Feature extraction method, system and terminal based on multi-central-domain radio frequency signals

Abstract

The invention discloses a feature extraction method and system based on multi-center-domain radio frequency signals and a terminal, and the method comprises the steps: obtaining multi-center-domain radio frequency signals, carrying out the splicing of the multi-center-domain radio frequency signals, obtaining a spliced radio frequency feature, mapping the spliced radio frequency feature to a source domain feature space, obtaining a domain specific feature, and carrying out the splicing of a risk factor feature, and obtaining a domain splicing feature; aligning the radio frequency signals between the source domain and the target domain by using the domain splicing features to obtain a loss function; constructing a domain-specific classifier, inputting the domain-specific features into the domain-specific classifier to output a loss function, and performing prediction by using a target domain of the domain-specific classifier to obtain the loss function; and training a classifier through a plurality of loss functions, and outputting sample features of the target domain. According to the method, feature extraction and splicing operation are performed on the radio frequency signals of the multiple central domains, then the loss function is constructed by using the distribution features to train the classifier, and the classifier integrating multiple specific fields is obtained, so that the feature accuracy of the target domain is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of signal feature processing, and particularly to a method, system, terminal and computer-readable storage medium for feature extraction based on multi-center domain radio frequency signals. Background Art

[0002] Ultrasonic radio frequency signals are of great significance for bone density detection. The heterogeneity of data from different centers has a great impact on the performance of the detection model. However, existing radio frequency research is based on data collection from a single center, which leads to poor cross-regional generalization ability of the model, resulting in inaccurate detection results for the target domain.

[0003] Therefore, the existing technology still needs to be improved and developed. Summary of the Invention

[0004] The main purpose of the present invention is to provide a method, system, terminal and computer-readable storage medium for feature extraction based on multi-center domain radio frequency signals, aiming to solve the problem in the existing technology that the model based on radio frequency research cannot process data from multiple center domains, and the model cannot overcome the influence of data heterogeneity, resulting in inaccurate detection results for the target domain.

[0005] To achieve the above object, the present invention provides a method for feature extraction based on multi-center domain radio frequency signals, and the method for feature extraction based on multi-center domain radio frequency signals includes the following steps:

[0006] Obtain effective radio frequency signals of a target domain and multiple source domains, and input each of the effective radio frequency signals into a constructed multi-layer deep residual network to output corresponding spliced radio frequency features;

[0007] Map each of the spliced radio frequency features to multiple source domain feature spaces respectively to obtain corresponding domain-specific features, and obtain risk factor features, and splice the risk factor features with all the domain-specific features respectively to obtain multiple domain-spliced features;

[0008] Use all the domain-spliced features to perform distribution alignment processing on the effective radio frequency signals between the multiple source domains and the target domain to obtain a first loss function;

[0009] Construct a domain-specific classifier, input the domain-specific features of all the source domains into the domain-specific classifier for classification to output multiple second loss functions, and use the domain-specific classifier to predict the processed target domain to obtain a third loss function;

[0010] Using the first loss function, multiple second loss functions, and the third loss function, train the domain-specific classifier to obtain an integrated domain-specific classifier, and output the processed sample features of the target domain through the integrated domain-specific classifier.

[0011] Optionally, in the feature extraction method based on multi-center domain radio frequency signals, the obtaining of the effective radio frequency signals of the target domain and multiple source domains specifically includes:

[0012] Obtain the radio frequency signals of the target domain and the radio frequency signals of multiple source domains;

[0013] Perform an interception operation on each of the radio frequency signals to obtain the effective radio frequency signals of the target domain and the effective radio frequency signals of each source domain.

[0014] Optionally, in the feature extraction method based on multi-center domain radio frequency signals, the obtaining of the effective radio frequency signals of the target domain and multiple source domains, and inputting each of the effective radio frequency signals into the constructed multi-layer deep residual network to output corresponding spliced radio frequency features specifically includes:

[0015] Obtain the effective radio frequency signals of the target domain and the effective radio frequency signals of multiple source domains, pair the effective radio frequency signal of the target domain with the effective radio frequency signals of each source domain respectively to obtain a plurality of radio frequency signal groups, where the number of the radio frequency signal groups is the same as the number of source domains;

[0016] Input all the radio frequency signal groups into the constructed multi-layer deep residual network respectively, and the multi-layer deep residual network performs convolution processing on all the radio frequency signal groups by using a plurality of convolutional layers to obtain corresponding first radio frequency features;

[0017] The multi-layer deep residual network performs pooling convolution processing on all the radio frequency signal groups by using a pooling layer and a convolutional layer to obtain corresponding second radio frequency features;

[0018] The multi-layer deep residual network splices each of the first radio frequency features and the corresponding second radio frequency features to obtain a plurality of spliced radio frequency features.

[0019] Optionally, in the feature extraction method based on multi-center domain radio frequency signals, the mapping of each of the spliced radio frequency features to multiple source domain feature spaces respectively to obtain corresponding domain-specific features, and the obtaining of risk factor features, and splicing the risk factor features with all the domain-specific features respectively to obtain a plurality of domain-spliced features specifically includes:

[0020] Construct the source domain feature space of each of the said source domains and the source domain feature space of the target domain, and map the concatenated RF features of each of the said source domains and the concatenated RF features of the target domain into the corresponding source domain feature spaces to obtain corresponding domain-specific features;

[0021] Input the domain-specific features into the pre-constructed domain feature extraction model, and extract risk factor features through the fully connected layer in the domain feature extraction model;

[0022] Through the domain feature extraction model, concatenate the risk factor features with all the domain-specific features respectively to obtain corresponding domain concatenated features.

