Training method of frequency difference estimation model
By enhancing the signal using an optimal stochastic resonance system in a low signal-to-noise ratio environment and training a frequency difference estimation model using a convolutional neural network, the problem of insufficient frequency difference estimation accuracy in existing technologies is solved, achieving higher prediction accuracy and generalization ability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 36TH RES INST OF CETC
- Filing Date
- 2026-01-21
- Publication Date
- 2026-04-10
AI Technical Summary
Existing frequency deviation estimation models are prone to overfitting when training data is limited, and their feature extraction capabilities are insufficient, resulting in poor frequency deviation prediction accuracy, especially in low signal-to-noise ratio environments where the estimation error is large.
By changing the location of the radiation source signal and conducting multiple experiments, the signal was enhanced using an optimal stochastic resonance system. A frequency difference estimation model was trained using a convolutional neural network, employing an input layer, a feature extraction layer, and a fully connected regression layer. The model was trained using training samples with enhanced signal.
It improves the prediction accuracy of the frequency difference estimation model in low signal-to-noise ratio environments, reduces the impact of noise, and enhances the model's generalization ability.
Smart Images
Figure CN121831676A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of radiation source signal positioning, and in particular to a training method of a frequency difference estimation model. BACKGROUND
[0002] FDOA (Frequency Difference of Arrival) extraction is the core technology of multi-station passive positioning system, and is widely used in radar, sonar, wireless communication, electronic reconnaissance and other fields. By measuring the frequency difference of signals arriving at different observation stations, the position and motion state of the radiation source can be accurately calculated, which has important value in military defense, disaster relief, unmanned vehicle monitoring and other scenarios.
[0003] Traditional neural networks are used for frequency difference regression tasks, such as shallow CNN or fully connected network, but they have the following defects: ① insufficient feature extraction capability: shallow networks are difficult to effectively extract deep time-frequency features related to frequency difference from raw signals; ② high risk of overfitting: when the training data is limited, the model is easy to fall into local optimum, and the generalization ability is poor, and the frequency difference estimation model trained on this basis has poor prediction accuracy for signal frequency difference.
[0004] Therefore, there is an urgent need for a new technical solution for training a frequency difference estimation model. SUMMARY
[0005] In view of the above analysis, the embodiments of the present application aim to provide a training method of a frequency difference estimation model to solve the problem of poor prediction accuracy of the existing frequency difference estimation model for signal frequency difference.
[0006] The embodiments of the present application provide a training method of a frequency difference estimation model, which comprises:
[0007] Through multiple experiments by changing the position of the radiation source signal, in each experiment, signals received by two observation stations are used, and the signals of the same position radiation source received by the two observation stations are used as a group of signals, and the frequency difference of each group of signals is labeled; Each group of signals is subjected to signal enhancement processing by using a pre-set optimal stochastic resonance system, each group of signals after signal enhancement processing and the frequency difference of each group of signals are used as a training sample, and a training set of the frequency difference estimation model is obtained; The frequency difference estimation model is trained based on the training set, and a trained frequency difference estimation model is obtained; wherein the frequency difference estimation model adopts a convolutional neural network.
[0008] Based on the further improvement of the above training method, the frequency difference estimation model comprises an input layer, a feature extraction layer and a fully connected regression layer. The input layer is used to receive each group of signals and transmit each group of signals to the feature extraction layer. The feature extraction layer is used to extract features from each group of signals and then transmit the extracted features to the fully connected regression layer. The fully connected regression layer is used to predict the frequency difference of each signal based on the extracted features of each signal.
[0009] Based on the further improvement of the above training method, the feature extraction layer includes multiple feature extraction sub-layers connected in sequence, and each feature extraction sub-layer includes a convolutional layer, a batch normalization layer and an activation function layer connected in sequence. Convolutional layers are used to perform convolution operations on the input signals and then transmit the convolutional signals to the batch normalization layer. The batch normalization layer is used to perform batch normalization on the input signal and then transmit the batch normalized signal to the activation function layer. The activation function layer is used to activate the input signal and then transmit the activated signal to the next layer of the network.
