Signal parameter estimation method under non-cooperative condition based on deep learning

Abstract

The invention discloses a deep learning-based signal parameter estimation method under a non-cooperative condition, and belongs to the technical field of information. According to the invention, a convolutional neural network is combined with a Transform network, a parameter estimation algorithm model of deep learning is provided, and parameter estimation of a modulation signal is completed; according to the architecture, local detail features of signals can be extracted, global features can also be captured, a full connection layer outputs a final parameter estimation value, and the signal parameter estimation effect is effectively improved; a multi-head attention mechanism module is introduced, signals can be expressed in multiple spaces, fine-grained information and global features of input signals are fully extracted, and deviation possibly caused by only one self-attention head structure is effectively balanced; the signal parameter estimation based on the deep learning network not only can reduce the workload of manual feature extraction, but also can obtain the same or even more excellent result as the manual feature extraction, so that the signal parameter estimation method is intelligent and precise.

Description

Technical Field

[0001] The present invention belongs to the field of information technology, and in particular relates to a signal parameter estimation method under non-collaborative conditions based on deep learning. Background Art

[0002] With the rapid development of electromagnetic signal processing methods and processing platforms, various platforms and instruments are embedded with information processing systems, making electromagnetic signals diverse, dense and complex. If the signals and parameters of both sides can be effectively acquired and accurately identified, it is of great significance to seize power control. With deep learning gradually emerging in various research fields, signal modulation identification and parameter estimation based on deep learning have become the latest mainstream research content in this field. Under non-cooperative conditions, the electromagnetic signal received by the receiver usually does not carry any prior information. In order to correctly demodulate the received modulated signal, in addition to identifying the modulation type of the signal, it is more important to obtain its parameter information. In general, the modulated signal to be tested is obtained by modulating the carrier through the original baseband signal. Therefore, after identifying the modulation mode of the signal, only by accurately estimating the parameters of the signal can the received modulated signal be correctly demodulated, and then the important information carried in the signal can be obtained.

[0003] The patent "A method for estimating the parameters of LFM pulse signals under aliasing conditions" (CN 117907941A) provides a method for estimating the parameters of LFM pulse signals under aliasing conditions, which is applied to the field of radar technology. The direct wave signal parameters of multiple radiation sources are similar, and the fractional Fourier transform (FrFT) is used to accumulate the energy of the direct wave signal to obtain the analytical expression after the peak LFM signal passes through to construct a multi-channel signal. The LFM signal is de-aliased in the FrFT domain through blind source separation, ultimately achieving the goal of improving the signal-to-noise ratio and separating different pulse signals with similar parameters, thereby achieving accurate parameter estimation.

[0004] The patent "Signal parameter estimation method, device, medium and equipment based on neural network" (CN 117596100A) provides a signal parameter estimation method based on neural network, which belongs to the field of communication technology. A pre-trained CNN model and a complex modulated signal to be estimated are obtained, and the real part and the imaginary part of the complex modulated signal are combined into a two-dimensional matrix. The two-dimensional matrix is ​​processed using the convolution module, the channel attention mechanism and the spatial attention mechanism in the convolution attention module, and the dilated causal convolution layer in the temporal convolution network module to obtain a feature matrix. The feature matrix is ​​processed using the first branch in the fully connected module to obtain a modulation type parameter, and the feature matrix is ​​processed using the second branch to obtain a symbol rate parameter.

[0005] The shortcomings of the existing technology are mainly reflected in: first, the traditional parameter estimation method is greatly affected by the signal-to-noise ratio and requires a large number of mathematical transformations; second, the high precision and effectiveness of the existing parameter estimation algorithms are all targeted at a certain type or class of specific LPI radar signals, which have certain limitations and generally have weak generalization capabilities. Summary of the invention

[0006] In view of the above-mentioned technical problems existing in the prior art, the present invention proposes a signal parameter estimation method under non-cooperative conditions based on deep learning, which has a reasonable design, overcomes the shortcomings of the prior art, and has good effects.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A signal parameter estimation method under non-cooperative conditions based on deep learning comprises the following steps:

[0009] Step 1: Obtain electromagnetic signals of different modulation types;

[0010] Step 2: Construct an intelligent parameter estimation model for electromagnetic signals; the specific steps are as follows:

[0011] Step 2.1: Input the modulated signal into a multi-channel convolutional network to achieve scale dimension transformation and feature extraction of signal features;

[0012] Step 2.2: Use the output of the multi-way convolutional network as the input of the Transformer encoder network, and use the multi-head attention mechanism to process the local features extracted by the multi-way convolutional network;

[0013] Step 2.3: Perform residual connection between the processed global features and the local features before processing by the multi-head self-attention mechanism, and perform normalization;

[0014] Step 2.4: The normalized features are used as the input of the feedforward neural network. After the feature transformation of the feedforward neural network, residual connection and normalization are performed again. The output obtained at this time is the output of the entire Transformer encoder network.