[0023] Optionally, for the feature extraction method based on multi-center domain RF signals, wherein, using all the domain concatenated features to perform distribution alignment processing on the effective RF signals of multiple said source domains and the effective RF signals between the target domains to obtain a first loss function, specifically including:

[0024] Obtain the source distribution p of the effective RF signals of all the said source domains and the target distribution q of the effective RF signals of the target domain, and calculate the maximum mean difference D(p, q) between the target domain and all the said source domains according to the multiple source distributions p and the target distribution q:

[0025]

[0026] wherein, i represents the i-th source domain, n i represents the number of source domain data of source domain i, m represents one of n i , represents the feature map obtained by mapping the original sample to the Hilbert space through the mapping function, S i represents the source domain data of the i-th source domain, represents S i each source domain data in, T represents the number of target domains, n T represents the T-th target domain, represents n T each target domain data in;

[0027]

[0028] wherein, k represents the kernel function of the mapping function, X Si represents all the source domain data in S i X represents T n T all the target domain data in, represents the feature map obtained by mapping the source domain data to the Hilbert space through the mapping function, It represents the feature map that maps the original samples of the target domain data to the Hilbert space through the mapping function;

[0029] Minimize the maximum mean discrepancy to achieve distribution alignment of the effective radio frequency signals between multiple source domains and the target domain, and obtain the first loss function L between all source domains and the target domain mmd :

[0030]

[0031] where N represents the number of source domains, and both represent the domain concatenation features, represents and the maximum mean discrepancy between.

[0032] Optionally, for the feature extraction method based on multi-center domain radio frequency signals, wherein, construct a domain-specific classifier, input the domain-specific features of all source domains into the domain-specific classifier for classification, output the second loss function of the domain-specific classifier, and use the domain-specific classifier to predict the processed target domain to obtain the third loss function, specifically including:

[0033] Construct a domain-specific classifier, input the domain-specific features of all source domains into the domain-specific classifier for classification, and calculate and output multiple second loss functions L of the domain-specific classifier according to each domain concatenation feature cls :

[0034]

[0035] where N represents the number of source domains, i represents the i-th source domain, S i represents the source domain data of the i-th source domain, E represents the cross-entropy loss, represents the domain concatenation feature, Y Si represents the true label of the source domain data of the i-th source domain, represents S i 's predicted label;

[0036] Input the domain-specific features of the target domain into the domain-specific classifier, and the domain-specific classifier classifies the domain-specific features of the target domain through all the second loss functions and outputs multiple prediction probabilities;

[0037] Construct the third loss function according to all the prediction probabilities and all the domain concatenation features:

[0038]

[0039] Among them, L disc represents the third loss function, and j represents the j-th source domain data other than i, represents S j 's predicted label.

[0040] Optionally, for the feature extraction method based on multi-center domain radio frequency signals, wherein, using the first loss function, multiple second loss functions and the third loss function to train the domain-specific classifier to obtain an integrated domain-specific classifier, and outputting the processed sample features of the target domain through the integrated domain-specific classifier, specifically including:

[0041] Construct a total loss function L according to the first loss function, multiple second loss functions and the third loss function:

[0042] L = L mmd + L cls + L disc ;

[0043] Input the total loss function into the domain-specific classifier for training to obtain an integrated domain-specific classifier, and input all the predicted probabilities into the integrated domain-specific classifier to output the processed sample features of the target domain.

[0044] In addition, to achieve the above object, the present invention also provides a feature extraction system based on multi-center domain radio frequency signals, wherein the feature extraction system based on multi-center domain radio frequency signals includes:

[0045] A signal splicing module, configured to obtain effective radio frequency signals of the target domain and multiple source domains, and input each of the effective radio frequency signals into a pre-constructed multi-layer deep residual network to output corresponding spliced radio frequency features;

[0046] A feature splicing module, configured to map each of the spliced radio frequency features into multiple source domain feature spaces to obtain corresponding domain-specific features, and obtain risk factor features, and splice the risk factor features with all the domain-specific features to obtain multiple domain-spliced features;

[0047] A data alignment module, configured to use all the domain-spliced features to perform distribution alignment processing on the effective radio frequency signals between the multiple source domains and the target domain to obtain a first loss function;

[0048] A feature prediction module, configured to construct a domain-specific classifier, input all the domain-specific features of the source domains into the domain-specific classifier for classification to output multiple second loss functions, and use the domain-specific classifier to predict the processed target domain to obtain a third loss function;

[0049] A feature output module, configured to use the first loss function, multiple second loss functions, and the third loss function to train the domain-specific classifier, obtain an integrated domain-specific classifier, and output sample features of the target domain through the integrated domain-specific classifier.

[0050] In addition, to achieve the above object, the present invention further provides a terminal, wherein the terminal includes: a memory, a processor, and a feature extraction program based on multi-center domain radio frequency signals stored on the memory and executable on the processor. When the feature extraction program based on multi-center domain radio frequency signals is executed by the processor, the steps of the above-mentioned feature extraction method based on multi-center domain radio frequency signals are implemented.

[0051] In addition, to achieve the above object, the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a feature extraction program based on multi-center domain radio frequency signals. When the feature extraction program based on multi-center domain radio frequency signals is executed by a processor, the steps of the above-mentioned feature extraction method based on multi-center domain radio frequency signals are implemented.

