Data-driven data processing method and system, electronic device and medium

Abstract

Embodiments of the present application provide a data processing method and system based on data driving, electronic equipment and medium, belonging to the technical field of computer. The method comprises: acquiring first sample data with labels, and inputting a corresponding channel state signal to a pre-set data recognition model; the data recognition model is obtained by training a plurality of unlabeled second sample data; the number of first sample data is less than the number of second sample data; the similarity of the amplitude value and the phase value of the channel state signal is calculated, the parameters of the classification module are adjusted, and a first loss value is obtained; the adjacent phase similarity and the adjacent amplitude similarity are calculated, and the parameters of the classification module are adjusted according to the adjacent phase similarity and the adjacent amplitude similarity, and a second loss value is obtained; the parameters of the feature extraction module are fixed, and the parameters of the classification module are fixed according to the adjusted first loss value and the second loss value, and a target data recognition model is obtained; the target data is input into the target data recognition model, and a data recognition result is obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, and particularly relates to a data processing method and system based on data driving, an electronic device and a medium. BACKGROUND

[0002] Personnel recognition or detection technology refers to a technical method of automatically recognizing, detecting and analyzing human bodies or information related to human bodies through computer technology and image processing technology. Personnel recognition or detection technology can extract and analyze various features of human bodies based on a camera or image data, realize acquisition and determination of identity, attributes and behaviors of human bodies, and has strong application value.

[0003] In a data-driven application scenario, such as a WiFi sensing scenario, WiFi signals are subjected to multipath propagation in a deployment environment, and then the signals after the propagation path are superimposed and analyzed by a data recognition model to detect or recognize a sensing target. However, due to the influence of the deployment environment on the device, and the influence of the differences of sensing individuals and the relative position and orientation of the device, in order to reduce the influence of the environment on the detection result, a large amount of sample data needs to be invested for training, and a data recognition model also needs to be set up for each scene, which greatly reduces the efficiency of data recognition. SUMMARY

[0004] The main purpose of the embodiments of the present application is to provide a data processing method and system based on data driving, an electronic device and a medium, which can improve the ability of the model to adapt to different environments, and improve the efficiency and accuracy of data recognition of the data recognition model.

[0005] To achieve the above object, a first aspect of the embodiment of the present application provides a data processing method based on data driving, which comprises: acquiring first sample data with labels and a channel state signal corresponding to the first sample data, and inputting the channel state signal into a pre-set data recognition model; wherein the data recognition model is obtained by feature learning training of a plurality of unlabeled second sample data; the number of the first sample data is less than the number of the second sample data; the data recognition model is provided with a feature extraction module and a classification module; the amplitude value and the phase value of the channel state signal are calculated through the data recognition model, and the similarity of the amplitude value and the phase value is calculated; the parameters of the classification module are adjusted according to the similarity of the amplitude value and the phase value, to obtain a first loss value; the adjacent phase value of the phase value is acquired, and the adjacent phase similarity of the phase value and the adjacent phase value is calculated; the adjacent amplitude value of the amplitude value is acquired, and the adjacent amplitude similarity of the amplitude value and the adjacent amplitude value is calculated; the parameters of the classification module are adjusted according to the adjacent phase similarity and the adjacent amplitude similarity, to obtain a second loss value; the parameters of the feature extraction module are fixed, and the parameters of the classification module are fixed according to the adjusted first loss value and the second loss value, to obtain a target data recognition model; target data is acquired, the target data is input into the target data recognition model, and a data recognition result is obtained.

[0006] According to some embodiments of the present application, after the first sample data with labels and the channel state signal corresponding to the first sample data are acquired, the method further comprises: establishing a local Gaussian noise model, and performing noise analysis on the channel state signal through the local Gaussian noise model to obtain a noise analysis result; identifying a noise signal according to the noise analysis result; mapping the noise signal to a high-dimensional feature space, and fitting the noise signal using a linear regression model to obtain a fitting result; removing the noise signal in the channel state signal according to the fitting result.

[0007] According to some embodiments of the present application, the data recognition model is trained by the following steps: obtaining second sample data, inputting the second sample data into the data recognition model; performing feature detection on the second sample data to obtain probability distribution values of the second sample data; analyzing channel state signals of the second sample data to obtain amplitude values and phase values corresponding to each of the second sample data; calculating the probability distribution of the amplitude values and the phase values in the probability distribution values to obtain amplitude probability distribution values and phase probability distribution values; calculating amplitude entropy corresponding to the amplitude probability distribution values and phase entropy corresponding to the phase probability distribution values according to the amplitude probability distribution values and the phase probability distribution values; calculating first loss values and second loss values according to the amplitude entropy and the phase entropy; calculating a total loss value according to the first loss values and the second loss values, and fixing parameters of the data recognition model according to the total loss value to obtain the trained data recognition model.

[0008] According to some embodiments of the present application, the data recognition model further comprises a time attention module, and the obtaining second sample data and inputting the second sample data into the data recognition model comprises: extracting amplitude entropy images and phase entropy images from each channel state signal of the second sample data; extracting subcarriers of the channel state signals and features of the channel state signals in the time dimension according to the amplitude entropy images and the phase entropy images to obtain attention feature maps of the subcarriers in the time dimension; multiplying the amplitude entropy images and the phase entropy images with the attention feature maps respectively to obtain amplitude entropy feature maps and phase entropy feature maps; and inputting the amplitude entropy feature maps and the phase entropy feature maps into the data recognition model.

[0009] According to some embodiments of the present application, the first loss value is a distortion loss value, the distortion loss value includes a consistency loss value, a sharpness loss value and a diversity loss value; the distortion loss value is calculated by the following steps: according to the amplitude probability distribution value and the phase probability distribution value, calculating an amplitude distribution divergence value and a phase distribution divergence value; according to the amplitude distribution divergence value and the phase distribution divergence value corresponding to each sample data, accumulating to obtain a divergence sum, and according to the divergence sum divided by the number of samples, obtaining an average divergence value; according to the average divergence value divided by 2, obtaining a consistency loss value; the amplitude entropy of each sample data is accumulated and then divided by the number of samples to obtain an average amplitude entropy; the phase entropy of each sample data is accumulated and then divided by the number of samples to obtain an average phase entropy; the average amplitude entropy and the average phase entropy are added and then divided by 2 to obtain a first entropy average value of each sample data; according to the first entropy average value multiplied by a first weight hyperparameter, obtaining a sharpness loss value; the amplitude probability distribution values corresponding to each sample data are added to obtain an amplitude probability distribution total value; the phase probability distribution values corresponding to each sample data are added to obtain a phase probability distribution total value; according to the amplitude probability distribution total value divided by the number of samples, obtaining an average amplitude probability distribution value; according to the phase probability distribution total value divided by the number of samples, obtaining an average phase probability distribution value; calculating a first amplitude entropy of the average amplitude probability distribution value and a first phase entropy of the average phase probability distribution value; the first amplitude entropy and the first phase entropy are added and then divided by 2 to obtain a second entropy average value corresponding to each sample data; the second entropy average value is multiplied by a second weight hyperparameter to obtain a diversity loss value; the consistency loss value, the sharpness loss value and the diversity loss value are added to obtain a distortion loss value.

[0010] According to some embodiments of the present application, the second loss value is a geometric consistency loss value, the geometric consistency loss value is calculated by the following steps: calculating the similarity of the amplitude probability distribution value and the phase probability distribution value; constructing a similarity matrix according to the similarity, and mapping the similarity matrix to a low-dimensional space to obtain a geometric embedding result; wherein the geometric embedding result is used to represent the position of each amplitude probability distribution value or phase probability distribution value in a geometric space; constructing a second loss function through the geometric embedding result and the similarity; according to the second loss function, calculating a second loss value.

