A high-precision estimation method for single-burst multidimensional parameters based on deep learning

Abstract

The present invention provides a high-precision single-burst multi-dimensional parameter estimation method based on deep learning, which belongs to the field of radio signal processing technology and solves the problem of low accuracy of existing single-burst time / frequency / space / energy multi-dimensional parameter estimation. The method includes: constructing a data set of corresponding dimensions based on the data features of signal samples in each dimension; constructing a corresponding specific parameter estimation model for the data set of each dimension; and training the corresponding specific parameter estimation model using the data set of each dimension respectively; using the trained specific parameter estimation model of each dimension as a branch, connecting the feature extraction module and parameter estimation module of each branch using an attention module to construct a multi-branch deep estimation model, and fine-tuning the multi-branch deep estimation model using the multi-dimensional data set; inputting the data features of the signal to be estimated in different dimensions into the multi-branch deep estimation model, and predicting and outputting the multi-dimensional parameter estimation results of time / frequency / space / energy.

Description

Technical Field

[0001] The present invention relates to the technical field of radio signal processing, and in particular to a single-burst multi-dimensional parameter high-precision estimation method based on deep learning. Background Art

[0002] In the field of modern communications, signal parameter estimation is the basis and key step for radio signal analysis and processing. This technology analyzes the received signal and accurately infers the key characteristic parameters of the signal, such as the direction of the incoming wave, carrier frequency deviation, and signal-to-noise ratio, providing a basis for subsequent signal processing and demodulation.

[0003] Traditional parameter estimation methods typically target specific parameters and signal types and are susceptible to factors such as noise and multipath fading. They are unable to meet the high-precision and high-efficiency requirements for parameter estimation in complex electromagnetic environments. With the development of artificial intelligence (AI), deep learning technology has demonstrated superior performance in signal processing. It can automatically extract deep signal features, overcoming the limitations of traditional parameter estimation methods in complex environments and becoming an effective approach to high-precision parameter estimation.

[0004] However, existing deep learning-based signal parameter estimation techniques typically rely on single feature input or single-task learning, lacking modeling of inter-feature correlations. This makes it difficult to fully exploit the multidimensional characteristics of signals in the time, frequency, spatial, and energy domains, limiting the accuracy of parameter estimation. Therefore, addressing the low accuracy of parameter estimation in complex electromagnetic environments and the need for traditional estimation methods to be designed for specific parameters and signal types, achieving high-precision, unified estimation of the multidimensional parameters of a single burst—time, frequency, spatial, and energy—is an urgent challenge. Summary of the Invention

[0005] In view of the above analysis, an embodiment of the present invention aims to provide a high-precision estimation method for single-burst multi-dimensional parameters based on deep learning, so as to solve the problem of low accuracy of existing single-burst time / frequency / space / energy multi-dimensional parameter estimation.

[0006] The present invention discloses a high-precision estimation method for single-burst multidimensional parameters based on deep learning, the method comprising:

[0007] According to the data characteristics of the signal samples in each dimension, a data set of the corresponding dimension is constructed;

[0008] For each dimension of the data set, a corresponding specific parameter estimation model is constructed; and the corresponding specific parameter estimation model is trained using the data set of each dimension respectively; wherein the specific parameter estimation model is composed of a feature extraction module and a parameter estimation module;

[0009] The trained parameter estimation model for each dimension is used as a branch. The attention module is used to connect the feature extraction module and the parameter estimation module of each branch to build a multi-branch depth estimation model. The multi-dimensional dataset is then used to fine-tune the multi-branch depth estimation model.

[0010] The data features of the newly received signal to be estimated in different dimensions are input into the trained multi-branch depth estimation model, and the multi-branch depth estimation model predicts and outputs the multi-dimensional parameter estimation results of time / frequency / space / energy.