[0052] In the present invention, effective radio frequency signals of a target domain and multiple source domains are obtained, and each effective radio frequency signal is input into a constructed multi-layer deep residual network to output corresponding spliced radio frequency features; each spliced radio frequency feature is respectively mapped to multiple source domain feature spaces to obtain corresponding domain-specific features, and risk factor features are obtained. The risk factor features are respectively spliced with all the domain-specific features to obtain multiple domain-spliced features; using all the domain-spliced features, distribution alignment processing is performed on the effective radio frequency signals of multiple source domains and the effective radio frequency signals between the target domains to obtain a first loss function; a domain-specific classifier is constructed, and all the domain-specific features of the source domains are input into the domain-specific classifier for classification to output multiple second loss functions, and the processed target domain is predicted using the domain-specific classifier to obtain a third loss function; using the first loss function, multiple second loss functions, and the third loss function to train the domain-specific classifier to obtain an integrated domain-specific classifier, and outputting sample features of the processed target domain through the integrated domain-specific classifier. The domain-specific classifier predicts the processed target domain to obtain a third loss function; using the first loss function, multiple second loss functions, and the third loss function to train the domain-specific classifier to obtain an integrated domain-specific classifier, and outputting sample features of the target domain through the integrated domain-specific classifier. By performing feature extraction and splicing operations on radio frequency signals in multiple center domains, and then using distribution features to construct a loss function to train the classifier, the present invention obtains an integrated classifier specific to multiple domains, thereby improving the feature accuracy of the target domain. Description of the Drawings

[0053] Figure 1 is a flowchart of a preferred embodiment of the method for feature extraction based on multi - center domain radio frequency signals according to the present invention;

[0054] Figure 2 is a network framework diagram of a preferred embodiment of the method for feature extraction based on multi - center domain radio frequency signals according to the present invention;

[0055] Figure 3 is a schematic diagram of a multi - layer deep residual network of a preferred embodiment of the method for feature extraction based on multi - center domain radio frequency signals according to the present invention;

[0056] Figure 4 is a schematic diagram of a residual module of a preferred embodiment of the method for feature extraction based on multi - center domain radio frequency signals according to the present invention;

[0057] Figure 5 is a structural diagram of a preferred embodiment of the feature extraction system based on multi - center domain radio frequency signals according to the present invention;

[0058] Figure 6 is a structural diagram of a preferred embodiment of the terminal according to the present invention. Detailed implementation manners

[0059] To make the objectives, technical solutions and advantages of the present invention clearer and more definite, the following further describes the present invention in detail with reference to the accompanying drawings and by way of examples. It should be understood that the specific examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0060] The method for feature extraction based on multi - center domain radio frequency signals in the preferred embodiment of the present invention, as Figure 1 shown, the method for feature extraction based on multi - center domain radio frequency signals includes the following steps:

[0061] Step S10, obtain the effective radio frequency signals of the target domain and multiple source domains, and input each of the effective radio frequency signals into the constructed multi - layer deep residual network to output the corresponding spliced radio frequency features.

[0062] Among them, as Figure 2 shown, the sources of signal collection include multiple source domains and one target domain. Since the signal data obtained during collection is long, resulting in a large signal sequence, which will interfere with the model during feature extraction and increase the risk of overfitting. Therefore, during the process of obtaining the effective radio frequency signals, first obtain the radio frequency signals of the target domain and multiple source domains; perform an interception operation on each of the radio frequency signals to obtain the effective radio frequency signals of the target domain and each of the source domains. Through the pre - processing operation of the radio frequency signals, the stability of the radio frequency signals input into the model is improved, thereby avoiding interference with the model and improving the accuracy of subsequent feature extraction.

[0063] Specifically, obtain the effective radio frequency signals of the target domain and the effective radio frequency signals of multiple source domains, pair the effective radio frequency signal of the target domain with the effective radio frequency signal of each source domain respectively to obtain a plurality of radio frequency signal groups, where the number of the radio frequency signal groups is the same as the number of the source domains; input all the radio frequency signal groups into the constructed multi-layer deep residual network respectively, and the multi-layer deep residual network performs convolution processing on all the radio frequency signal groups by using a plurality of convolutional layers to obtain corresponding first radio frequency features; the multi-layer deep residual network performs pooling convolution processing on all the radio frequency signal groups by using a pooling layer and a convolutional layer to obtain corresponding second radio frequency features; the multi-layer deep residual network splices each of the first radio frequency features and the corresponding second radio frequency feature to obtain a plurality of spliced radio frequency features.

[0064] Among them, as Figure 3 shown, there are four channels input into the multi-layer deep residual network, and the finally output result is obtained by concatenating the radio frequency characteristics of the four channels to obtain the spliced radio frequency features. For the effective radio frequency signals input into each channel, they will be processed in two ways, and the two obtained radio frequency features are spliced to obtain the spliced radio frequency features of each channel.

[0065] Further, as Figure 4 shown, the first processing method is to perform convolution operation on the effective radio frequency signal through two identical convolutional layers, which also includes a batch normalization layer and an activation function, to obtain a first radio frequency feature that is half of the original length of the effective radio frequency signal; the second processing method is to perform processing through a pooling layer and a convolutional layer to obtain a second radio frequency feature that is half of the original length of the effective radio frequency signal, and finally splice the first radio frequency feature and the second radio frequency feature to obtain the spliced radio frequency feature corresponding to each effective radio frequency signal. Through the multi-layer deep residual network, the feature extraction ability can be improved, overfitting of the model can be avoided, noise can be removed by noise reduction to avoid interference to the model, and finally normalization is performed to unify the numerical range, improving the stability of the model during training.

[0066] Step S20: Map each of the spliced radio frequency features to a plurality of source domain feature spaces respectively to obtain corresponding domain-specific features, and obtain risk factor features, and splice the risk factor features with all the domain-specific features respectively to obtain a plurality of domain-spliced features.

[0067] Specifically, construct the source domain feature space of each of the source domains and the source domain feature space of the target domain, and map the concatenated RF features of each source domain and the concatenated RF features of the target domain into the corresponding source domain feature spaces to obtain corresponding domain-specific features; input the domain-specific features into the pre-constructed domain feature extraction model, and extract risk factor features through the fully connected layer in the domain feature extraction model; through the domain feature extraction model, concatenate the risk factor features with all the domain-specific features respectively to obtain corresponding domain concatenated features.