[0015] Step 2.5: The features output by the Transformer encoder network are calculated through the fully connected layer to output the final parameter estimate;

[0016] Step 3: Train the neural network and save the optimal model;

[0017] Step 4: The unknown parameter modulation signal outputs the parameters to be estimated of the modulation signal through the optimal model.

[0018] Preferably, in step 2.1, the four modules of the multi-channel convolutional network are composed of 3-layer, 3-layer, 9-layer, and 3-layer convolutional blocks, respectively. The convolution kernel size of the multi-channel convolutional network is set to 3×3, the step size is 1, and the original modulated signal of 2×2000 is input into the network for multi-channel convolution processing to achieve scale transformation of signal features and local feature extraction. The output on each feature dimension is processed using a nonlinear activation function and a normalization layer to improve the nonlinear expression ability of the network.

[0019] Preferably, in step 2.3, the local features of the modulated signal extracted by the multi-channel convolutional network are input into the Transformer encoder network, and after a multi-head attention mechanism, the local features are residually connected with the global features and normalized.

[0020] Preferably, the calculation formula of the multi-head attention mechanism is as shown in (1) and (2):

[0021] Y=MutiAttention[Q,K,V]=concat(h1,h2,...h k )W o (1);

[0022] h i =attention(QW i Q ,KW i K ,VW i V ) (2);

[0023] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, and W i Q , W i K , W i V are the parameter matrices, h i is the output of the attention mechanism.

[0024] Preferably, the calculation steps of the multi-head attention mechanism are as follows:

[0025] First, transform Q, K, and V through k sets of linear mappings learned separately;

[0026] Then these k groups of transformed Q, K and V are sent to the single-layer dot product attention mechanism in parallel;

[0027] Finally, the outputs of these k groups of attention mechanisms are concatenated together and then subjected to a linear transformation to obtain the final output.

[0028] Preferably, in step 3, the estimation effect of the signal parameter estimation model can be evaluated by the relative mean square error RMSE between the true value and the estimated value; RMSE is defined as the difference between the true parameter and the predicted estimated value divided by the true value, and the obtained value is squared and then averaged, as shown in formula (3):

[0029]

[0030] in, is the estimated parameter value, y i is the true value, and m is the sample size.

[0031] Preferably, in step 4, the parameters of the optimal model are loaded, and the modulated signal data of the parameters to be estimated are input into the optimal network model, and the output parameters of the neural network are the signal parameters to be estimated.

[0032] Beneficial technical effects brought by the present invention:

[0033] In view of the problems of electromagnetic signals such as complex forms, large differences in timing information, and diverse parameter dimensions, the study explored multi-type signal features and fully utilized the unique advantages of deep learning in feature extraction and expression. This paper proposed a signal parameter estimation method under non-collaborative conditions based on deep learning.

[0034] Convolutional neural networks are widely used in detection and recognition tasks due to their powerful feature extraction capabilities. However, due to the limitation of the receptive field, the convolution operation cannot effectively capture the global features of the modulated signal, and the feature extraction of long-distance time series information in signal parameter estimation is not comprehensive enough. The Transformer model introduces a self-attention mechanism, which enables the model to mine long-distance feature dependencies, but it is easy to ignore local detail information. Therefore, the present invention combines the advantages of both and proposes a parameter estimation algorithm model for a multi-channel convolutional network to complete the parameter estimation of the signal. This architecture can extract local detail features of the signal and capture global features. The fully connected layer outputs the final parameter estimation value, which effectively improves the signal parameter estimation effect. The present invention introduces a multi-head attention mechanism module, which can represent electromagnetic signals in multiple spaces, fully extract the fine-grained information and global features of the input signal, and effectively balance the deviation that may be caused by only one self-attention head structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the method of the present invention;

[0036] Figure 2 It is a diagram of multiple types of modulation signals;

[0037] in, Figure 2 (a) is a schematic diagram of linear frequency modulation; Figure 2(b) is a schematic diagram of a nonlinear frequency modulation signal; Figure 2 (c) is a schematic diagram of a normal pulse; Figure 2 (d) is a schematic diagram of Barker code; Figure 2 (e) is a schematic diagram of a noise amplitude modulation signal; Figure 2 (f) is a schematic diagram of a multi-tone interference signal;