[0011] On the basis of the above solution, the present invention also makes the following improvements:

[0012] Furthermore, the datasets of each dimension are used to train the corresponding specific parameter estimation model, and the following are performed:

[0013] The specific parameter estimation model is trained using the data features in the data set of the corresponding dimension as the input of the feature extraction model in the specific parameter estimation model, and the corresponding labels as the predicted output of the parameter estimation model in the specific parameter estimation model to obtain a trained specific parameter estimation model.

[0014] Furthermore, the feature extraction module and the parameter estimation module are connected in sequence, the feature extraction module is used to extract deep features of data features, and the parameter estimation module is used to estimate signal parameters from the extracted deep features.

[0015] Furthermore, a multi-branch depth estimation model is constructed as shown, and the following is executed:

[0016] The specific parameter estimation model of each dimension is regarded as a branch respectively. The feature vectors output by the feature extraction modules of each branch are added and input into the attention module. The vectors output by the attention module are input into the parameter estimation modules of each branch respectively to construct a multi-branch depth estimation model.

[0017] Furthermore, the multi-dimensional dataset is used to fine-tune the multi-branch depth estimation model, performing:

[0018] Migrating model parameters of the specific parameter estimation model of each dimension to the multi-branch depth estimation model;

[0019] Use multi-dimensional datasets to fine-tune the network parameters of the multi-branch depth estimation model.

[0020] Furthermore, the multi-branch depth estimation model predicts and outputs the multi-dimensional parameter estimation results of time / frequency / space / energy, and executes:

[0021] Input the data features of each dimension into the feature extraction module of the corresponding dimension in the trained multi-branch depth estimation model, and the parameter estimation module of the corresponding dimension in the multi-branch depth estimation model predicts and outputs the parameter prediction result of the corresponding dimension;

[0022] By combining the parameter prediction results of different dimensions, we can obtain the multi-dimensional parameter estimation results of time / frequency / space / energy.

[0023] Furthermore, the prediction parameters of each dimension are: signal direction, signal start time and duration, signal carrier frequency deviation, signal-to-noise ratio and signal bandwidth; among them,

[0024] The signal start time and duration are time domain parameters, the signal carrier frequency deviation and signal bandwidth are frequency domain parameters, the signal arrival direction is a spatial domain parameter, and the signal-to-noise ratio is an energy domain parameter.

[0025] Furthermore, the constructed data sets of various dimensions include: signal IQ data set, signal amplitude data set, signal phase data set, signal covariance eigenvalue data set and signal power spectrum data set.

[0026] Further, construct data sets of various dimensions and execute:

[0027] The IQ matrix of each signal sample is used as the data feature and the signal direction as the corresponding label to construct the signal IQ dataset;

[0028] The instantaneous amplitude of each signal sample is used as the data feature, and the signal start time and duration are used as labels to construct a signal amplitude dataset;

[0029] The instantaneous phase of each signal sample is used as the data feature and the signal carrier frequency deviation is used as the label to construct a signal phase dataset;

[0030] The eigenvalue matrix corresponding to the covariance matrix of each signal sample is used as the data feature and the signal-to-noise ratio is used as the label to construct a signal covariance eigenvalue dataset;

[0031] The power spectrum matrix constructed by the power spectrum of each signal sample is used as the data feature and the signal bandwidth is used as the label to construct a signal power spectrum dataset.

[0032] Furthermore, the signal samples are determined as follows:

[0033] A multi-antenna reception method or simulation method is used to obtain radio signals with different modulation styles and frequencies, and the burst segments are extracted and processed to obtain signal samples with fixed length.

[0034] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0035] (1) The present invention adopts multiple forms of input and fully integrates the key characteristics of the signal in time, frequency, space, and energy multi-dimensional space, achieving unified high-precision estimation of single burst multi-dimensional parameters, significantly improving the estimation accuracy and system robustness.

[0036] (2) The present invention proposes a high-precision estimation method for single-burst multi-dimensional parameters based on deep learning. It uses a neural network to learn deep features from signal data, overcoming the problem that traditional estimation methods estimate specific parameters and signal types and are easily affected by factors such as noise and multipath fading, and realizes high-precision unified estimation of time / frequency / space / energy multi-dimensional parameters.