[0011] According to some embodiments of the present application, the similarity of the amplitude probability distribution value and the phase probability distribution value is calculated, including: obtaining a phase value of the second sample data and adjacent phase values of the phase value, an amplitude value and adjacent amplitude values of the amplitude value; wherein the adjacent phase values and the adjacent amplitude values are multiple; a first similarity of the phase value and one of the adjacent phase values is calculated according to a positive similarity function, and a second similarity of the amplitude value and the multiple adjacent amplitude values is calculated; and a normalized similarity is obtained by dividing the first similarity by an accumulated value of the second similarity.

[0012] To achieve the above object, the second aspect of the present application proposes a data processing system based on data driving, which comprises: a first sample data acquisition module, configured to acquire labeled first sample data and a channel state signal corresponding to the first sample data, and input the channel state signal into a pre-set data recognition model; wherein the data recognition model is obtained by feature learning and training of multiple unlabeled second sample data; the number of the first sample data is less than the number of the second sample data; the data recognition model is provided with a feature extraction module and a classification module; a similarity calculation module, configured to calculate an amplitude value and a phase value of the channel state signal through the data recognition model, and calculate a similarity of the amplitude value and the phase value; a first loss value acquisition module, configured to adjust parameters of the classification module according to the similarity of the amplitude value and the phase value, and obtain a first loss value; an adjacent phase similarity acquisition module, configured to obtain adjacent phase values of the phase value, and calculate an adjacent phase similarity of the phase value and the adjacent phase values; an adjacent amplitude similarity acquisition module, configured to obtain adjacent amplitude values of the amplitude value, and calculate an adjacent amplitude similarity of the amplitude value and the adjacent amplitude values; a second loss value acquisition module, configured to adjust the parameters of the classification module according to the adjacent phase similarity and the adjacent amplitude similarity, and obtain a second loss value; a target data recognition model acquisition module, configured to fix the parameters of the feature extraction module, and fix the parameters of the classification module according to the adjusted first loss value and the second loss value, and obtain a target data recognition model; and a data recognition result acquisition module, configured to acquire target data, input the target data into the target data recognition model, and obtain a data recognition result.

[0013] To achieve the above object, the third aspect of the present application proposes an electronic device, which comprises a memory and a processor, the memory stores a computer program, and the processor implements the data processing method based on data driving of the first aspect of the present application when executing the computer program.

[0014] To achieve the above object, a computer readable medium is provided in the fourth aspect of the embodiments of the present application, the medium stores a computer program, and the computer program is executed by a processor to implement the data processing method based on data driving in the first aspect of the present application.

[0015] The data processing method, system, electronic device and medium based on data driving provided by the present application can learn the sample data features through the data recognition model, fix the parameters of the feature extraction module of the model, and train the labeled sample data, so that the data recognition model can adjust the parameters of the classification module, and then has the ability to classify the input sample data. Therefore, only a small amount of sample data needs to be input to train the data recognition model, so that the data recognition model has strong feature extraction and data classification capabilities, thereby not being limited to the environment and the availability of sample data, so that the model is suitable for various environments, and the efficiency and accuracy of the data recognition model for target data recognition are greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a structural schematic diagram of the data processing system based on data driving provided by the embodiments of the present application;

[0017] Figure 2 is a flowchart of the data processing method based on data driving provided by the embodiments of the present application;

[0018] Figure 3 is a flowchart after the first sample data with labels and the channel state signal corresponding to the first sample data are obtained provided by the embodiments of the present application;

[0019] Figure 4 is a flowchart of the data recognition model training step provided by the embodiments of the present application;

[0020] Figure 5 is Figure 4 a flowchart of step S301 in

[0021] Figure 6 is a schematic diagram of the time attention module provided by the embodiments of the present application;

[0022] Figure 7 is a step flowchart of calculating the warping loss value provided by the embodiments of the present application;

[0023] Figure 8 is a step flowchart of calculating the geometric consistency loss value provided by the embodiments of the present application;

[0024] Figure 9 is Figure 8 a flowchart of step S601 in

[0025] Figure 10 FIG. 7 is a process schematic diagram of a data recognition model pre-training stage provided by an embodiment of the present application;

[0026] Figure 11 FIG. 8 is a process schematic diagram of a data recognition model fine-tuning stage provided by an embodiment of the present application;

[0027] Figure 12 FIG. 9 is a process schematic diagram of a data recognition model decision-making stage provided by an embodiment of the present application;

[0028] Figure 13 FIG. 10 is a functional module schematic diagram of a data processing system based on data driving provided by an embodiment of the present application;

[0029] Figure 14 FIG. 11 is a hardware structure schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0031] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification and claims and the above drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0033] In a data-driven application scenario, the samples of data can be recognized. Specifically, the labels or categories of the samples can be recognized by analyzing and learning the data samples using algorithms or models. Such a data-driven method can help to automate and optimize the process of label recognition, and improve accuracy and efficiency.

[0034] However, in some more complex perception scenarios, traditional data-driven models often fail to perform more complex perception tasks. The relative position, orientation and environmental characteristics of the perception target and the perception device will affect the transmission of the signal, so that the effective features of the signal are lost in the received signal. Therefore, the model needs to perform effective feature extraction on the received signal. Moreover, since a model for filtering data is usually only applicable to a specific environment or relative position, it cannot effectively filter in other environments or positions. Therefore, a model can only be applied to one scene, which not only increases the difficulty of deployment, but also increases the cost of data recognition.

[0035] In addition, in some cases, it is difficult to obtain sample data, such as a model for detecting the falling action of a person. The acquisition of sample data for a person falling is often unacceptable. Therefore, the limitation of the amount of sample data greatly reduces the accuracy of the model in recognizing data.

[0036] Therefore, the embodiments of the present application provide a data processing method, system, electronic device and medium based on data driving, which can learn the characteristics of sample data through a data recognition model, fix the parameters of the feature extraction module of the model, and train the labeled sample data. The data recognition model can adjust the parameters of the classification module, thereby having the ability to classify the input sample data. Therefore, only a small amount of sample data is needed to train the data recognition model, so that the data recognition model has strong feature extraction and data classification capabilities. Therefore, it is not necessary to be limited to the availability of the environment and sample data, so that the model is suitable for a variety of environments, greatly improving the efficiency and accuracy of the data recognition model in recognizing target data.

[0037] The data processing method, system, electronic device and medium based on data driving provided by the embodiments of the present application are specifically described through the following embodiments. First, the data processing system based on data driving in the embodiments of the present application is described.

[0038] Please refer to Figure 1 In some embodiments, the data processing system based on data driving includes a first pre-training module 101, a data acquisition module 102, a data processing module 103, a second pre-training module 104, a data updating module 105, a decision output module 106, and a control module 107.

[0039] In some embodiments, the control module 107 can be the nerve center and command center of the system. The control module 107 can generate operation control signals according to instruction operation codes and timing signals to complete the control of instruction fetching and instruction execution. For example, the control module 107 can control the first pre-training module 101, the data acquisition module 102, the data processing module 103, the second pre-training module 104, the data updating module 105, and the decision output module 106 according to instructions to ensure that the data recognition model can smoothly train and recognize sample data.

[0040] In some embodiments, the first pre-training module 101 is used to pre-train the data recognition model in the pre-training stage. Specifically, the data recognition model can be trained by the amplitude values and phase values of the unmarked sample data and related parameters such as amplitude probability distribution and phase probability distribution, so that the data recognition model has good feature extraction capability and sample classification capability. In some embodiments, the first pre-training module 101 can continuously optimize the parameters of the data recognition model by calculating the first loss function value and the second loss function value during training until the convergence condition is reached or the preset training times are reached, so that the data recognition model has good feature extraction capability and classification capability.