[0038] Figure 3 Schematic diagram of a multi-way convolutional network;

[0039] Figure 4 Schematic diagram of the multi-head self-attention mechanism. DETAILED DESCRIPTION

[0040] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0041] The process of the method of the present invention is as follows Figure 1 As shown, the specific implementation is as follows:

[0042] (1) Obtaining electromagnetic signals of different modulation types

[0043] Signal types have become more complex and diverse with the development of multi-system radiation sources and the changing communication environment. Different signals have different formation mechanisms and require very different parameters to be estimated. For example, the main parameters to be estimated for linear frequency modulation signals are the starting frequency and frequency modulation slope; the main parameters to be estimated for phase coded signals are the signal's carrier frequency, symbol width, code rate, and coding method; the main parameters to be estimated for FSK signals are the coding frequency interval and symbol sequence.

[0044] like Figure 2 As shown in (a)-(f) of FIG. 1 , the present invention simulates and generates multiple types of modulation signals, such as linear frequency modulation signals, nonlinear frequency modulation signals, binary phase shift keying signals, quadrature amplitude modulation signals, Barker code signals, single-tone signals, multi-tone signals, frequency hopping signals, etc. The ratio of the training set to the test set is set to 8:2.

[0045] (2) Construct a signal intelligent parameter estimation model.

[0046] First, a multi-channel convolutional network is used to extract local features of the modulated signal; second, the output of the multi-channel convolutional network is used as the input of the Transformer encoder network to extract global features; finally, the global features extracted by the Transformer encoder network are output through the fully connected layer to obtain the final parameter estimation value. The specific process is as follows:

[0047] Step 1: Set the convolution kernel size of the multi-channel convolutional network to 3×3 and the step size to 1. Input the original modulated signal of 2×2000 into the network for multi-channel convolution processing. Use nonlinear activation function and normalization layer to process the output on each feature dimension to improve the nonlinear expression ability of the network and reduce the adverse effects of abnormal values ​​on training.

[0048] Step 2: The four modules of the multi-channel convolutional network are composed of 3-layer, 3-layer, 9-layer, and 3-layer convolutional blocks, respectively. Figure 3 As shown in Figure 2, the feature map extracted in step 1 is used through a multi-channel convolutional network to achieve signal feature scale transformation and feature extraction.

[0049] Step 3: Input the local features of the signal extracted by the multi-channel convolutional network into the Transformer encoder network and pass through the multi-head attention mechanism, such as Figure 4 As shown in the figure, the features are residually connected and normalized to improve the stability of network training.

[0050] When the self-attention mechanism model extracts features from the information at the current position, it tends to over-focus on its own position. In order to avoid this defect and enhance the fitting ability of the model, the present invention proposes a multi-head self-attention mechanism. The calculation formula of the multi-head attention mechanism is as follows:

[0051] Y=MutiAttention[Q,K,V]=concat(h1,h2,...h k )W o

[0052] h i =attention(QW i Q ,KW i K ,VW i V )

[0053] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, and W i Q , W i K , W i V are the parameter matrices, h i is the output of the attention mechanism.

[0054] First, Q, K, and V are transformed through k sets of linear mappings learned separately, and then these k sets of transformed Q, K, and V are sent to a single-layer dot product attention mechanism in parallel. Finally, the outputs of these k sets of attention mechanisms are concatenated together and then subjected to a linear transformation to obtain the final output. The use of a multi-head self-attention mechanism can represent signals in multiple spaces, thereby fully extracting multiple levels of spatial features of the input signal and effectively balancing the deviations that may be caused by a structure with only one self-attention head.

[0055] Step 4: Use the normalized features as the input of the feedforward network. After the feature transformation of the feedforward neural network, perform residual connection and normalization again. The output obtained at this time is the output of the entire Transformer encoder network.

[0056] Step 5: The global features output by the Transformer encoder are calculated through the fully connected layer to output the final parameter estimate.

[0057] (3) Train the neural network and save the optimal model

[0058] The estimation effect of the signal parameter estimation model can be evaluated by the relative mean square error RMSE between the true value and the estimated value. RMSE is defined as the difference between the true parameter and the predicted estimated value divided by the true value, and the obtained value is squared and then averaged. The formula is as follows:

[0059]

[0060] in, is the estimated parameter value, y i is the true value, and m is the sample size.

[0061] (4) Parameters to be estimated for the output modulation signal

[0062] Load the parameters of the optimal model, input the modulated signal data of the parameters to be estimated into the optimal network model, and the output parameters of the network are the signal parameters to be estimated.