[0037] (3) The present invention adopts a specific parameter estimation model weight migration method, which effectively reduces the training time and difficulty of the multi-branch depth estimation model.

[0038] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. Throughout the drawings, the same reference symbols denote the same components.

[0040] Figure 1 A flowchart of a single-burst multi-dimensional parameter high-precision estimation method based on deep learning provided by an embodiment of the present invention;

[0041] Figure 2 This is a structural diagram of a multi-branch depth estimation model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0043] The specific embodiment of the present invention discloses a single burst multi-dimensional parameter high-precision estimation method based on deep learning, the flow chart of the method is as follows Figure 1 As shown, the method includes the following steps.

[0044] Step S1: Acquire radio signals of different modulation patterns and frequencies, extract burst segments, and process them to obtain signal samples of fixed length.

[0045] Preferably, in this embodiment, a multi-antenna reception method is used to repeatedly collect radio signals with different modulation patterns and frequencies, or a simulation method that matches the multi-antenna reception method is used to obtain radio signals with different modulation patterns and frequencies. The multi-antenna reception method is used to obtain the direction of arrival of the radio signal, and the collection of different modulation patterns and frequencies is used to change the modulation pattern and frequency during signal transmission, thereby increasing the diversity of signal samples.

[0046] Typically, the acquired radio signal contains multiple bursts. Each burst is extracted from the original continuous radio signal containing multiple bursts. Zero padding is then performed on each burst to obtain a fixed-length signal sample. Therefore, each signal sample contains a complete burst.

[0047] Step S2: Construct a data set of the corresponding dimension based on the data features of the signal sample in each dimension.

[0048] Preferably, in this embodiment, five dimensional data sets are set, including: a signal IQ data set, a signal amplitude data set, a signal phase data set, a signal covariance eigenvalue data set, and a signal power spectrum data set.

[0049] The signal direction, signal start time and duration, signal carrier frequency deviation, signal-to-noise ratio and bandwidth of each signal sample are obtained respectively to construct each data set.

[0050] (1) Construction of signal IQ dataset

[0051] Extract the IQ components of each signal sample separately to form an IQ matrix; use the IQ matrix of each signal sample as the data feature and the signal direction as the corresponding label to construct a signal IQ data set .

[0052] For example, the IQ matrix of the signal samples Expressed as:

[0053] (1)

[0054] in, 、 Represents the first The first of the antennas I component and Q component of each sampling point; , Indicates the number of antennas; , The number of sampling points representing the signal samples.

[0055] Therefore, the signal IQ dataset It can be expressed as:

[0056] (2)

[0057] in, 、 Respectively represent The IQ matrix of the signal samples, the direction of the signal wave, is the number of signal samples.

[0058] (2) Construction of signal amplitude dataset

[0059] The instantaneous amplitude of each signal sample is used as the data feature, and the signal start time and duration are used as labels to construct a signal amplitude dataset. .

[0060] The instantaneous amplitude of the signal sample Expressed as:

[0061] (3)

[0062] in, Indicates the number of signal samples sampling points.

[0063] Signal Amplitude Dataset Expressed as:

[0064] (4)

[0065] in, Indicates the The instantaneous amplitude of a signal sample, 、 Respectively represent The signal start time and duration of each signal sample.

[0066] (3) Construction of signal phase dataset

[0067] The instantaneous phase of each signal sample is used as the data feature and the signal carrier frequency deviation is used as the label to construct the signal phase dataset. .

[0068] The instantaneous phase of the signal sample Expressed as:

[0069] (5)

[0070] Signal Phase Dataset Expressed as:

[0071] (6)

[0072] in, 、 Respectively represent The instantaneous phase of a signal sample and the signal carrier frequency deviation.