[0041] In some embodiments, the data acquisition module 102 is used to acquire or collect the first sample data, the second sample data, and the target data, and input the corresponding data into the data recognition model for training or recognition.

[0042] In some embodiments, the data processing module 103 can perform data cleaning operations on the data, such as removing outliers, filling missing values, and removing duplicates for the target data to ensure the quality and consistency of the target data; or mapping the sample data from high dimension to low dimension, or adjusting the similarity of the sample data, etc.

[0043] In some embodiments, the second pre-training module 104 is used to train a small amount of labeled samples, so that the data recognition model can recognize the label of the corresponding environment or scene, correspond the extracted sample features to the label, and fix the parameters of the feature extraction module and fine-tune the parameters of the classification module.

[0044] In some embodiments, the data updating module 105 is used to adjust and update the parameters of the data recognition model according to the first loss function and the second loss function during training.

[0045] In some embodiments, the decision output module 106 is used to output the recognition result of the target data. For example, if the action is recognized, the corresponding action is output, if the identity is recognized, the corresponding identity information such as name is output, etc.

[0046] The data processing method based on data driving in the embodiments of the present application can be illustrated by the following embodiments.

[0047] It should be noted that in each specific embodiment of the present application, when it is necessary to process relevant data related to the identity or characteristics of the user according to user information, user behavior data, user historical data, and user location information, the user's permission or consent will be obtained first, for example, when accessing the user's stored data and the user's cache data access request, the user's permission or consent will be obtained first. Moreover, the collection, use and processing of these data will comply with relevant laws, regulations and standards. In addition, when the embodiments of the present application need to obtain sensitive personal information of the user, the separate permission or separate consent of the user will be obtained through a pop-up window or by jumping to a confirmation page, and after obtaining the separate permission or separate consent of the user, the necessary user-related data for the normal operation of the embodiments of the present application will be obtained.

[0048] Figure 2 is an optional flowchart of the data processing method based on data driving provided by the embodiments of the present application, Figure 2 The method in can include but not limited to steps S101 to S108.

[0049] Step S101, acquiring labeled first sample data and channel state signals corresponding to the first sample data, and inputting the channel state signals into a pre-set data recognition model; wherein the data recognition model is obtained by feature learning and training of a plurality of unlabeled second sample data; the number of first sample data is less than the number of second sample data; the data recognition model is provided with a feature extraction module and a classification module;

[0050] Step S102, calculating the amplitude value and the phase value of the channel state signal by the data recognition model, and calculating the similarity of the amplitude value and the phase value;

[0051] Step S103, adjusting the parameters of the classification module according to the similarity of the amplitude value and the phase value, to obtain a first loss value;

[0052] Step S104, acquiring adjacent phase values of the phase value, and calculating the adjacent phase similarity of the phase value and the adjacent phase value;

[0053] Step S105, acquiring adjacent amplitude values of the amplitude value, and calculating the adjacent amplitude similarity of the amplitude value and the adjacent amplitude value;

[0054] Step S106, adjusting the parameters of the classification module according to the adjacent phase similarity and the adjacent amplitude similarity, to obtain a second loss value;

[0055] Step S107, fixing the parameters of the feature extraction module, and fixing the parameters of the classification module according to the adjusted first loss value and the second loss value, to obtain a target data recognition model;

[0056] Step S108, obtaining target data, inputting the target data into the target data recognition model, and obtaining a data recognition result.

[0057] In some embodiments, the classification module can predict the category to which the sample belongs according to the feature vector of the sample. In some embodiments, the classification module is a classifier, where the classifier can be a linear classifier and a nonlinear classifier, the linear classifier can be a logistic regression, a support vector machine, etc., and the nonlinear classifier can be a decision tree, a neural network, etc.

[0058] It can be understood that the feature extraction module refers to an algorithm or model used to extract useful features from input data in the field of machine learning and computer vision. In some embodiments, the feature extraction module can be a feature extractor, which can extract features in different ways, for example, an image-based feature extractor can use various filters, edge detectors, texture descriptors, etc. to capture structures and patterns in images; while a text-based feature extractor can use bag-of-words model, TF-IDF, etc. to extract keywords and phrases in text.

[0059] It can be understood that the training or recognition of the sample using the data recognition model can be divided into three stages. Specifically, the first stage is a pre-training stage, corresponding to the second sample data, at this time the data recognition model is trained on a large amount of unlabeled sample data, and the parameters of the first loss function and the second loss function are adjusted in the process, so that the data recognition model can have strong feature extraction and classification capabilities without sample labels; the second stage is a fine-tuning stage, corresponding to the first sample data, at this time, only a small amount of sample data with labels, such as 1 to 3 sample data, is needed, so that the data recognition model can fix the parameters of the feature extraction module and fine-tune the parameters of the classification module, and the data recognition model can quickly extract features and classify labels according to the labels of the sample data, that is, the data recognition model can perform action recognition or other recognition such as identity recognition through training of a small amount of samples in any scene; the third stage is a decision stage, corresponding to the target data, at this time, the data recognition model has completed the training of the target recognition of a specific scene, such as human action training, and can directly recognize the input target data, that is, can be directly put into use.

[0060] In some embodiments, before the first sample data with labels is input into the data recognition model, the data recognition model has been trained by the second sample data to obtain strong feature extraction capability and sample classification capability. Specifically, the data recognition model continuously optimizes parameters in the training process according to a twist loss function, and obtains strong feature extraction capability for sample data. The similarity of the amplitude probability distribution and the phase probability distribution of the sample data is calculated by a geometric consistency loss function, thereby simulating the relationship between the sample data and the sample labels, so that the data recognition model has the ability to recognize labels according to sample data.

[0061] In some embodiments, the number of the first sample data is less than the number of the second sample data, because the second sample data is used to train the initial data recognition model, and the initial data recognition model needs to be trained by a large amount of sample data to have strong feature recognition capability and sample classification capability. It can be understood that the second sample data is unlabeled sample data without labels, because the role of the second sample data is to enable the data recognition model to have the ability to be applicable to any behavior recognition scene or detection scene through training, so that when the data recognition model is trained by the second sample data, the parameters of the feature extraction module will be fixed, but the parameters of the classification module will not be fixed.

[0062] It can be understood that the first sample data is labeled sample data, and further, the first sample data is any data that needs to be classified and recognized. For example, the first sample data can be action data, and the corresponding label is falling down, waving hand, etc. The first sample data can also be related to personnel detection data, such as portrait data, and the label is name, etc. It can be understood that since the data recognition model has strong feature extraction capability and sample classification capability after being trained by the second sample data, that is, it has strong feature extraction capability and sample classification capability through semi-supervised learning, therefore, the first sample data can be 1 to 3, for example, 2, so that the data recognition model can directly recognize the corresponding data according to the corresponding sample data and labels. It can be understood that in order to also increase the number of first sample data according to actual needs, the present application does not make specific limitations thereto.

[0063] In some embodiments, the amplitude value and the phase value of the channel state signal can be calculated by the data recognition model, and the similarity of the amplitude value and the phase value can be calculated. In some embodiments, the similarity of the amplitude value and the phase value can be calculated by Euclidean distance, cosine similarity, etc., and the parameters of the classification module can be adjusted according to the similarity, so that the classification module increases the extraction ability of the characteristics of the same sample data. For example, when the similarity of the amplitude value and the phase value of the sample data reaches a preset threshold, the parameters of the classification module can be adjusted, that is, by adjusting the parameters of the classification module, the data recognition model can quickly extract the same characteristics when identifying the same sample data, so as to improve the classification ability of the sample data and improve the efficiency of subsequent training of the data classification model.