[0010] Based on further improvements to the above training method, the feature extraction layer includes 3-5 feature extraction sub-layers.
[0011] Based on the further improvement of the above training method, the loss function of the frequency difference estimation model is: ; in, Indicates the number of training samples. Indicates the first The signal frequency difference label value in each training sample Indicates the first Predicted signal frequency difference values from each training sample. This represents the minimum constant.
[0012] Based on the further improvement of the above training method, the optimal stochastic resonance system is pre-set through the following steps: Multiple signal samples were collected by two observation stations; each signal sample included signals collected by both observation stations from the same radiation source. All possible stochastic resonance systems were determined based on the signal-to-noise ratio of the signals received at the two observation stations. By using multiple signal samples to screen all possible stochastic resonance systems, the optimal stochastic resonance system is obtained.
[0013] Based on the further improvement of the above training method, the step of using multiple signal samples to screen all possible stochastic resonance systems to obtain the optimal stochastic resonance system includes: The evaluation values of all possible random resonance systems are obtained by evaluating all possible random resonance systems respectively using multiple signal samples. The optimal random resonance system is determined according to the evaluation values of all possible random resonance systems.
[0014] Based on the further improvement of the training method, the evaluation value of each possible random resonance system is determined by the following steps, comprising: Multiple signal samples are input into each possible random resonance system for signal enhancement processing, and multiple signal enhancement samples are obtained. The normalized cross-correlation coefficient of each signal enhancement sample in the multiple signal enhancement samples is calculated, and multiple normalized cross-correlation coefficients are obtained. The average value of the multiple normalized cross-correlation coefficients is taken as the evaluation value of each possible random resonance system.
[0015] Based on the further improvement of the training method, the normalized cross-correlation coefficient of each signal enhancement sample is calculated by the following formula: ; Wherein, The normalized cross-correlation coefficient of each signal enhancement sample is denoted as, The acquisition signal sequence of one observation station in each signal enhancement sample is denoted as, The acquisition signal average sequence of one observation station in each signal enhancement sample is denoted as, The acquisition signal of another observation station in each signal enhancement sample is denoted as, The acquisition signal average sequence of another observation station in each signal enhancement sample is denoted as.
[0016] Based on the further improvement of the training method, the optimal random resonance system is determined according to the evaluation values of all possible random resonance systems, comprising: The evaluation values of all possible random resonance systems are sorted in descending order, and the sorted evaluation values are obtained. The maximum evaluation value is selected from the sorted evaluation values, and the random resonance system corresponding to the maximum evaluation value is taken as the optimal random resonance system.
[0017] Compared with the prior art, the present application can at least realize one of the following beneficial effects: The optimal random resonance system is set in advance to perform signal enhancement processing on each group of signals, which reduces the noise in each group of signals after signal enhancement processing, and the each group of signals after signal enhancement processing is used as a training sample to train the frequency difference estimation model, thereby further improving the prediction accuracy of the frequency difference estimation model on signal frequency difference.
[0018] The technical solutions in the present application can be combined with each other to realize more preferred combination solutions. Other features and advantages of the present application will be described in the following description, and some advantages will become apparent from the description or can be understood by implementing the present application. The objects and other advantages of the present application can be realized and obtained by the content particularly pointed out in the description and the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS
[0019] The accompanying drawings are included to provide a further understanding of the present application, and are incorporated herein and constitute a part of the detailed description. The same reference numbers in different drawings refer to the same elements throughout the drawings.
[0020] Figure 1 A flowchart of a training method of a frequency difference estimation model provided for an embodiment of the present application is shown in the figure. Figure 2 A structural diagram of a frequency difference estimation model provided for an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION
[0021] The preferred embodiments of the present application will be described in detail below with reference to the accompanying drawings, wherein the accompanying drawings form a part of the present application and are used to explain the principles of the embodiments of the present application, but are not used to limit the scope of the present application.
[0022] In a low signal-to-noise ratio application scenario, it is often difficult to accurately estimate the frequency difference of different observation stations, so that the accuracy error of estimating the position of a radiation source in the low signal-to-noise ratio application scenario is large, and it is difficult to meet the application needs.