[0068] Among them, after feature representation with a multi-layer deep residual network as a common feature extractor, the concatenated RF features of each source domain and the target domain are respectively mapped into their respective source domain feature spaces, so as to extract corresponding domain-specific features. This process can be realized by a domain-specific feature extractor (i.e., the above-mentioned domain feature extraction model), which is composed of two convolutional layers with a kernel length of 1 and one convolutional layer with a kernel length of 3, and does not share weights during training, so as to map each pair of source domain and target domain data into a domain-specific feature space.

[0069] Furthermore, use the fully connected layer of the domain feature extraction model to extract risk factor features, and concatenate them with the domain-specific features respectively, and output corresponding domain concatenated features to improve the discriminability of the output features.

[0070] Step S30: Use all the domain concatenated features to perform distribution alignment processing on the effective RF signals of multiple source domains and the effective RF signals between the target domains to obtain a first loss function.

[0071] Specifically, obtain the source distribution p of the effective RF signals of all the source domains and the target distribution q of the effective RF signals of the target domain, and calculate the maximum mean difference D(p, q) between the target domain and all the source domains according to the multiple source distributions p and the target distribution q:

[0072]

[0073] where i represents the i-th source domain, n i represents the number of source domain data of source domain i, m represents one of the n i and represents the feature map obtained by mapping the original sample into the Hilbert space through the mapping function, S i represents the source domain data of the i-th source domain, represents S i each source domain data in, T represents the number of target domains, n T represents the T-th target domain, represents n Teach piece of target domain data therein;

[0074]

[0075] where k represents the kernel function of the mapping function, X Si represents all source domain data in S i and X T represents all target domain data in n T ; represents the feature map obtained by mapping the original samples of the source domain data to the Hilbert space through the mapping function, and represents the feature map obtained by mapping the original samples of the target domain data to the Hilbert space through the mapping function; minimizing the maximum mean discrepancy to achieve distribution alignment of effective radio frequency signals between multiple source domains and the target domain, and obtaining the first loss function L between all source domains and the target domain mmd :

[0076]

[0077] where N represents the number of source domains, and both represent domain concatenated features, represents and the maximum mean discrepancy between.

[0078] Specifically, after extracting the domain concatenated features, data between the source domain and the target domain is aligned. This process can use MMD (Maximum Mean Discrepancy) to evaluate the data difference between the two domains. Specifically, MMD can test whether two data sets belong to two different distributions. If the mean discrepancy reaches the maximum, it means they come from two different distributions. If the mean discrepancy is 0, it means the two data sets come from the same distribution. In this embodiment, by minimizing the maximum mean discrepancy, data alignment between the target domain and the source domain is achieved, and at the same time, the MMD loss (i.e., the first loss function) between the source domain and the target domain is obtained, thus completing the domain alignment in the first stage, significantly reducing the data heterogeneity between different domains, and improving the accuracy of the model output results.

[0079] Step S40: Construct a domain-specific classifier, input the domain-specific features of all source domains into the domain-specific classifier for classification, output multiple second loss functions, and use the domain-specific classifier to predict the processed target domain to obtain a third loss function.

[0080] Among them, after completing the data alignment processing in different fields in the first stage, the alignment operation in the second stage can be performed, that is, the classifiers in different source fields are aligned to reduce the differences in predicting the target domain and reduce the impact of the heterogeneity of multi-domain data on the model performance.

[0081] Specifically, a domain-specific classifier is constructed, and the domain-specific features of all the source fields are input into the domain-specific classifier for classification. According to the concatenated features of each domain, multiple second loss functions \(L\) of the domain-specific classifier are calculated and output. cls :

[0082]

[0083] Among them, \(N\) represents the number of source domains, \(i\) represents the \(i\)-th source domain, \(S\) i represents the source domain data of the \(i\)-th source domain, \(E\) represents the cross-entropy loss, represents the concatenated domain features, represents the true label of the source domain data of the \(i\)-th source domain, represents \(S\) i 's predicted label;

[0084] The domain-specific features of the target domain are input into the domain-specific classifier, and the domain-specific classifier classifies the domain-specific features of the target domain through all the second loss functions and outputs multiple prediction probabilities.

[0085] According to all the prediction probabilities and all the concatenated domain features, a third loss function is constructed:

[0086]

[0087] Among them, \(L\) disc represents the third loss function, \(j\) represents the \(j\)-th source domain data other than \(i\), represents \(S\) j 's predicted label.

[0088] Among them, for the concatenated domain features of each source domain, the softmax classifier (i.e., the above-mentioned domain-specific classifier) is used as the domain-specific classifier for classification, and the cross-entropy loss is used to represent the loss of the classifier (i.e., the second loss function). In an ideal situation, the predicted target domain samples by different domain-specific classifiers should obtain the same prediction results. However, due to the different data distributions in the source domains, the domain-specific classifier may misclassify the target domain samples, especially for the target domain samples with high class similarity, which need to be corrected. Therefore, the second alignment stage is performed in this process.

[0089] Among them, after inputting the domain-specific features of all source domains into the domain-specific classifier, the probability value of predicting the target domain samples is obtained. The alignment operation in the second stage is to minimize the differences between all probability values, and then the absolute value of the difference in the probability output for each target domain sample is used as the difference loss (i.e., the third loss function). Based on all the loss functions constructed above, the classifier can be trained to improve the accuracy of the classifier in predicting the target domain.

[0090] Step S50: Use the first loss function, multiple second loss functions, and the third loss function to train the domain-specific classifier to obtain an integrated domain-specific classifier, and output the processed sample features of the target domain through the integrated domain-specific classifier.

[0091] Specifically, according to the first loss function, multiple second loss functions, and the third loss function, a total loss function L is constructed:

[0092] L = L mmd + L cls + L disc ;

[0093] Input the total loss function into the domain-specific classifier for training to obtain an integrated domain-specific classifier, and input all the prediction probabilities into the integrated domain-specific classifier to output the processed sample features of the target domain.