[0064] In some embodiments, the adjacent phase similarity of the phase value and the adjacent phase value, and the adjacent amplitude similarity of the amplitude value and the adjacent amplitude value can be calculated, and the parameters of the classification module can be adjusted according to the adjacent phase similarity and the adjacent amplitude similarity. It can be understood that if the adjacent phase similarity is high, the adjacent amplitude similarity should also be the same as or similar to the adjacent phase similarity, and at this time, the two sample data can be classified into the same type of sample data by adjusting the parameters of the classification module, so as to realize efficient classification of the sample data. It can be understood that if the adjacent phase similarity is low, the adjacent amplitude similarity should also be the same as or similar to the adjacent phase similarity, and at this time, the two sample data can be classified into different types of sample data by adjusting the parameters of the classification module, so as to realize efficient classification of the sample data.

[0065] It can be understood that in order to enable the data recognition model to extract common characteristics of the amplitude entropy and the phase entropy, a first loss function, i.e., a twist loss function, can be set to adjust the similarity of the amplitude entropy and the phase entropy to obtain a first loss value, so that the data recognition model can extract common characteristics of the amplitude entropy and the phase entropy. In some embodiments, the similarity can be adjusted by using a similarity measurement method, such as cosine similarity, Euclidean distance, etc.

[0066] In some embodiments, the first loss value corresponds to a first loss function (Twist loss function) including a consistency loss value, a sharpness loss value, and a diversity loss value. It can be understood that in the process of training the data recognition model, in order to improve the extraction ability of the data recognition model to the characteristics, the consistency loss value and the sharpness loss value should be as low as possible, and the diversity loss value should be as high as possible.

[0067] In some embodiments, a neighboring phase value of the phase value can be obtained, and a neighboring similarity between the phase value and the neighboring phase value can be calculated. The neighboring similarity can be used as an adjustment target to adjust the amplitude value and the neighboring amplitude value under the first sample data, so as to obtain a second loss value. Alternatively, a neighboring amplitude value of the amplitude value can be obtained, and a neighboring similarity between the amplitude value and the neighboring amplitude value can be calculated. The neighboring similarity can be used as an adjustment target to adjust the phase value and the neighboring phase value under the first sample data, so as to obtain a second loss value.

[0068] In some embodiments, a neighboring similarity between the amplitude probability distribution value and the neighboring amplitude probability distribution value can also be calculated, and the neighboring similarity can be used as an adjustment target to adjust the similarity between the phase probability distribution value and the neighboring phase probability distribution value. Alternatively, a neighboring similarity between the phase probability distribution value and the neighboring phase probability distribution value can be calculated, and the neighboring similarity can be used as an adjustment target to adjust the similarity between the amplitude probability distribution value and the neighboring amplitude probability distribution value. It can be understood that the adjustment based on the similarity between the amplitude values or the phase values of the sample data can make the geometric structure between the amplitude values and the phase values consistent as much as possible, so as to facilitate the classification of the sample data by the data recognition model. Therefore, during the training of the data recognition model by the second sample data, the adjustment between the amplitude values and the phase values can not necessarily require that the sample data must be accompanied by sample labels, so as to limit the data recognition model to a certain scene, or to make the data recognition model spend a large amount of data for training or design a new model each time the new environment is identified. It is only necessary to make the data recognition model have classification ability through similarity adjustment at the beginning.

[0069] It can be understood that, when the first sample data is used to train the data recognition model, the feature extraction capability of the data recognition model for the data is already good enough, so that the parameters of the first loss function can not be adjusted and can be directly fixed. Alternatively, the parameters of the first loss function can be fine-tuned and fixed, and the first loss value is recalculated. The embodiments of the present application do not make specific limitations on this.

[0070] It can be understood that, after the first sample data is used to train the data recognition model, the data recognition model already has the recognition capability for the scene or environment corresponding to the first sample data, so that the target data can be directly obtained and input into the trained data recognition model, so as to obtain a data recognition result, such as judging that the corresponding person's name is Zhang San, or judging that the corresponding human body action is falling down.

[0071] The data processing method and system based on data driving, the electronic device and the medium provided in the application can learn the sample data characteristics through the data recognition model, fix the parameters of the feature extraction module of the model, and train the labeled sample data, so that the data recognition model can adjust the parameters of the classification module, and then has the ability to classify the input sample data. Therefore, only a small amount of sample data needs to be input to train the data recognition model, so that the data recognition model has strong feature extraction and data classification capabilities, thereby not being limited to the environment and the availability of sample data, so that the model is suitable for various environments, and the efficiency and accuracy of the data recognition model for target data recognition are greatly improved.

[0072] Reference Figure 3 In some embodiments, after obtaining the labeled first sample data and the channel state signal corresponding to the first sample data, the method further includes, but is not limited to, steps S201 to S204:

[0073] In step S201, a local Gaussian noise model is established, and noise analysis is performed on the channel state signal through the local Gaussian noise model to obtain a noise analysis result.

[0074] In step S202, the noise signal is identified according to the noise analysis result.

[0075] In step S203, the noise signal is mapped to a high-dimensional feature space, and a linear regression model is used to fit the noise signal to obtain a fitting result.

[0076] In step S204, the noise signal in the channel state signal is removed according to the fitting result.

[0077] In some embodiments, the characteristics of the first sample data can be described by establishing a local Gaussian noise model. The local Gaussian noise model can collect channel state signal samples within a period of time and statistically analyze the noise distribution. In some embodiments, the collected time can be divided into multiple time periods, so that the local Gaussian noise model can better simulate the noise distribution in each region. In some embodiments, the established local Gaussian noise model is used to analyze the noise of the channel state signal to obtain a noise analysis result. In some embodiments, the channel state signal can also be processed. Specifically, the signal can be divided into different local regions, and the noise in each region can be analyzed according to the local Gaussian noise model to obtain a noise analysis result.

[0078] In some embodiments, the characteristics of the noise signal can be identified according to the noise analysis result, such as the power spectral density, frequency distribution, etc. of the noise. Further, the noise signal is mapped to a high-dimensional feature space, for example, the noise signal is converted into a set of wavelet coefficients with different frequencies and phases using wavelet transform, and then the wavelet coefficients are fitted using a linear regression model, and through the fitting, a fitting result is obtained, which is a mathematical description of the noise signal in the high-dimensional feature space. In some embodiments, the noise signal can also be converted into a set of eigenvectors with higher energy using the principal component analysis (PCA) method, and the noise signal is fitted in the high-dimensional feature space using a linear regression model to obtain a fitting result, which is a mathematical description of the noise signal in the channel state signal or channel state diagram.

[0079] Further, the noise signal in the channel state signal or channel state diagram is removed according to the fitting result. Exemplarily, the influence of the noise can be reduced or eliminated by subtracting the fitting result from the original signal, thereby improving the clarity and reliability of the channel state signal or channel state diagram.

[0080] Please refer to Figure 4 In some embodiments, the data recognition model is trained by, but not limited to, the following steps S301 to S307:

[0081] Step S301, obtaining second sample data, and inputting the second sample data into the data recognition model;

[0082] Step S302, performing feature detection on the second sample data to obtain a probability distribution value of the second sample data;

[0083] Step S303, analyzing the channel state signal of the second sample data to obtain an amplitude value and a phase value corresponding to each second sample data;

[0084] Step S304, calculating the probability distribution of the amplitude value and the phase value in the probability distribution value to obtain an amplitude probability distribution value and a phase probability distribution value;

[0085] Step S305, calculating an amplitude entropy corresponding to the amplitude probability distribution value and a phase entropy corresponding to the phase probability distribution value according to the amplitude probability distribution value and the phase probability distribution value;

[0086] Step S306, calculating a first loss value and a second loss value according to the amplitude entropy and the phase entropy;

[0087] Step S307, calculating a total loss value according to the first loss value and the second loss value, and fixing the parameters of the data recognition model according to the total loss value to obtain the trained data recognition model.