[0023] One specific embodiment of the present application discloses a training method of a frequency difference estimation model, as shown in the figure, the training method comprises: Figure 1 Step S1: multiple tests are performed by changing the position of a radiation source signal, in each test, signals received by two observation stations are used, the signals of the same position radiation source received by the two observation stations are used as a group of signals, and the frequency difference of each group of signals is labeled; Step S2: a pre-set optimal stochastic resonance system is used to perform signal enhancement processing on each group of signals, each group of signals after the signal enhancement processing and the frequency difference of each group of signals are used as a training sample, and a training set of the frequency difference estimation model is obtained; Step S3: the frequency difference estimation model is trained based on the training set, and a trained frequency difference estimation model is obtained; wherein the frequency difference estimation model adopts a convolutional neural network.
[0024] Specifically, as shown in the figure, in step S1, multiple tests are performed by using two observation stations, multiple tests can be performed at the same position, and signals can be randomly generated in each test, and the two observation stations are used for receiving. Figure 1 In step S1, multiple tests are performed by using two observation stations, multiple tests can be performed at the same position, and signals can be randomly generated in each test, and the two observation stations are used for receiving.
[0025] Specifically, a group of signals can be obtained through each test, and the difference between the spectral peak values of the signals of the two stations is measured by using a high-resolution spectrum analyzer as a reference for frequency difference marking.
[0026] Specifically, as shown in FIG. 1, in step S1, each group of signals includes the sampling signals of the two observation stations. When the application environment of the two observation stations is a low signal-to-noise ratio environment, the noise in the collected signals is large, which seriously affects the frequency difference estimation of the two observation stations. Figure 1
[0027] Specifically, as shown in FIG. 2, in step S2, an optimal stochastic resonance system is set in advance. The optimal stochastic resonance system converts part of the energy of the noise in the signal to be estimated into signal energy, so as to enhance the weak characteristic signal, improve the signal-to-noise ratio of the signal, and improve the frequency difference estimation of the two observation stations. Figure 1
[0028] Preferably, the optimal stochastic resonance system is set in advance by the following steps: Collecting a plurality of signal samples by the two observation stations; wherein each signal sample includes the signals collected by the two observation stations on the same radiation source signal; Determining all possible stochastic resonance systems based on the signal reception signal-to-noise ratios of the two observation stations; Screening all possible stochastic resonance systems by using the plurality of signal samples to obtain the optimal stochastic resonance system.
[0029] Specifically, the same radiation source signal can be collected by the two observation stations respectively, and the signals in different time periods are taken as a signal sample. The sampling is discrete sampling, and each signal sample includes a continuous fixed number of sampling points.
[0030] Specifically, the same radiation source signal can also be collected by the plurality of observation stations respectively, and the signals of any two observation stations in the same time period are taken as a plurality of signal samples. The sampling is discrete sampling, and each signal sample includes a continuous fixed number of sampling points.
[0031] It is worth noting that the signals collected by the two observation stations on the same radiation source signal in the same time period are included in the same signal sample, which will not be described here.
[0032] It can be understood that the linear stiffness coefficient a and the nonlinear stiffness coefficient b are included in the stochastic resonance system, and each combination of the linear stiffness coefficient a and the nonlinear stiffness coefficient b is taken as a stochastic resonance system.
[0033] Specifically, after obtaining the plurality of signal samples, the signal-to-noise ratio is determined in combination with the application scenarios in which the two observation stations are located, that is, the signal receiving signal-to-noise ratios of the two observation stations are determined, the feasibility range of the linear stiffness coefficient a and the feasibility range of the nonlinear stiffness coefficient b in the stochastic resonance system are determined according to the signal receiving signal-to-noise ratios of the two observation stations, and then the combination of all linear stiffness coefficients a and nonlinear stiffness coefficients b is determined as all possible stochastic resonance systems.
[0034] Specifically, all possible stochastic resonance systems are screened using the plurality of signal samples to obtain an optimal stochastic resonance system.