[0063] The key points and protection points of the present invention include:

[0064] The present invention combines convolutional neural networks with Transformer networks and proposes a deep learning parameter estimation algorithm model to complete the parameter estimation of modulated signals. This architecture can extract local detail features of the signal and capture global features. The fully connected layer outputs the final parameter estimation value, effectively improving the signal parameter estimation effect.

[0065] The present invention introduces a multi-head attention mechanism module, which can represent signals in multiple spaces and fully extract multiple levels of spatial features of the input signal, effectively balancing the deviations that may be caused by only one self-attention head structure.

[0066] Compared with traditional methods, signal parameter estimation based on deep learning networks can not only reduce the huge workload of manual feature extraction, but also achieve the same or even better results as manual feature extraction, making the signal parameter estimation method more intelligent and accurate.

[0067] The present invention protects signal identification, parameter estimation units or functional components in signal monitoring, analysis and identification equipment.

[0068] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A signal parameter estimation method under non-cooperative conditions based on deep learning, characterized in that: The steps include: Step 1: Obtain electromagnetic signals of different modulation types; Step 2: Construct an intelligent parameter estimation model for electromagnetic signals; the specific steps are as follows: Step 2.1: Input the modulated signal into a multi-channel convolutional network to achieve scale dimension transformation and feature extraction of signal features; Step 2.2: Use the output of the multi-way convolutional network as the input of the Transformer encoder network, and use the multi-head attention mechanism to process the local features extracted by the multi-way convolutional network; Step 2.3: Perform residual connection between the processed global features and the local features before processing by the multi-head self-attention mechanism, and perform normalization; Step 2.4: The normalized features are used as the input of the feedforward neural network. After the feature transformation of the feedforward neural network, residual connection and normalization are performed again. The output obtained at this time is the output of the entire Transformer encoder network. Step 2.5: The features output by the Transformer encoder network are calculated through the fully connected layer to output the final parameter estimate; Step 3: Train the neural network and save the optimal model; Step 4: The unknown parameter modulation signal outputs the parameters to be estimated of the modulation signal through the optimal model.

2. The signal parameter estimation method under non-cooperative conditions based on deep learning according to claim 1, characterized in that: In step 2.1, the four modules of the multi-way convolutional network are composed of 3-layer, 3-layer, 9-layer, and 3-layer convolutional blocks respectively. The convolution kernel size of the multi-way convolutional network is set to 3×3, the step size is 1, and the original modulated signal of size 2×2000 is input into the network for multi-channel convolution processing to achieve scale transformation and local feature extraction of signal features; nonlinear activation functions and normalization layers are used to process the output on each feature dimension to improve the nonlinear expression ability of the network.

3. The signal parameter estimation method under non-cooperative conditions based on deep learning according to claim 1, characterized in that: In step 2.3, the local features of the modulated signal extracted by the multi-channel convolutional network are input into the Transformer encoder network. After the multi-head attention mechanism, the local features are residually connected with the global features and normalized.

4. The signal parameter estimation method under non-cooperative conditions based on deep learning according to claim 3, characterized in that: The calculation formulas of the multi-head attention mechanism are shown in (1) and (2): Y=MutiAttention[Q,K,V]=concat(h1,h2,...h k )W o (1); h i =attention(QW i Q ,KW i K ,VW i V ) (2); Among them, Q is the query matrix, K is the key matrix, V is the value matrix, and W i Q , W i K , W i V are the parameter matrices, h i is the output of the attention mechanism.

5. The signal parameter estimation method under non-cooperative conditions based on deep learning according to claim 4, characterized in that: The calculation steps of the multi-head attention mechanism are as follows: First, transform Q, K, and V through k sets of linear mappings learned separately; Then these k groups of transformed Q, K and V are sent to the single-layer dot product attention mechanism in parallel; Finally, the outputs of these k groups of attention mechanisms are concatenated together and then subjected to a linear transformation to obtain the final output.

6. The signal parameter estimation method under non-cooperative conditions based on deep learning according to claim 1, characterized in that: In step 3, the estimation effect of the signal parameter estimation model can be evaluated by the relative mean square error RMSE between the true value and the estimated value; RMSE is defined as the difference between the true parameter and the predicted estimated value divided by the true value, and the obtained value is squared and then averaged, as shown in formula (3): in, is the estimated parameter value, y i is the true value, and m is the sample size.

7. The signal parameter estimation method under non-cooperative conditions based on deep learning according to claim 1, characterized in that: In step 4, the parameters of the optimal model are loaded, and the modulated signal data of the parameters to be estimated are input into the optimal network model. The output parameters of the neural network are the signal parameters to be estimated.

Images