[0088] In some embodiments, the second sample data refers to sample data used for initially training the data recognition model, and enabling the data recognition model to have feature extraction capability and classification capability on sample data. That is to say, before the second sample data is input into the data recognition model, the data recognition model has not been trained on sample data. In some embodiments, the loss function of the data recognition model can be set as a first loss function corresponding to a first loss value and a second loss function corresponding to a second loss value. In some embodiments, the first loss function can be a twist loss function (i.e., a Twist loss function) corresponding to a twist loss value, and the second loss function can be a geometric consistency loss function.

[0089] In some embodiments, feature detection can be performed on the second sample data to obtain the probability distribution value of the second sample data, so as to understand the occurrence frequency and distribution of each feature and perform statistical analysis on the data. Illustratively, the probability distribution value of the sample data can be obtained through statistical analysis, machine learning, etc. Illustratively, the features can be extracted from the second sample data, and then the features of the second sample data are modeled through statistical methods or machine learning algorithms to obtain the probability distribution value.

[0090] In some embodiments, when analyzing the channel state signal of the second sample data, the characteristics of the channel state signal can be analyzed by observing the change of the channel state signal over time, and specifically, the amplitude value and the phase value can be obtained by observing the waveform of each data point. It can be understood that the amplitude value represents the amplitude of the signal, which can be obtained by calculating the absolute value of each data point. For example, if the second sample data is [0.1, -0.3, 0.2, 0.4, -0.5], the corresponding amplitude value is [0.1, 0.3, 0.2, 0.4, 0.5]. In some embodiments, the phase value represents the offset of the signal over time, which can be obtained by calculating the phase angle of each data point, and the phase angle can be calculated by the inverse tangent function. For example, if the second sample data is [0.1, -0.3, 0.2, 0.4, -0.5], the corresponding phase value is the inverse tangent value of [0.1, -0.3, 0.2, 0.4, -0.5].

[0091] In some embodiments, a probability density function can be set to describe the probability distribution. Specifically, according to the given probability density function, the value range of the amplitude value or the phase value is determined, and the area in the given range is calculated to obtain the amplitude probability distribution value or the phase probability distribution value. That is to say, the probability density function can be integrated with the amplitude value or the phase value to obtain the result of integration, which represents the probability of the amplitude value or the phase value in the probability distribution.

[0092] In some embodiments, for a given amplitude probability distribution value, the probability of each amplitude value can be calculated, and the calculated probability is taken as a logarithm, the logarithm is multiplied by the corresponding probability value, and the sum of all product values is taken to obtain the inverse number, i.e. the corresponding amplitude entropy.

[0093] In some embodiments, for a given phase probability distribution value, the probability of each phase value can be calculated, and the calculated probability is taken as a logarithm, the logarithm is multiplied by the corresponding probability value, and the sum of all product values is taken to obtain the inverse number, i.e. the corresponding phase entropy.

[0094] In some embodiments, the loss function can be constructed by the amplitude entropy and the phase entropy, or the parameters associated with the amplitude entropy and the phase entropy, such as the amplitude probability distribution value and the phase probability distribution value, etc. The first loss value (i.e. the distortion loss value) and the second loss value (i.e. the geometric consistency loss value) are calculated, and then the total loss value is calculated according to the first loss value and the second loss value. In the model training, the parameters are adjusted constantly, so that the total loss value decreases, until the convergence condition of the data recognition model or the preset training times are reached, and then the training is stopped. At this time, the data recognition model has a strong feature extraction capability and a classification capability for data.

[0095] Exemplarily, the total loss function can be represented as:

[0096] L = λL t + μL g

[0097] Where λ and μ are initial hyperparameters, L t is the first loss function, and L g is the second loss function.

[0098] Please refer to Figure 5 In some embodiments, the data recognition model further includes a time attention module, and step S301 includes but is not limited to steps S401 to S404:

[0099] Step S401, extracting an amplitude entropy image and a phase entropy image from each channel state signal of the second sample data;

[0100] Step S402, extracting a subcarrier of the channel state signal and a feature of the channel state signal in the time dimension according to the amplitude entropy image and the phase entropy image, to obtain an attention feature map of the subcarrier in the time dimension;

[0101] Step S403, multiplying the amplitude entropy image and the phase entropy image with the attention feature map respectively to obtain an amplitude entropy feature map and a phase entropy feature map;

[0102] Step S404, inputting the amplitude entropy feature map and the phase entropy feature map into the data recognition model.

[0103] In some embodiments, the amplitude entropy is used to describe the diversity of amplitude changes in the signal or image, and the phase entropy image utilizes the phase information to describe the features of the image. In some embodiments, the channel state signal can be subjected to time domain discrete Fourier transform to obtain a frequency domain representation, and from the frequency domain representation, the amplitude and phase information of the channel state signal can be obtained. Illustratively, the amplitude or phase at each frequency point can be subjected to statistics and the probability distribution thereof can be calculated, and the entropy value can be calculated according to the probability distribution to obtain the amplitude entropy image or the phase entropy image.

[0104] In some embodiments, the amplitude entropy feature map is an image obtained by calculating the amplitude entropy values of the channel state signal at different times, and the phase entropy feature map is an image obtained by calculating the phase entropy values of the channel state signal at different times. It can be understood that the amplitude entropy feature map can be used to analyze the frequency distribution, periodicity, noise, etc. of the channel state signal, and the phase entropy feature map can be used to analyze the phase change, phase synchronism, etc. of the channel state signal, and the amplitude entropy feature map and the phase entropy feature map can be used in combination to provide a comprehensive description of the features of the channel state signal.

[0105] In some embodiments, in the pre-training stage, in order to enable the feature extraction module to extract key feature information from the channel state signal without labels, a subcarrier-time attention mechanism, i.e. a time attention module, can be added before the input layer of the CNN network of the data recognition model, as shown in the following formula: Figure 6 In some embodiments, the amplitude entropy image and the phase entropy image are multiplied by the attention feature map after the time attention module calculates the attention feature map of the subcarriers in the time dimension corresponding to the amplitude entropy image and the phase entropy image of the channel state signal, to obtain the adjusted attention feature map, i.e. the amplitude entropy feature map and the phase entropy feature map added with the attention feature, and the amplitude entropy feature map and the phase entropy feature map are input into the data recognition model. In some embodiments, the amplitude entropy feature map and the phase entropy feature map are weighted entropy images, which can more effectively extract important feature information from the channel state signal, and help to improve the performance and accuracy of the data recognition model.

[0106] Illustratively, for example, the data recognition model needs to recognize sample data, for example, the sample data is human falling related image signal data, then the amplitude entropy image and the phase entropy image of the falling action are multiplied by the attention feature map in the preprocessing stage through the time attention module to obtain the adjusted amplitude entropy feature map and the phase entropy feature map, so that the data recognition model pays more attention to the key features of the falling action in the recognition process, thereby improving the accuracy of action recognition.