[0035] Preferably, the screening of all possible stochastic resonance systems using the plurality of signal samples to obtain an optimal stochastic resonance system comprises: evaluating all possible stochastic resonance systems using the plurality of signal samples to obtain evaluation values of all possible stochastic resonance systems; determining the optimal stochastic resonance system according to the evaluation values of all possible stochastic resonance systems.
[0036] Specifically, all possible stochastic resonance systems are numbered, and each numbered stochastic resonance system is evaluated using the plurality of signal samples to determine the evaluation value corresponding to each numbered stochastic resonance system, that is, the evaluation values of all possible stochastic resonance systems are determined.
[0037] Preferably, the evaluation value of each possible stochastic resonance system is determined by the following steps, comprising: inputting the plurality of signal samples into each possible stochastic resonance system for signal enhancement processing to obtain a plurality of signal enhancement samples; calculating the normalized cross-correlation coefficient of each signal enhancement sample in the plurality of signal enhancement samples to obtain a plurality of normalized cross-correlation coefficients; taking the average value of the plurality of normalized cross-correlation coefficients as the evaluation value of each possible stochastic resonance system.
[0038] Specifically, when determining the evaluation value of a numbered stochastic resonance system, the plurality of signal samples are respectively subjected to signal enhancement processing by the numbered stochastic resonance system to obtain a plurality of signal enhancement samples.
[0039] Specifically, the normalized cross-correlation coefficient is calculated for each signal enhancement sample, and the plurality of signal enhancement samples are calculated to obtain a plurality of normalized cross-correlation coefficients.
[0040] Preferably, the normalized cross-correlation coefficient of each signal enhancement sample is calculated by the following formula: ; wherein, a normalized cross-correlation coefficient of each signal enhancement sample, a collected signal sequence of one observation station in each signal enhancement sample, a collected signal average sequence of one observation station in each signal enhancement sample, a collected signal of another observation station in each signal enhancement sample, a collected signal average sequence of another observation station in each signal enhancement sample.
[0041] Specifically, for a plurality of sampling points including two observation stations in each signal enhancement sample, the number of sampling points of the two observation stations is the same, and the normalized cross-correlation coefficient of each signal enhancement sample is calculated by
[0042] Specifically, after obtaining a plurality of normalized cross-correlation coefficients corresponding to a plurality of signal enhancement samples, the average value of the plurality of normalized cross-correlation coefficients is taken as the evaluation value of a random resonance system with a certain number.
[0043] Specifically, by traversing all numbered random resonance systems, the evaluation values of all possible random resonance systems can be obtained.
[0044] It is worth noting that the same random resonance system is evaluated by a plurality of signal enhancement samples, which improves the robustness and reliability of the evaluation value and lays a foundation for training the frequency difference estimation model.
[0045] Specifically, after obtaining the evaluation values of all possible random resonance systems, the evaluation values of all possible random resonance systems are analyzed to determine the optimal random resonance system.
[0046] Preferably, the optimal random resonance system is determined according to the evaluation values of all possible random resonance systems, comprising: sorting the evaluation values of all possible random resonance systems in descending order to obtain sorted evaluation values; selecting the maximum evaluation value from the sorted evaluation values, and taking the random resonance system corresponding to the maximum evaluation value as the optimal random resonance system.
[0047] Specifically, the evaluation values of all possible random resonance systems are compared, then sorted in descending order to obtain sorted evaluation values, and the random resonance system corresponding to the maximum evaluation value is taken as the optimal random resonance system as the optimal random resonance system in step S2.
[0048] Specifically, as Figure 1 As shown, in step S2, each group of signals is input to a pre-set optimal random resonance system. The optimal random resonance system performs signal enhancement processing on each group of signals to obtain each group of signals after signal enhancement processing. Finally, each group of signals after signal enhancement processing and the frequency difference of each group of signals are used as a training sample, thereby obtaining multiple training samples as a training set.
[0049] It is worth noting that a convolutional neural network is a type of feedforward neural network. Its core structure includes convolutional layers, pooling layers, and fully connected layers. Convolutional layers extract local features through sliding kernels, pooling layers reduce dimensionality while retaining key information, and fully connected layers complete regression or classification tasks. Using a convolutional neural network as the structure of a frequency difference estimation model can utilize the convolutional kernels in the convolutional neural network to extract features from signal bit transitions and spectral peaks, resulting in higher accuracy of the predicted frequency difference estimate.