[0094] Among them, a total loss function is constructed using three loss functions, and the domain-specific classifier is trained through the total loss function to obtain an integrated domain-specific classifier to identify the samples of the target domain and obtain the labels (or probabilities) of the target domain samples.

[0095] Furthermore, in this embodiment, the dataset used comes from five centers and is ultrasonic radiofrequency signals collected by a QUS (quantitative ultrasound system) bone densitometer provided by five hospitals. The first dataset contains a total of 90 normal samples, 90 osteopenia samples, and 69 osteoporosis samples; the second dataset contains a total of 35 normal samples, 74 osteopenia samples, and 104 osteoporosis samples; the third dataset contains a total of 71 normal samples, 38 osteopenia samples, and 102 osteoporosis samples; the fourth dataset contains a total of 55 normal samples, 74 osteopenia samples, and 101 osteoporosis samples; the fifth dataset contains a total of 152 normal samples, 169 osteopenia samples, and 102 osteoporosis samples; each ultrasonic radiofrequency signal has 4 channels, the length of each signal cycle is 1180, and the size of the signal after preprocessing becomes 4×118000.

[0096] Further, the environment in which this embodiment is executed is on a Linux server. Feature extraction is performed through 5-fold cross-validation, and the same 10-fold cross-validation is used for differential diagnosis. Further, the model is trained using the PyTorch (an open-source deep learning framework for machine learning and deep learning) framework on a single TITAN RTX GPU with 24GB of memory. During the training process, the learning rate is set to 10 -5 , and the number of iterations is set to 200. An early stopping mechanism is set to avoid overfitting, and the batch size is set to 16.

[0097] Further, the classification tasks performed in this embodiment include binary classification and ternary classification experiments. Among them, the classification metrics include four classification evaluation metrics: accuracy (ACC), specificity (SPE), sensitivity (SEN), and the area under the ROC curve (AUC):

[0098]

[0099] Among them, TP (true positive), TN (true negative), FP (false positive), and FN (false negative) represent the number of samples of true positive, true negative, false positive, and false negative, respectively.

[0100] The present invention performs feature extraction and splicing operations on the radio frequency signals in multiple central domains, and then uses the distribution features to construct a loss function to train the classifier, obtaining an integrated multi-domain-specific classifier, thereby improving the feature accuracy of the target domain.

[0101] Further, as Figure 5 shown, based on the above feature extraction method for multi-central domain radio frequency signals, the present invention also correspondingly provides a feature extraction system for multi-central domain radio frequency signals. Among them, the feature extraction system for multi-central domain radio frequency signals includes:

[0102] A signal splicing module 51, configured to obtain the effective radio frequency signals of the target domain and multiple source domains, and input each of the effective radio frequency signals into a constructed multi-layer deep residual network, and output corresponding spliced radio frequency features;

[0103] A feature splicing module 52, configured to map each of the spliced radio frequency features to a plurality of source domain feature spaces respectively, obtain corresponding domain-specific features, and acquire risk factor features, and splice the risk factor features with all the domain-specific features respectively to obtain a plurality of domain-spliced features;

[0104] A data alignment module 53, configured to perform distribution alignment processing on the effective radio frequency signals of the plurality of source domains and the effective radio frequency signals between the target domains by using all the domain-spliced features, to obtain a first loss function;

[0105] A feature prediction module 54, configured to construct a domain-specific classifier, input the domain-specific features of all the source domains into the domain-specific classifier for classification, output a plurality of second loss functions, and use the domain-specific classifier to predict the processed target domain, to obtain a third loss function;

[0106] A feature output module 55, configured to train the domain-specific classifier by using the first loss function, the plurality of second loss functions, and the third loss function, to obtain an integrated domain-specific classifier, and output the sample features of the processed target domain through the integrated domain-specific classifier.

[0107] Further, as Figure 6 shown, based on the above feature extraction method and system for multi-center domain radio frequency signals, the present invention further correspondingly provides a terminal, and the terminal includes a processor 10, a memory 20, and a display 30. Figure 6 Only some components of the terminal are shown, but it should be understood that it is not required to implement all the shown components, and more or fewer components can be alternatively implemented.

[0108] The memory 20 may be an internal storage unit of the terminal in some embodiments, such as a hard disk or a memory of the terminal. The memory 20 may also be an external storage device of the terminal in other embodiments, such as a plug-in hard disk equipped on the terminal, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Further, the memory 20 may also include both the internal storage unit and the external storage device of the terminal. The memory 20 is used to store application software installed on the terminal and various types of data, such as program codes for installing the terminal. The memory 20 may also be used to temporarily store data that has been output or will be output. In one embodiment, a feature extraction program 40 for multi-center domain radio frequency signals is stored on the memory 20, and the feature extraction program 40 for multi-center domain radio frequency signals can be executed by the processor 10, so as to implement the feature extraction method for multi-center domain radio frequency signals in the present application.

[0109] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chips, which are used to run the program code stored in the memory 20 or process data, such as executing the feature extraction method based on multi - center domain radio frequency signals, etc.

[0110] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch - type liquid crystal display, and an OLED (Organic Light - Emitting Diode) touch device, etc. The display 30 is used to display the information of the terminal and to display a visual user interface. The components of the terminal communicate with each other through a system bus.