[0107] Please refer to Figure 7 In some embodiments, the first loss value is a twist loss value, and the twist loss value includes a consistency loss value, a sharpness loss value, and a diversity loss value; the twist loss value is calculated by, but not limited to, the following steps S501 to S504:

[0108] Step S501, according to the amplitude probability distribution value and the phase probability distribution value, the amplitude distribution divergence value and the phase distribution divergence value are calculated; the amplitude distribution divergence value and the phase distribution divergence value corresponding to each sample data are accumulated to obtain a divergence sum, and the average divergence value is obtained by dividing the divergence sum by the number of samples; the consistency loss value is obtained by dividing the average divergence value by 2;

[0109] Step S502, the amplitude entropy of each sample data is accumulated and then divided by the number of samples to obtain the amplitude entropy average value; the phase entropy of each sample data is accumulated and then divided by the number of samples to obtain the phase entropy average value; the first entropy average value of each sample data is obtained by adding the amplitude entropy average value and the phase entropy average value and then dividing by 2; the sharpness loss value is obtained by multiplying the first entropy average value by the first weight hyperparameter;

[0110] Step S503, the amplitude probability distribution values corresponding to each sample data are added to obtain the amplitude probability distribution total value; the phase probability distribution values corresponding to each sample data are added to obtain the phase probability distribution total value; the amplitude probability distribution average value is obtained by dividing the amplitude probability distribution total value by the number of samples; the phase probability distribution average value is obtained by dividing the phase probability distribution total value by the number of samples; the first amplitude entropy of the amplitude probability distribution average value and the first phase entropy of the phase probability distribution average value are calculated; the second entropy average value corresponding to each sample data is obtained by adding the first amplitude entropy and the first phase entropy and then dividing by 2; the diversity loss value is obtained by multiplying the second entropy average value by the second weight hyperparameter;

[0111] Step S504, the consistency loss value, the sharpness loss value, and the diversity loss value are added to obtain the twist loss value.

[0112] In some embodiments, the twist loss value is the Twist loss value, which is used to measure the gap between the prediction results generated by the data recognition model in the training process and the true results. In some embodiments, assuming that the probability distribution values obtained from the amplitude and the phase are P1 and P2 respectively in the pre-training stage, the Twist loss function can be represented as:

[0113]

[0114] Where, D KL(·||·) represents the Kullback-Leibler divergence (KL divergence) between two probability distributions, which is used to measure the difference between the distribution predicted by the data recognition model and the true distribution, H(·) represents the information entropy of a certain probability distribution, B represents the number of samples during training, P 1,i represents the amplitude probability distribution value of the i-th sample, P 2,i represents the phase probability distribution value of the i-th sample, P k,i represents the class probability distribution of the i-th sample in a training batch in the amplitude or phase branch, and a and β are weight hyperparameters.

[0115] For the Twist loss function, represents the consistency loss value, represents the sharpness loss value, represents the diversity loss value. It can be understood that, in the process of training the data recognition model, in order to improve the feature extraction ability of the data recognition model, the consistency loss value and the sharpness loss value should be as low as possible, and the diversity loss value should be as high as possible.

[0116] Specifically, the consistency loss value measures the recognition ability of the data recognition model for similar sample data. When the consistency loss value is high, it indicates that the data recognition model can accurately cluster similar sample data and maintain consistency. Therefore, reducing the consistency loss value can improve the feature extraction ability of the data recognition model, so that similar sample data is more easily recognized by the data recognition model. In some embodiments, in order to make the predictions of the amplitude distribution probability and the phase distribution probability as consistent as possible, i.e., to keep the amplitude and phase of the sample data as consistent as possible, the value of the consistency term can be minimized, so that the data recognition model can extract similar information from the amplitude and phase.

[0117] The sharpness of the sample data refers to the distinctness and distinguishability between different categories of sample data, and the sharpness loss value measures the distinguishability of the data recognition model for different categories of sample data. It can be understood that the output distribution can be made to tend to sharpness by minimizing the amplitude entropy and phase entropy of each sample data. That is to say, each sample data is assigned a certain value, i.e., a vector in an ideal case, and at the same time, the features of sample data assigned to the same category are made more compact, i.e., similar sample features are made closer.

[0118] The diversity loss value measures the processing capability of the data recognition model for the difference between different samples. It can be understood that, when processing the diversity term, the prediction result of the sample data should be able to cover different samples, and try to avoid the network predicting all data categories as the same category. Specifically, it can be achieved by maximizing the first amplitude entropy of the amplitude probability distribution average value and the first phase entropy of the phase probability distribution average value. It can be understood that the first amplitude entropy of the amplitude probability distribution average value and the first phase entropy of the phase probability distribution average value can refer to the first amplitude entropy and the first phase entropy corresponding to all sample data within the same sample data.

[0119] Please refer to Figure 8 In some embodiments, the second loss value is a geometric consistency loss value, and the geometric consistency loss value is calculated by, but not limited to, the following steps S601 to S604:

[0120] Step S601, calculating the similarity of the amplitude probability distribution value and the phase probability distribution value;

[0121] Step S602, constructing a similarity matrix according to the similarity, and mapping the similarity matrix to a low-dimensional space to obtain a geometric embedding result; wherein the geometric embedding result is used to represent the position of each amplitude probability distribution value or phase probability distribution value in the geometric space;

[0122] Step S603, constructing a second loss function by the geometric embedding result and the similarity;

[0123] Step S604, calculating the second loss value according to the second loss function.

[0124] In some embodiments, since only using the warping function cannot make the data recognition model have strong sample data classification capability, it is necessary to construct a second loss function, so that the data recognition model can maintain the geometric structure of the amplitude entropy and the phase entropy of the sample data by utilizing the geometric structure between the sample data. It can be understood that when the geometric structure of the amplitude entropy and the phase entropy is consistent, it means that the amplitude feature and the phase feature have similar distribution and change trend in the data, and such consistency can make the classification model better capture the intrinsic characteristics of the data, thereby improving the accuracy of classification.

[0125] In some embodiments, the amplitude probability distribution value can be calculated from the amplitude value, the phase probability distribution value can be calculated from the phase value, and the similarity between the amplitude probability distribution value and the adjacent amplitude probability distribution value of the same batch of sample data, and the similarity between the phase probability distribution value and the adjacent phase probability distribution value of the same sample data can be calculated. It can be understood that, since the amplitude value and the phase value belong to the same sample data, the similarity between the amplitude probability distribution value and the adjacent amplitude probability distribution value should be consistent with the phase of the similarity between the phase probability distribution value and the adjacent phase probability distribution value.

[0126] Alternatively, the amplitude entropy can be calculated from the amplitude value, the phase entropy can be calculated from the phase value, and the similarity between the amplitude entropy and the adjacent amplitude entropy of the same batch of sample data, and the similarity between the phase entropy and the adjacent phase entropy of the same sample data can be calculated. It can be understood that, since the amplitude value and the phase value belong to the same sample data, the similarity between the amplitude entropy and the adjacent amplitude entropy should be consistent with the phase of the similarity between the phase entropy and the adjacent phase entropy.

[0127] In some embodiments, the similarity between the amplitude probability distribution values of multiple data in the same batch of sample data and the similarity between the phase probability distribution values can be calculated, and the similarity between the phase probability distribution values can be adjusted based on the similarity between the amplitude probability distribution values, that is, the geometric structure of the sample data can be directly or indirectly adjusted to make the characteristics between the same sample more obvious.

[0128] In some embodiments, a similarity matrix can be constructed according to the calculated similarity between the multiple sample data and the adjacent sample data, and the similarity matrix can be mapped to a low-dimensional space by a dimension reduction technique to obtain a geometric embedding result. In some embodiments, the dimension reduction technique can be principal component analysis, t-SNE, etc.

[0129] In some embodiments, a loss function can be constructed from the geometric embedding result and the similarity. Specifically, a target function can be defined, which includes two parts, the first part is the error of the geometric embedding result, that is, the position of the original probability distribution value is approximated by the result of the geometric embedding, and the second part is the error of the similarity, that is, the original similarity is approximated by the calculated similarity.