[0050] Specifically, such as Figure 1 As shown, in step S3, the frequency difference estimation model composed of the convolutional neural network is trained using the training set obtained in step S2 to obtain the trained frequency difference estimation model.
[0051] Preferably, the frequency difference estimation model includes an input layer, a feature extraction layer, and a fully connected regression layer; The input layer receives each set of signals and transmits each set of signals to the feature extraction layer. The feature extraction layer is used to extract features from each group of signals and then transmit the extracted features to the fully connected regression layer. The fully connected regression layer is used to predict the frequency difference of each signal based on the extracted features of each signal.
[0052] Specifically, such as Figure 2 As shown, the frequency difference estimation model provided in this embodiment of the invention includes an input layer, a feature extraction layer, and a fully connected regression layer connected in sequence. The input layer receives each set of signals and transmits each set of signals to the feature extraction layer. The feature extraction layer extracts features from each set of signals to obtain the extracted features of each set of signals. In the fully connected regression layer, the frequency difference prediction value of each set of signals is predicted based on the extracted features of each set of signals.
[0053] Preferably, the feature extraction layer includes multiple feature extraction sub-layers connected in sequence, and each feature extraction sub-layer includes a convolutional layer, a batch normalization layer and an activation function layer connected in sequence. Convolutional layers are used to perform convolution operations on the input signals and then transmit the convolutional signals to the batch normalization layer. The batch normalization layer is used to perform batch normalization on the input signal and then transmit the batch normalized signal to the activation function layer. The activation function layer is used to activate the input signal and then transmit the activated signal to the next layer of the network.
[0054] Specifically, such as Figure 2 As shown, the feature extraction sub-layers in the frequency difference estimation model provided in this embodiment of the invention have the same structure, and each feature extraction sub-layer includes a convolutional layer, a batch normalization layer and an activation function layer connected in sequence. The convolutional layer performs convolution operation on the input signal, the batch normalization layer performs batch normalization operation on the input signal, and the activation function layer performs activation operation on the input signal.
[0055] Preferably, the feature extraction layer includes 3-5 feature extraction sub-layers.
[0056] Specifically, after determining the network structure of the frequency difference estimation model, the frequency difference estimation model is trained using multiple training samples from the training set obtained in step S2.
[0057] Preferably, the loss function of the frequency difference estimation model is: ; in, Indicates the number of training samples. Indicates the first Signal frequency difference label values in each training sample Indicates the first Predicted signal frequency difference values from each training sample. This represents the minimum constant.
[0058] It is worth noting that since the "relative error" of frequency difference is more important than the "absolute error," for example, the error is 1Hz when the frequency difference is 10Hz, and the error is 1Hz when the frequency difference is 100Hz. The former has a larger relative error. Therefore, the training method of the frequency difference estimation model provided in this embodiment of the invention adopts the relative mean square error loss function.
[0059] Specifically, in step S3, when training the frequency difference estimation model, a training cutoff condition is set in advance, such as when the number of training rounds reaches a threshold or the loss function used meets the threshold, and the training is completed to obtain the trained frequency difference estimation model.
[0060] Compared with the prior art, the frequency difference estimation model training method provided in this embodiment of the invention performs signal enhancement processing on each group of signals through a pre-set optimal stochastic resonance system, which reduces the noise in each group of signals after signal enhancement processing. The frequency difference estimation model is trained using each group of signals after signal enhancement processing as training samples, which further improves the prediction accuracy of the frequency difference estimation model for signal frequency difference.
[0061] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiment methods can be completed by instructing the relevant hardware by a computer program, and the program can be stored in a computer readable storage medium. The computer readable storage medium is a disk, an optical disk, a read-only memory, a random access memory, etc.
[0062] The above description is merely preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.