[0111] In one embodiment, when the processor 10 executes the feature extraction program 40 based on multi - center domain radio frequency signals in the memory 20, the following steps are implemented:

[0112] Obtain the effective radio frequency signals of the target domain and multiple source domains, and input each of the effective radio frequency signals into the constructed multi - layer deep residual network to output corresponding spliced radio frequency features;

[0113] Map each of the spliced radio frequency features to multiple source - domain feature spaces respectively to obtain corresponding domain - specific features, and obtain risk - factor features. Splice the risk - factor features with all the domain - specific features respectively to obtain multiple domain - spliced features;

[0114] Use all the domain - spliced features to perform distribution alignment processing on the effective radio frequency signals of multiple source domains and the effective radio frequency signals between the target domain to obtain a first loss function;

[0115] Construct a domain - specific classifier, input all the domain - specific features of the source domains into the domain - specific classifier for classification, output multiple second loss functions, and use the domain - specific classifier to predict the processed target domain to obtain a third loss function;

[0116] Use the first loss function, multiple second loss functions, and the third loss function to train the domain - specific classifier to obtain an integrated domain - specific classifier, and output the sample features of the processed target domain through the integrated domain - specific classifier.

[0117] Among them, the obtaining of the effective radio frequency signals of the target domain and multiple source domains specifically includes:

[0118] Obtain the radio frequency signals of the target domain and the radio frequency signals of multiple source domains;

[0119] Intercept each of the radio frequency signals to obtain the effective radio frequency signals in the target domain and the effective radio frequency signals in each source domain.

[0120] Among them, obtaining the effective radio frequency signals in the target domain and multiple source domains, and inputting each of the effective radio frequency signals into the constructed multi-layer deep residual network to output the corresponding spliced radio frequency features specifically includes:

[0121] Obtain the effective radio frequency signals in the target domain and the effective radio frequency signals in multiple source domains, pair the effective radio frequency signal in the target domain with the effective radio frequency signals in each source domain respectively to obtain a plurality of radio frequency signal groups, where the number of the radio frequency signal groups is the same as the number of source domains;

[0122] Input all the radio frequency signal groups into the constructed multi-layer deep residual network respectively. The multi-layer deep residual network performs convolution processing on all the radio frequency signal groups by using a plurality of convolutional layers to obtain the corresponding first radio frequency features;

[0123] The multi-layer deep residual network performs pooling convolution processing on all the radio frequency signal groups by using a pooling layer and a convolutional layer to obtain the corresponding second radio frequency features;

[0124] The multi-layer deep residual network splices each of the first radio frequency features and the corresponding second radio frequency features to obtain a plurality of spliced radio frequency features.

[0125] Among them, mapping each of the spliced radio frequency features to a plurality of source domain feature spaces respectively to obtain the corresponding domain-specific features, and obtaining the risk factor features, and splicing the risk factor features with all the domain-specific features respectively to obtain a plurality of domain-spliced features specifically includes:

[0126] Construct the source domain feature spaces of each source domain and the source domain feature space of the target domain, and map the spliced radio frequency features of each source domain and the spliced radio frequency feature of the target domain into the corresponding source domain feature spaces to obtain the corresponding domain-specific features;

[0127] Input the domain-specific features into the constructed domain feature extraction model, and extract the risk factor features through the fully connected layer in the domain feature extraction model;

[0128] Through the domain feature extraction model, splice the risk factor features with all the domain-specific features respectively to obtain the corresponding domain-spliced features.

[0129] Among them, using all the domain-spliced features to perform distribution alignment processing on the effective radio frequency signals in multiple source domains and the effective radio frequency signals between the target domain to obtain the first loss function specifically includes:

[0130] Obtain the source distribution p of the effective radio frequency signals of all the source domains and the target distribution q of the effective radio frequency signals of the target domain. According to the multiple source distributions p and the target distribution q, calculate the maximum mean discrepancy D(p, q) between the target domain and all the source domains:

[0131]

[0132] where i represents the i-th source domain, n i represents the number of source domain data of the source domain i, m represents one of the n i among them, represents the feature map obtained by mapping the original samples to the Hilbert space through the mapping function, S i represents the source domain data of the i-th source domain, represents S i each source domain data in it, T represents the number of target domains, n T represents the T-th target domain, represents n T each target domain data in it;

[0133]

[0134] where k represents the kernel function of the mapping function, represents all the source domain data in S i X T represents n T all the target domain data in it, represents the feature map obtained by mapping the source domain data to the Hilbert space through the mapping function for the original samples, represents the feature map obtained by mapping the target domain data to the Hilbert space through the mapping function for the original samples;

[0135] Minimize the maximum mean discrepancy to achieve the distribution alignment of the effective radio frequency signals between the multiple source domains and the target domain, and obtain the first loss function L between all the source domains and the target domain mmd :

[0136]

[0137] where N represents the number of source domains, and both represent the domain concatenation features, represents and the maximum mean discrepancy between them.

[0138] Among them, for constructing the domain-specific classifier, inputting all the domain-specific features of the source domains into the domain-specific classifier for classification, outputting the second loss function of the domain-specific classifier, and using the domain-specific classifier to predict the processed target domain to obtain the third loss function, specifically including:

[0139] Construct a domain-specific classifier, input all the domain-specific features of the source domains into the domain-specific classifier for classification, calculate and output multiple second loss functions \(L\) of the domain-specific classifier according to the concatenated features of each domain cls :

[0140]

[0141] Among them, \(N\) represents the number of source domains, \(i\) represents the \(i\)-th source domain, \(S\) i represents the source domain data of the \(i\)-th source domain, \(E\) represents the cross-entropy loss, represents the concatenated domain features, \(Y\) Si represents the true label of the source domain data of the \(i\)-th source domain, represents \(S\) i 's predicted label;

[0142] Input the domain-specific features of the target domain into the domain-specific classifier, and the domain-specific classifier classifies the domain-specific features of the target domain through all the second loss functions and outputs multiple prediction probabilities;

[0143] Construct the third loss function according to all the prediction probabilities and all the concatenated domain features:

[0144]

[0145] Among them, \(L\) disc represents the third loss function, \(j\) represents the \(j\)-th source domain data other than \(i\), represents \(S\) j 's predicted label.