[0130] In some embodiments, the second loss function can be a distance loss function, an intra-class consistency loss function, an inter-class separability loss, a local structure loss, and the like, and the present embodiments are not limited in this regard. By way of example, a distance loss function is described in detail as the second loss function. In some embodiments, after obtaining the similarity of the amplitude probability distribution value and the similarity of the phase probability distribution value of each sample, the relationship between the two in the low-dimensional space can be maintained by minimizing the distance between the samples. In some embodiments, the Euclidean distance between the mapped samples can be minimized, for example, using a mean square error loss or a triangular inequality loss method.

[0131] In some embodiments, the distance between the similarity of the amplitude probability distribution value and the similarity of the phase probability distribution value can be minimized in the low-dimensional space. Specifically, let the amplitude probability distribution value of sample 1 be A1, the adjacent amplitude probability distribution value be B1, and the calculated similarity be S1; let the phase probability distribution value of sample 1 be A2, the adjacent probability distribution value be B2, and the calculated similarity be S2. Further, after mapping to the low-dimensional space, S1 can target the geometric structure of S2 for training and learning, so that the phase probability distribution value of sample 1 is close to the characteristics of the amplitude probability distribution value; or S2 can target the geometric structure of S1 for training and learning, so that the amplitude probability distribution value of sample 1 is close to the characteristics of the phase probability distribution value. In this way, the classification learning of the data recognition model under unlabeled samples can be completed, so as to facilitate the classification of sample data according to the label based on a small amount of labeled sample data (such as 1 to 3 sample data) in the future.

[0132] Please refer to Figure 9 In some embodiments, step S601 includes but is not limited to steps S701 to S703:

[0133] Step S701, obtaining the phase value and the adjacent phase value of the phase value, the amplitude value and the adjacent amplitude value of the amplitude value of the second sample data; wherein the adjacent phase value and the adjacent amplitude value are multiple;

[0134] Step S702, calculating the first similarity of the phase value and one of the adjacent phase values according to the positive value similarity function, and calculating the second similarity of the amplitude value and the multiple adjacent amplitude values;

[0135] Step S703, obtaining the normalized similarity by dividing the first similarity by the cumulative value of the second similarity.

[0136] In some embodiments, the probability distribution of the phase entropy and the probability distribution of the amplitude entropy in the sample data can be obtained, and the similarity between the sample data can be calculated by a positive similarity function. In some embodiments, the loss function of the second part of the data recognition model is a geometric consistency loss, and the second loss value is the geometric consistency loss value. It can be understood that the distortion loss function can be used to enable the model to have the feature extraction capability, but cannot learn the classification task by a small amount of labeled sample data, so the data recognition model is trained by the geometric loss function in the second stage to enable the data recognition model to have the label classification capability of the sample data.

[0137] In some embodiments, the similarity of the amplitude probability distribution value and the phase probability distribution value of each sample in the sample data can be calculated, and the geometric features in the amplitude and the geometric features in the phase of the sample data are made consistent based on the similarity between the sample data, so as to achieve the effect of classifying the sample data.

[0138] In some embodiments, if the amplitude probability distribution of a sample data exhibits similar features with the amplitude probability distribution of an adjacent sample, the phase probability distribution of the same sample data should also exhibit similar features with the phase probability distribution of the adjacent sample. For example, if the amplitude probability distribution of sample No. 1 exhibits similar features with the amplitude probability distribution of sample No. 2, the phase probability distribution of sample No. 1 should also exhibit similar features with the phase probability distribution of sample No. 2. Therefore, the similarity between the sample data can be calculated by a positive similarity function, and the geometric structure of the amplitude entropy and the phase entropy of the channel state signal is made consistent as much as possible.

[0139] For example, the similarity between the sample data can be calculated by the amplitude distribution probability corresponding to the amplitude entropy and the phase distribution probability corresponding to the phase entropy. In some embodiments, for a sample x k with a distribution P i , the geometric embedding result can be represented as:

[0140]

[0141] where q i|j,k represents the geometric structure similarity, which can be calculated by the following formula:

[0142]

[0143] where K(·||·) is a positive similarity function, and the calculation formula is as follows:

[0144]

[0145] For example, for K(P k,i , P​k,j ), the similarity can be calculated as:

[0146]

[0147] Next, the application of the above formula in the present application is illustrated using an example:

[0148] In some embodiments, assuming there are four amplitude probability distribution values in a batch of sample data, A1, B1, C1, and D1, the similarity S1 between A1 and B1 can be calculated by a positive similarity function. Then, taking A1 as the center, the similarities S1, S2, and S3 between A1 and B1, C1, and D1, respectively, are calculated. Finally, S1 is divided by the sum of S1, S2, and S3 to normalize the similarity. The specific formula is as follows:

[0149]

[0150] Next, the technical solution of the present application is discussed in a general process using an example.

[0151] Please refer to Figure 10 , Figure 11 and Figure 12 , Figure 10 , Figure 11 and Figure 12The process diagram of the data recognition model from the pre-training stage, the fine-tuning stage to the decision stage is shown. In some embodiments, in the pre-training stage of the data recognition model, the data recognition model is mainly trained by the sample data, so that the data recognition model has feature extraction capability. Specifically, a large amount of unlabeled sample data (including amplitude entropy feature map and phase entropy feature map) can be input into the data recognition model, the feature extraction module of the data recognition model extracts features from the sample data, and maps the sample data to a high-dimensional space, in which the sample data is classified by a classifier. Further, the output of the classifier can be converted into amplitude probability distribution and phase probability distribution, which can be achieved by using a softmax function (normalized exponential function). Then, the parameters of the feature extraction module and the parameters of the classification module are adjusted to adjust the loss function, so that the feature vectors of the unlabeled sample data have certain tightness and discrimination in the feature space. Thus, the model training in the pre-training stage is completed, and the model has semi-supervised learning ability. It can be understood that in this process, the performance of the data recognition model can be optimized by using a back propagation algorithm, that is, the gradient of the loss function with respect to the model parameters is calculated, and the gradient descent optimization algorithm is used to update the parameters of the data recognition model, so that the data recognition model can gradually optimize and improve the accuracy of data prediction. Through multiple iterations of optimization, the data recognition model can gradually learn the feature extraction and classification ability suitable for the task, and use the feature information of the unlabeled sample data to improve the model performance.

[0152] In some embodiments, the loss function is composed of two parts, including a Twist loss function and a geometric loss function, wherein the Twist loss function is used for feature extraction of data, and the geometric loss function is used for obtaining the geometric structure between samples. It can be understood that the parameters of the feature extraction module can be fixed in the pre-training stage, or the parameters of the feature extraction module can be fixed in the fine-tuning stage.

[0153] It can be understood that after the data recognition model is pre-trained, the model can be applied to any occasion requiring identity recognition or action recognition, and is not limited by the specific environment. Only a small amount of labeled sample data needs to be trained, and the parameters in the classification module can be adjusted according to the training results.

[0154] In some embodiments, in the fine-tuning stage of the data recognition model, the input sample data is labeled sample data, specifically sample data of different scenes or different tasks, such as action scenes, which are labeled as actions such as falling, or identity recognition scenes, such as person detection scenes, which are labeled as the name of the person. It can be understood that the sample data at this time is a small amount of sample data, and since the data recognition model has been given strong feature extraction capability in the pre-training stage, it is not necessary to train the feature extraction capability of the data recognition model. In some embodiments, the feature extraction module can extract features from the sample data, and since the model has been given the feature extraction capability in the pre-training stage, the parameters of the feature extraction module can not be adjusted at this time. In some embodiments, after the sample data is extracted, the sample is classified by the classification module to output a set of predicted class labels. It can be understood that in order to evaluate and optimize the classification result, the classification result can be converted into a discriminant vector, specifically, the discriminant vector is a vector that encodes the classification result, which can be used to measure the accuracy and performance of the classification result.