Claims
1. A training method for a frequency difference estimation model, characterized in that, The training method includes: Multiple experiments were conducted by changing the location of the radiation source signal. In each experiment, two observation stations were used to receive the signal. The signals from the same radiation source received by the two observation stations were taken as a group of signals, and the frequency difference of each group of signals was marked. The signal enhancement process is performed on each group of signals using a pre-set optimal stochastic resonance system. Each group of signals after signal enhancement and the frequency difference of each group of signals are used as training samples to obtain the training set of the frequency difference estimation model. The frequency difference estimation model is trained based on the training set to obtain the trained frequency difference estimation model; the frequency difference estimation model adopts a convolutional neural network.
2. The training method according to claim 1, characterized in that, The frequency difference estimation model includes an input layer, a feature extraction layer, and a fully connected regression layer; The input layer receives each set of signals and transmits each set of signals to the feature extraction layer. The feature extraction layer is used to extract features from each group of signals and then transmit the extracted features to the fully connected regression layer. The fully connected regression layer is used to predict the frequency difference of each signal based on the extracted features of each signal.
3. The training method according to claim 2, characterized in that, The feature extraction layer includes multiple feature extraction sub-layers connected in sequence, and each feature extraction sub-layer includes a convolutional layer, a batch normalization layer and an activation function layer connected in sequence. Convolutional layers are used to perform convolution operations on the input signals and then transmit the convolutional signals to the batch normalization layer. The batch normalization layer is used to perform batch normalization on the input signal and then transmit the batch normalized signal to the activation function layer. The activation function layer is used to activate the input signal and then transmit the activated signal to the next layer of the network.
4. The training method according to claim 3, characterized in that, The feature extraction layer comprises 3-5 feature extraction sub-layers.
5. The training method according to any one of claims 1-4, characterized in that, The loss function of the frequency deviation estimation model is: ； in, Indicates the number of training samples. Indicates the first Signal frequency difference label values in each training sample Indicates the first Predicted signal frequency difference values from each training sample. This represents the minimum constant.
6. The training method according to claim 1, characterized in that, The optimal stochastic resonance system is pre-configured through the following steps: Multiple signal samples were collected by two observation stations; each signal sample included signals collected by both observation stations from the same radiation source. All possible stochastic resonance systems were determined based on the signal-to-noise ratio of the signals received at the two observation stations. By using multiple signal samples to screen all possible stochastic resonance systems, the optimal stochastic resonance system is obtained.
7. The training method according to claim 6, characterized in that, The process of using multiple signal samples to screen all possible stochastic resonance systems to obtain the optimal stochastic resonance system includes: By using multiple signal samples, all possible stochastic resonance systems are evaluated, and evaluation values for all possible stochastic resonance systems are obtained. The optimal stochastic resonance system is determined based on the evaluation values of all possible stochastic resonance systems.
8. The training method according to claim 7, characterized in that, The evaluation value for each possible stochastic resonance system is determined through the following steps: Multiple signal samples are input into each possible stochastic resonance system for signal enhancement processing to obtain multiple signal enhancement samples; Calculate the normalized cross-correlation coefficient for each signal enhancement sample in the multiple signal enhancement samples to obtain multiple normalized cross-correlation coefficients; The average of multiple normalized cross-correlation coefficients is used as the evaluation value for each possible stochastic resonance system.
9. The training method according to claim 8, characterized in that, The normalized cross-correlation coefficient for each signal enhancement sample is calculated using the following formula: ； in, This represents the normalized cross-correlation coefficient for each signal enhancement sample. This represents the sequence of signals acquired by one observation station in each signal enhancement sample. This represents the sequence of averaged signals collected from one observation station in each signal enhancement sample. This represents the signal acquired by another observation station in each signal enhancement sample. The sequence of averaged signals from another observation station in each signal enhancement sample.
10. The training method according to claim 7, characterized in that, The step of determining the optimal stochastic resonance system based on the evaluation values of all possible stochastic resonance systems includes: Sort the evaluation values of all possible stochastic resonance systems in descending order to obtain the sorted evaluation values; Select the highest evaluation value from the sorted evaluation values, and take the stochastic resonance system corresponding to the highest evaluation value as the optimal stochastic resonance system.