[0146] Among them, for training the domain-specific classifier using the first loss function, multiple second loss functions and the third loss function to obtain an integrated domain-specific classifier, and outputting the sample features of the processed target domain through the integrated domain-specific classifier, specifically including:

[0147] Construct the total loss function \(L\) according to the first loss function, multiple second loss functions and the third loss function:

[0148] \(L = L\) mmd +\(L\) cls +\(L\) disc ;

[0149] Input the total loss function into the domain-specific classifier for training to obtain an integrated domain-specific classifier, and input all the prediction probabilities into the integrated domain-specific classifier to output the processed sample features of the target domain.

[0150] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a feature extraction program based on multi-center domain radio frequency signals. When the feature extraction program based on multi-center domain radio frequency signals is executed by a processor, the steps of the above-mentioned feature extraction method based on multi-center domain radio frequency signals are implemented.

[0151] In summary, the present invention provides a feature extraction method and related devices based on multi-center domain radio frequency signals. The method includes: obtaining valid radio frequency signals of a target domain and multiple source domains, and inputting each of the valid radio frequency signals into a constructed multi-layer deep residual network to output corresponding spliced radio frequency features; mapping each of the spliced radio frequency features into multiple source domain feature spaces to obtain corresponding domain-specific features, and obtaining risk factor features, and splicing the risk factor features with all the domain-specific features to obtain multiple domain-spliced features; using all the domain-spliced features to perform distribution alignment processing on the valid radio frequency signals of the multiple source domains and the valid radio frequency signals between the target domains to obtain a first loss function; constructing a domain-specific classifier, inputting all the domain-specific features of the source domains into the domain-specific classifier for classification to output multiple second loss functions, and using the domain-specific classifier to predict the processed target domain to obtain a third loss function; using the first loss function, the multiple second loss functions and the third loss function to train the domain-specific classifier to obtain an integrated domain-specific classifier, and outputting the processed sample features of the target domain through the integrated domain-specific classifier. The classifier predicts the processed target domain to obtain a third loss function; using the first loss function, the multiple second loss functions and the third loss function to train the domain-specific classifier to obtain an integrated domain-specific classifier, and outputting the sample features of the target domain through the integrated domain-specific classifier. By performing feature extraction and splicing operations on the radio frequency signals of the multi-center domain, and then using the distribution features to construct a loss function to train the classifier, the present invention obtains an integrated multiple domain-specific classifiers, thereby improving the feature accuracy of the target domain.

[0152] It should be noted that in this text, the term "comprise", "include" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or terminal comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or terminal. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or terminal comprising such element.

[0153] Of course, those of ordinary skill in the art can understand that all or part of the processes of implementing the above-described embodiment methods can be completed by instructing relevant hardware (such as a processor, a controller, etc.) through a computer program. The program can be stored in a computer-readable storage medium readable by a computer. When the program is executed, it can include the processes of the above-described method embodiments. The computer-readable storage medium can be a memory, a magnetic disk, an optical disk, etc.

[0154] It should be understood that the application of the present invention is not limited to the above examples. For those of ordinary skill in the art, improvements or transformations can be made according to the above description. All such improvements and transformations should fall within the protection scope of the appended claims of the present invention.

Claims

1. A feature extraction method based on multi-center domain radio frequency signals, characterized in that: The feature extraction method based on multi-center domain radio frequency signals includes: Acquire valid radio frequency signals of a target domain and multiple source domains, input each of the valid radio frequency signals into a constructed multi-layer deep residual network, and output corresponding concatenated radio frequency features; Mapping each of the spliced ​​radio frequency features to multiple source domain feature spaces to obtain corresponding domain-specific features, and obtaining risk factor features, and splicing the risk factor features with all the domain-specific features to obtain multiple domain spliced ​​features; Using all the domain splicing features, distribution alignment processing is performed on a plurality of valid radio frequency signals in the source domains and valid radio frequency signals in the target domain to obtain a first loss function; Constructing a domain-specific classifier, inputting all domain-specific features of the source domain into the domain-specific classifier for classification, outputting a plurality of second loss functions, and using the domain-specific classifier to predict the processed target domain to obtain a third loss function; The domain-specific classifier is trained by using the first loss function, a plurality of the second loss functions and the third loss function to obtain an integrated domain-specific classifier, and the processed sample features of the target domain are output by the integrated domain-specific classifier.

2. The feature extraction method based on multi-center domain radio frequency signals according to claim 1 is characterized in that: The obtaining of effective radio frequency signals of the target domain and the plurality of source domains specifically includes: Acquire a radio frequency signal of a target domain and radio frequency signals of multiple source domains; An interception operation is performed on each of the radio frequency signals to obtain a valid radio frequency signal of the target domain and a valid radio frequency signal of each of the source domains.

3. The feature extraction method based on multi-center domain radio frequency signals according to claim 2 is characterized in that: The step of obtaining effective RF signals of the target domain and multiple source domains, inputting each of the effective RF signals into the constructed multi-layer deep residual network, and outputting corresponding concatenated RF features specifically includes: Acquire a valid radio frequency signal of a target domain and valid radio frequency signals of multiple source domains, and pair the valid radio frequency signal of the target domain with the valid radio frequency signal of each of the source domains to obtain multiple radio frequency signal groups, wherein the number of the radio frequency signal groups is the same as the number of the source domains; Inputting all the RF signal groups into the constructed multi-layer deep residual network respectively, and the multi-layer deep residual network performs convolution processing on all the RF signal groups using multiple convolution layers to obtain corresponding first RF features; The multi-layer deep residual network uses a pooling layer and a convolution layer to perform pooling convolution processing on all the RF signal groups to obtain corresponding second RF features; The multi-layer deep residual network concatenates each of the first RF features with the corresponding second RF features to obtain multiple concatenated RF features.