[0073] (4) Construction of signal covariance eigenvalue dataset

[0074] The eigenvalue matrix corresponding to the covariance matrix of each signal sample (i.e., the eigenvalue of the covariance matrix) is used as the data feature and the signal-to-noise ratio is used as the label to construct the signal covariance eigenvalue dataset. .

[0075] By length The signal samples are shifted, intercepted and stacked point by point to obtain a matrix , expressed as:

[0076] (7)

[0077] Calculate the sample covariance matrix , expressed as:

[0078] (8)

[0079] right Perform eigenvalue decomposition to obtain the eigenvalue matrix .

[0080] Signal covariance eigenvalue dataset Expressed as:

[0081] (9)

[0082] in, 、 Respectively represent The eigenvalue matrix and signal-to-noise ratio corresponding to the covariance matrix of the signal samples.

[0083] (5) Construction of signal power spectrum dataset

[0084] The power spectrum matrix constructed by the power spectrum of each signal sample is used as the data feature and the signal bandwidth is used as the label to construct the signal power spectrum dataset. .

[0085] The signal sample The signal of the sampling point Power spectrum Expressed as:

[0086] (10)

[0087] in, is the Fourier transform. Extract The real and imaginary parts of are used to construct the power spectrum matrix, which is expressed as:

[0088] (11)

[0089] Signal power spectrum dataset Expressed as:

[0090] (12)

[0091] in, 、 Respectively represent The power spectrum matrix of the signal samples and the signal bandwidth.

[0092] Step S3: for each dimension of the data set, construct a corresponding specific parameter estimation model; and use the data set of each dimension to train the corresponding specific parameter estimation model respectively; wherein the specific parameter estimation model consists of a feature extraction module and a parameter estimation module.

[0093] In a specific parameter estimation model, the feature extraction module and the parameter estimation module are connected in sequence. The feature extraction module is used to extract deep features of the data features, and the parameter estimation module is used to estimate specific signal parameters from the extracted deep features. During specific implementation, an adapted specific parameter estimation model can be selected based on the characteristics of the dataset in each dimension. For example, the network structure of the feature extraction module and parameter estimation module in a certain dimension can be designed according to the method in Table 1.

[0094] Table 1 Network structure in feature extraction module and parameter estimation module

[0095]

[0096] As shown in Table 1, the feature extraction module includes a convolution layer Conv, a maximum pooling layer MaxPooling and two residual blocks Resblock. 、 Represent the size of the convolution kernel and pooling window respectively, Adjust according to the dimension of the input data, represents the step length, Indicates the number of convolution kernels. The parameter estimation module is composed of three cascaded fully connected layers (FC) with different numbers of neurons. It should be noted that this is just a simple example of an estimation model. In actual applications, the convolutional layer of the feature extraction module should also be followed by a batch normalization layer and an activation function.

[0097] Corresponding to the five-dimensional data set constructed in step S2, in this step, the five-dimensional specific parameter estimation models constructed are respectively expressed as: a specific parameter estimation model for the wave direction estimation dimension, a specific parameter estimation model for the start and duration estimation dimension, a specific parameter estimation model for the carrier frequency offset estimation dimension, a specific parameter estimation model for the signal-to-noise ratio estimation dimension, and a specific parameter estimation model for the bandwidth estimation dimension. Therefore, using the data set As a dataset for training a specific parameter estimation model for the incoming wave direction estimation dimension, the dataset As a dataset for training specific parameter estimation models for the start and duration estimation dimensions, the dataset As a specific parameter estimation model for training carrier frequency offset estimation dimension, the dataset As a dataset for training a specific parameter estimation model for the signal-to-noise ratio estimation dimension, the dataset A dataset used to train a specific parameter estimation model for bandwidth estimation dimensions.

[0098] Specifically, in the process of separately training the parameter estimation model of each dimension, the data features in the data set of the corresponding dimension are used as the input of the feature extraction module in the specific parameter estimation model, and the corresponding labels are used as the predicted output of the parameter estimation module in the specific parameter estimation model. The specific parameter estimation model is trained to obtain a trained parameter estimation model.

