling layer and the average pooling layer respectively, obtains two feature maps after dimensionality reduction, and superimposes the two feature maps after dimensionality reduction into a fused feature map; Step S434: applying a convolution layer and an activation function layer with the stacking layer as input to calculate the spatial attention map, and calculating the spatial attention map by applying the convolution layer and the activation function layer; Step S435: multiply the spatial attention map and the fusion feature map element by element, adjust the response in the feature map according to the attention degree of different spatial positions, and obtain the final output feature map; Step S44: Fully connected layer. There are three fully connected layers at the end of the network. The neurons in the fully connected layer are fully connected to each activation function of the previous layer, converting the feature map learned by the neural network into the final classification result.
2. A time-frequency synchronization method for a wireless network according to claim 1, characterized in that: In step S5, the time-frequency synchronization performance evaluation includes the following steps: Step S51: time-frequency offset, using root mean square error and mean absolute error as metrics to evaluate the difference between the actual value and the estimated value; Step S52: Robustness evaluation, evaluating whether the time-frequency synchronization model can maintain stable performance for different data distributions, different transmission channel conditions and noise environments.
3. The method for time-frequency synchronization for a wireless network according to claim 1, characterized in that: In step S1, the data collection includes the following steps: Step S11: Node selection, determining the nodes that need to collect time-frequency synchronization signal data; Step S12: preparing data acquisition equipment, including equipment for acquiring signal data, including an acquisition card and a radio frequency receiver; Step S13: signal acquisition, installing and configuring data acquisition equipment on the selected node, and starting to acquire time-frequency synchronization signal data within a specific time period, including the following steps: Step S131: clock offset recording, recording the clock offset information of each node, and calculating the clock offset by comparing the difference between the local clock of the node and the reference clock; Step S132: Propagation delay measurement, using the timestamp method to determine the sending and receiving time of the signal, and using the known synchronization signal sending time and receiving time to calculate the delay of the signal during the propagation process; Step S133: Signal amplitude measurement, recording the amplitude information of the signal at the receiving node, and measuring the amplitude of the signal by the power of the received signal.
4. The method for time-frequency synchronization for a wireless network according to claim 1, characterized in that: In step S2, the data preprocessing is specifically processing the collected time-frequency synchronization signal data, including data cleaning, denoising and filtering.
5. A time-frequency synchronization system for a wireless network, used to implement a time-frequency synchronization method for a wireless network as claimed in any one of claims 1 to 4, characterized in that: It includes a data collection module, a data preprocessing module, a time-frequency synchronization model building module, a time-frequency synchronization model training module and a time-frequency synchronization performance evaluation module.
6. A time-frequency synchronization system for a wireless network according to claim 5, characterized in that: The data collection module is specifically used to perform node selection and equipment preparation, and collect time-frequency synchronization signal data; The data preprocessing module specifically processes the collected time-frequency synchronization signal data, including data cleaning, denoising and filtering; The time-frequency synchronization model construction module specifically establishes a data set of received signals, quantifies signal classification standards and generates a label set, builds a distributed deep learning network, preprocesses and converts received signals, performs discrete Fourier transform and discrete short-time Fourier transform on baseband IQ signals, and performs feature extraction; The time-frequency synchronization model training module is specifically a distributed node jointly training a deep learning network, calculating a global loss function and performing adaptive feature fusion, introducing a spatial attention module for adaptive feature refinement, calculating a spatial attention map, multiplying it with the fusion feature map, obtaining a final output feature map, and then using a fully connected layer to obtain a classification result; The time-frequency synchronization performance evaluation module specifically performs time-frequency offset and robustness evaluation on the time-frequency synchronization performance.