4. The feature extraction method based on multi-center domain radio frequency signals according to claim 1 is characterized in that: The step of mapping each of the spliced ​​radio frequency features to multiple source domain feature spaces to obtain corresponding domain-specific features, and obtaining risk factor features, and splicing the risk factor features with all the domain-specific features to obtain multiple domain splicing features specifically includes: Constructing a source domain feature space of each of the source domains and a source domain feature space of the target domain, and mapping the concatenated RF features of each of the source domains and the concatenated RF features of the target domain to the corresponding source domain feature space to obtain corresponding domain-specific features; Inputting the domain-specific features into the constructed domain feature extraction model, and extracting risk factor features through the fully connected layer in the domain feature extraction model; Through the domain feature extraction model, the risk factor features are spliced ​​with all the domain-specific features respectively to obtain corresponding domain splicing features.

5. The feature extraction method based on multi-center domain radio frequency signals according to claim 4 is characterized in that: The method of utilizing all the domain splicing features to perform distribution alignment processing on the effective RF signals of the multiple source domains and the effective RF signals of the target domain to obtain a first loss function specifically includes: Obtain a source distribution p of all effective RF signals in the source domain and a target distribution q of effective RF signals in the target domain, and calculate a maximum average difference D(p,q) between the target domain and all the source domains based on the multiple source distributions p and the target distribution q: Where i represents the i-th source domain, n i represents the number of source domain data from source domain i, and m represents the number of source domain data from n i One of them, It represents the feature map of the original sample mapped to the Hilbert space by the mapping function, S i represents the source domain data of the i-th source domain, Indicates S i Each source domain data in, T represents the number of target domains, n T represents the Tth target domain, Indicates n T Each target domain data in; Wherein, k represents the kernel function of the mapping function, Indicates S i All source domain data in X T Indicates n T All target domain data in It means that the source domain data is mapped to the feature map of the Hilbert space through the mapping function. It means that the target domain data is mapped to the feature map of the Hilbert space by the mapping function. The maximum average difference is minimized to achieve distribution alignment of effective radio frequency signals between multiple source domains and the target domain, and a first loss function L between all the source domains and the target domain is obtained. mmd : Where N represents the number of source domains, and Both represent domain splicing features. express and The maximum average difference between .

6. The feature extraction method based on multi-center domain radio frequency signals according to claim 5 is characterized in that: The constructing of a domain-specific classifier, inputting all domain-specific features of the source domain into the domain-specific classifier for classification, outputting a second loss function of the domain-specific classifier, and using the domain-specific classifier to predict the processed target domain to obtain a third loss function specifically includes: Construct a domain-specific classifier, input all the domain-specific features of the source domain into the domain-specific classifier for classification, and calculate and output multiple second loss functions L of the domain-specific classifier according to each of the domain splicing features. cls : Where N represents the number of source domains, i represents the i-th source domain, and S i represents the source domain data of the i-th source domain, E represents the cross entropy loss, represents the domain splicing feature, represents the true label of the source domain data of the i-th source domain, Indicates S i The predicted label of Inputting the domain-specific features of the target domain into the domain-specific classifier, wherein the domain-specific classifier classifies the domain-specific features of the target domain by using all the second loss functions and outputs a plurality of prediction probabilities; Based on all the predicted probabilities and all the domain concatenation features, a third loss function is constructed: Among them, L disc represents the third loss function, j represents the jth source domain data except i, Indicates S j The predicted label of .

7. The feature extraction method based on multi-center domain radio frequency signals according to claim 6 is characterized in that: The using the first loss function, the plurality of the second loss functions and the third loss function to train the domain-specific classifier to obtain an integrated domain-specific classifier, and outputting the processed sample features of the target domain through the integrated domain-specific classifier specifically includes: According to the first loss function, the plurality of the second loss functions and the third loss function, a total loss function L is constructed: L=L mmd +L cls +L disc ; The total loss function is input into the domain-specific classifier for training to obtain an integrated domain-specific classifier, and all the predicted probabilities are input into the integrated domain-specific classifier, and the processed sample features of the target domain are output.

8. A feature extraction system based on multi-center domain radio frequency signals, characterized in that: The feature extraction system based on multi-center domain radio frequency signals includes: A signal splicing module, used to obtain effective RF signals from a target domain and multiple source domains, and input each of the effective RF signals into a constructed multi-layer deep residual network, and output corresponding spliced ​​RF features; A feature splicing module, used to map each of the spliced ​​RF features to multiple source domain feature spaces to obtain corresponding domain-specific features, and obtain risk factor features, and splice the risk factor features with all the domain-specific features to obtain multiple domain splicing features; A data alignment module, used to utilize all the domain splicing features to perform distribution alignment processing on the effective radio frequency signals between the plurality of source domains and the target domain to obtain a first loss function; a feature prediction module, configured to construct a domain-specific classifier, input all domain-specific features of the source domain into the domain-specific classifier for classification, output a plurality of second loss functions, and use the domain-specific classifier to predict the processed target domain to obtain a third loss function; A feature output module is used to train the domain-specific classifier using the first loss function, multiple second loss functions and the third loss function to obtain an integrated domain-specific classifier, and output the sample features of the target domain through the integrated domain-specific classifier.

9. A terminal, characterized in that: The terminal includes: a memory, a processor, and a feature extraction program based on multi-center domain radio frequency signals stored in the memory and executable on the processor. When the feature extraction program based on multi-center domain radio frequency signals is executed by the processor, the steps of the feature extraction method based on multi-center domain radio frequency signals as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a feature extraction program based on multi-center domain radio frequency signals, and when the feature extraction program based on multi-center domain radio frequency signals is executed by the processor, the steps of the feature extraction method based on multi-center domain radio frequency signals according to any one of claims 1 to 7 are implemented.

Images