[0099] For example, the mean square error can be used as the loss function. The result of signal sample prediction output by parameter estimation model is: , the corresponding true label is , at this time, the loss function Expressed as:

[0100] (13)

[0101] When the training loss reaches the minimum, the parameter estimation model training is completed, and the model parameters of the specific parameter estimation model that has completed the training are saved.

[0102] Step S4: Take the trained specific parameter estimation model of each dimension as a branch, use the attention module to connect the feature extraction module of each branch with the parameter estimation network, build a multi-branch depth estimation model, and use the multi-dimensional data set to fine-tune the multi-branch depth estimation model.

[0103] In this embodiment, after the specific parameter estimation model of each dimension is trained, the specific parameter estimation model of each dimension is respectively used as a branch (i.e., one dimension corresponds to one branch), and the feature vectors output by the feature extraction module of each branch are added and input into the attention module. The vectors output by the attention module are respectively input into the parameter estimation module of each branch to realize parameter estimation of multiple dimensions. The constructed multi-branch depth estimation model is as follows: Figure 2 As shown in the figure. In the constructed multi-branch depth estimation model, the model parameters of the specific parameter estimation model of each dimension are also migrated to the multi-branch depth estimation model. That is, based on the correspondence between different branches in the multi-branch depth estimation model and the specific parameter estimation model, the model parameters of the specific parameter estimation model saved after training are imported.

[0104] Preferably, in this embodiment, the feature vectors output by feature extraction models of different dimensions have the same dimension, and the feature vectors of different dimensions are summed up by matrix addition.

[0105] More specifically, the feature extraction module of all branches can be called a multi-branch feature extraction module. Figure 2 The multi-branch feature extraction module includes five branches. Branches 1, 2, 3, 4, and 5 have the same architecture as the feature extraction modules in the specific parameter estimation model of the arrival direction estimation dimension, the specific parameter estimation model of the start and duration estimation dimension, the specific parameter estimation model of the carrier frequency offset estimation dimension, the specific parameter estimation model of the signal-to-noise ratio estimation dimension, and the specific parameter estimation model of the bandwidth estimation dimension, respectively. They correspond to five different training data: the IQ matrix, the instantaneous amplitude, the instantaneous phase, the eigenvalue matrix corresponding to the covariance matrix, and the power spectrum matrix. In addition, the parameter estimation module of all branches can also be called a multi-branch parameter estimation module, see Figure 2 The multi-branch parameter estimation module also includes five branches. Each branch has the same architecture as the parameter estimation module of the above five specific parameter estimation models (the structure and parameters of each layer are the same), and outputs the incoming wave direction, start and duration, carrier frequency deviation, signal-to-noise ratio and bandwidth estimation results respectively.

[0106] In this embodiment, the attention module adopts the QKV attention mechanism, assuming that the feature vector output by the multi-branch feature extraction module is , Indicates the batch size, represents the sequence length, Represents the dimension of the feature vector, through three sets of trainable weight matrices , , Performing linear transformation yields:

[0107] (14)

[0108] in, 、 、 , is the dimension of attention calculation. The attention score is calculated as:

[0109] (15)

[0110] The vector output of the attention module is represented as:

[0111] (16)

[0112] The attention module dynamically assigns importance weights to different features, enabling the model to focus on the most critical features across different branches. The attention module will be trained further in subsequent fine-tuning training. The output vectors from the attention module are split into five paths and fed into the parameter estimation modules for each of the five dimensions in the multi-branch depth estimation model.

[0113] Preferably, in this embodiment, a multi-dimensional data set is used to fine-tune the multi-branch depth estimation model, performing:

[0114] Step S41: Migrating the model parameters of the specific parameter estimation model of each dimension to the multi-branch depth estimation model.

[0115] Step S42: fine-tune the network parameters of the multi-branch depth estimation model using the multi-dimensional data set.

[0116] Specifically, the five constructed datasets are used to fine-tune the multi-branch depth estimation model, and the weighted mean square error is used as the loss function. 、 、 、 and Five data sets are used to fine-tune the multi-branch depth estimation model. The mean square error of different branches is , , then the loss function of the multi-branch depth estimation model is Expressed as:

[0117] (17)

[0118] in, 、 、 、 、 Represents the weights of the five dimensions.

[0119] Step S5: Input the data features of the newly received signal to be estimated in different dimensions into the trained multi-branch depth estimation model, and the multi-branch depth estimation model predicts and outputs the multi-dimensional parameter estimation results of time / frequency / space / energy.

[0120] Specifically, the parameter estimation modules of different dimensions in the multi-branch depth estimation model can respectively predict the output signal direction, signal start time and duration, signal carrier frequency offset, signal-to-noise ratio, and signal bandwidth. Among them, signal start time and duration are time-domain parameters, signal carrier frequency offset and signal bandwidth are frequency-domain parameters, signal direction is a spatial-domain parameter, and signal-to-noise ratio is an energy-domain parameter.

[0121] Specifically, step S5 is implemented in the following manner.

[0122] Step S51: For the newly received signal to be estimated, according to the construction method of the data set of each dimension, the data features of each dimension are obtained respectively, including: IQ matrix, instantaneous amplitude, instantaneous phase, eigenvalue matrix corresponding to the covariance matrix, and power spectrum matrix.

[0123] Step S52: The data features of each dimension are input into the feature extraction module of the corresponding dimension in the trained multi-branch depth estimation model, and the parameter estimation module of the corresponding dimension in the multi-branch depth estimation model predicts and outputs the parameter estimation results of the corresponding dimension; the parameter estimation results of different dimensions are combined to obtain the multi-dimensional parameter estimation results of time / frequency / space / energy.

[0124] In summary, the single-burst multi-dimensional parameter high-precision estimation method based on deep learning provided by this embodiment collects signal data of different modulation styles and frequencies by using multi-antenna reception, extracts signal bursts and processes them, constructs data sets of different dimensions and corresponding specific parameter estimation models, and uses the corresponding data sets for separate training. Based on the specific parameter estimation model, a multi-dimensional parameter estimation model with a multi-branch structure is constructed, and the multi-branch depth estimation model is fine-tuned and trained using weight migration. By introducing a multi-branch depth estimation model, this embodiment can fuse the features extracted by the specific models of each dimension, alleviate the information isolation that exists when the specific parameter estimation models of each dimension are trained separately, and improve the generalization ability. At the same time, it can avoid the problem that the specific parameter estimation models of each dimension need to be loaded and inferred independently during the inference phase, thereby simplifying the deployment process. Afterwards, the signal data to be estimated is input into the trained multi-branch depth estimation model to obtain the multi-dimensional parameter estimation result. The solution of the embodiment of the present invention uses a neural network to learn the deep features of the signal, which can achieve unified and high-precision estimation of the signal's time / frequency / space / energy multi-dimensional parameters.

[0125] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0126] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A high-precision single-burst multidimensional parameter estimation method based on deep learning, characterized in that: The method comprises: According to the data characteristics of the signal samples in each dimension, a data set of the corresponding dimension is constructed; For each dimension of the data set, a corresponding specific parameter estimation model is constructed; and the corresponding specific parameter estimation model is trained using the data set of each dimension respectively; wherein the specific parameter estimation model is composed of a feature extraction module and a parameter estimation module; The trained parameter estimation model for each dimension is used as a branch. The feature vectors output by the feature extraction modules of each branch are added and input into the attention module. The vectors output by the attention module are input into the parameter estimation modules of each branch respectively to construct a multi-branch depth estimation model. The multi-dimensional dataset is then used to fine-tune the multi-branch depth estimation model and train the attention module. The data features of the newly received signal to be estimated in different dimensions are input into the trained multi-branch depth estimation model, and the multi-branch depth estimation model predicts and outputs the multi-dimensional parameter estimation results of time / frequency / space / energy.

2. The single burst multi-dimensional parameter high-precision estimation method based on deep learning according to claim 1 is characterized in that The datasets of each dimension are used to train the corresponding specific parameter estimation model, and the following is performed: The specific parameter estimation model is trained using the data features in the data set of the corresponding dimension as the input of the feature extraction model in the specific parameter estimation model, and the corresponding labels as the predicted output of the parameter estimation model in the specific parameter estimation model to obtain a trained specific parameter estimation model.

3. The single burst multi-dimensional parameter high-precision estimation method based on deep learning according to claim 2 is characterized in that The feature extraction module and the parameter estimation module are connected in sequence. The feature extraction module is used to extract deep features of data features, and the parameter estimation module is used to estimate signal parameters from the extracted deep features.

4. The single burst multi-dimensional parameter high-precision estimation method based on deep learning according to any one of claims 1 to 3, characterized in that: The multi-dimensional dataset is used to fine-tune the multi-branch depth estimation model, performing: Migrating model parameters of the specific parameter estimation model of each dimension to the multi-branch depth estimation model; Use multi-dimensional datasets to fine-tune the network parameters of the multi-branch depth estimation model.

5. The single burst multi-dimensional parameter high-precision estimation method based on deep learning according to claim 4 is characterized in that: The multi-branch depth estimation model predicts and outputs the multi-dimensional parameter estimation results of time / frequency / space / energy, and executes: Input the data features of each dimension into the feature extraction module of the corresponding dimension in the trained multi-branch depth estimation model, and the parameter estimation module of the corresponding dimension in the multi-branch depth estimation model predicts and outputs the parameter prediction result of the corresponding dimension; By combining the parameter prediction results of different dimensions, we can obtain the multi-dimensional parameter estimation results of time / frequency / space / energy.

6. The single burst multi-dimensional parameter high-precision estimation method based on deep learning according to claim 5 is characterized in that: The prediction parameters of each dimension are: signal direction, signal start time and duration, signal carrier frequency deviation, signal-to-noise ratio and signal bandwidth; among them, The signal start time and duration are time domain parameters, the signal carrier frequency deviation and signal bandwidth are frequency domain parameters, the signal arrival direction is a spatial domain parameter, and the signal-to-noise ratio is an energy domain parameter.

7. The single burst multi-dimensional parameter high-precision estimation method based on deep learning according to claim 6 is characterized in that: The constructed datasets of various dimensions include: signal IQ dataset, signal amplitude dataset, signal phase dataset, signal covariance eigenvalue dataset and signal power spectrum dataset.

8. The single burst multi-dimensional parameter high-precision estimation method based on deep learning according to claim 7 is characterized in that: Build data sets of various dimensions and execute: The IQ matrix of each signal sample is used as the data feature and the signal direction as the corresponding label to construct the signal IQ dataset; The instantaneous amplitude of each signal sample is used as the data feature, and the signal start time and duration are used as labels to construct a signal amplitude dataset; The instantaneous phase of each signal sample is used as the data feature and the signal carrier frequency deviation is used as the label to construct a signal phase dataset; The eigenvalue matrix corresponding to the covariance matrix of each signal sample is used as the data feature and the signal-to-noise ratio is used as the label to construct a signal covariance eigenvalue dataset; The power spectrum matrix constructed by the power spectrum of each signal sample is used as the data feature and the signal bandwidth is used as the label to construct a signal power spectrum dataset.

9. The single burst multi-dimensional parameter high-precision estimation method based on deep learning according to claim 1 is characterized in that The signal samples are determined as follows: A multi-antenna reception method or simulation method is used to obtain radio signals with different modulation styles and frequencies, and the burst segments are extracted and processed to obtain signal samples with fixed length.

Images