[0155] Further, according to the discriminant vector and the label corresponding to the input sample data, the loss function is input. It can be understood that the loss function is used to measure the prediction error of the classifier, and by minimizing the loss function, the parameters of the classifier can be optimized to improve the accuracy and performance of the classification. It can be understood that after the data recognition model converges or reaches a certain stopping condition, the fine-tuning stage of the data recognition model is completed, at this time, it means that the data recognition model has been given the classification capability corresponding to the corresponding sample data, therefore, the parameters of the classification module are fixed.

[0156] In some embodiments, in the decision stage of the data recognition model, that is, after the data recognition model has been trained and has the ability to recognize data, real-time samples (i.e. real-time extracted data) are input to detect real data, the sample is feature extracted to obtain the amplitude entropy feature map and the phase entropy feature map, the feature extraction module is used to extract data features, and then the classification module is used to divide the categories of the features, and the discriminant vector is output to represent the relevance of the features. In some embodiments, the results of each discriminant vector can be fused using a fusion function, which can weight and sum the confidence or relevance of multiple discriminant vectors to obtain the final data recognition result.

[0157] It can be understood that the real-time sample input in the decision stage should be the same type of sample data as the sample data in the fine-tuning stage, for example, both are person recognition data.

[0158] Through implementation of the above overall process, only a small amount of samples needs to be input to train the data recognition model, so that the data recognition model has strong feature extraction capability and data classification capability, thereby not being limited to the availability of the environment and sample data, so that the model is adapted to various environments, and the efficiency and accuracy of data recognition of the data recognition model on target data are greatly improved.

[0159] Please refer to Figure 13 The embodiment of the application also provides a data processing system based on data driving, which can implement the data processing method based on data driving. The data processing system based on data driving comprises:

[0160] The first sample data acquisition module 1301 is configured to acquire first sample data with labels and channel state signals corresponding to the first sample data, and input the channel state signals to a pre-set data recognition model. The data recognition model is obtained by feature learning and training of a plurality of un-labeled second sample data. The number of the first sample data is less than the number of the second sample data. The data recognition model is provided with a feature extraction module and a classification module.

[0161] The similarity calculation module 1302 is configured to calculate the amplitude value and the phase value of the channel state signals by using the data recognition model, and calculate the similarity of the amplitude value and the phase value.

[0162] The first loss value acquisition module 1303 is configured to adjust the parameters of the classification module according to the similarity of the amplitude value and the phase value, and obtain a first loss value.

[0163] The adjacent phase similarity acquisition module 1304 is configured to acquire adjacent phase values of the phase value, and calculate the adjacent phase similarity of the phase value and the adjacent phase values.

[0164] The adjacent amplitude similarity acquisition module 1305 is configured to acquire adjacent amplitude values of the amplitude value, and calculate the adjacent amplitude similarity of the amplitude value and the adjacent amplitude values.

[0165] The second loss value acquisition module 1306 is configured to adjust the parameters of the classification module according to the adjacent phase similarity and the adjacent amplitude similarity, and obtain a second loss value.

[0166] The target data recognition model acquisition module 1307 is configured to fix the parameters of the feature extraction module, and fix the parameters of the classification module according to the adjusted first loss value and the second loss value, and obtain a target data recognition model.

[0167] The data recognition result acquisition module 1308 is configured to acquire target data, input the target data into the target data recognition model, and obtain a data recognition result.

[0168] In some embodiments, the classification module can predict the class to which the sample belongs according to the feature vector of the sample. In some embodiments, the classification module is a classifier, where the classifier can be a linear classifier and a nonlinear classifier, the linear classifier can be a logistic regression, a support vector machine, etc., and the nonlinear classifier can be a decision tree, a neural network, etc.

[0169] It can be understood that the feature extraction module refers to an algorithm or model used to extract useful features in input data in the field of machine learning and computer vision. In some embodiments, the feature extraction module can be a feature extractor, which can extract features in different ways, for example, an image-based feature extractor can use various filters, edge detectors, texture descriptors, etc. to capture structures and patterns in images; while a text-based feature extractor can use bag-of-words model, TF-IDF, etc. to extract keywords and phrases in text.

[0170] It can be understood that the training or identification of the sample using the data recognition model can be divided into three stages. Specifically, the first stage is the pre-training stage, corresponding to the second sample data, at this time the data recognition model is trained by a large amount of unlabeled sample data, and the parameters of the first loss function and the second loss function are adjusted in this process, so that the data recognition model can also have strong feature extraction ability and classification ability without sample labels; the second stage is the fine-tuning stage, corresponding to the first sample data, at this time, only a small amount of sample data with labels, such as 1 to 3 sample data, is needed, so that the data recognition model can fix the parameters of the feature extraction module and fine-tune the parameters of the classification module, and the data recognition model can quickly extract features and classify labels according to the sample data labels. That is to say, the data recognition model can perform action recognition or other recognition such as identity recognition in any scene through the training of a small amount of samples; the third stage is the judgment stage, corresponding to the target data, at this time, the data recognition model has completed the training of the target recognition of a specific scene, such as human action training, and can directly identify the input target data, that is, can be directly put into use.

[0171] In some embodiments, before the first sample data with labels is input into the data recognition model, the data recognition model has been trained by the second sample data to obtain strong feature extraction capability and sample classification capability. Specifically, the data recognition model continuously optimizes parameters in the training process according to a twist loss function, and obtains strong feature extraction capability for sample data. The similarity of the amplitude probability distribution and the phase probability distribution of the sample data is calculated by a geometric consistency loss function, thereby simulating the relationship between the sample data and the sample labels, so that the data recognition model has the ability to recognize labels according to sample data.

[0172] In some embodiments, the number of the first sample data is less than the number of the second sample data, because the second sample data is used to train the initial data recognition model, and the initial data recognition model needs to be trained by a large amount of sample data to have strong feature recognition capability and sample classification capability. It can be understood that the second sample data is unlabeled sample data without labels, because the role of the second sample data is to enable the data recognition model to have the ability to be applicable to any behavior recognition scene or detection scene through training, so that when the data recognition model is trained by the second sample data, the parameters of the feature extraction module will be fixed, but the parameters of the classification module will not be fixed.

[0173] It can be understood that the first sample data is labeled sample data, and further, the first sample data is any data that needs to be classified and recognized. For example, the first sample data can be action data, and the corresponding label is falling down, waving hand, etc. The first sample data can also be related to personnel detection data, such as portrait data, and the label is name, etc. It can be understood that since the data recognition model has strong feature extraction capability and sample classification capability after being trained by the second sample data, that is, it has strong feature extraction capability and sample classification capability through semi-supervised learning, therefore, the first sample data can be 1 to 3, for example, 2, so that the data recognition model can directly recognize the corresponding data according to the corresponding sample data and labels. It can be understood that in order to also increase the number of first sample data according to actual needs, the present application does not make specific limitations thereto.

[0174] In some embodiments, the amplitude value and the phase value of the channel state signal can be calculated by the data recognition model, and the similarity of the amplitude value and the phase value can be calculated. In some embodiments, the similarity of the amplitude value and the phase value can be calculated by Euclidean distance, cosine similarity, etc., and the parameters of the classification module can be adjusted according to the similarity, so that the classification module increases the extraction ability of the characteristics of the same sample data. For example, when the similarity of the amplitude value and the phase value of the sample data reaches a preset threshold, the parameters of the classification module can be adjusted, that is, by adjusting the parameters of the classification module, the data recognition model can quickly extract the same characteristics when identifying the same sample data, so as to improve the classification ability of the sample data and improve the efficiency of subsequent training of the data classification model.

[0175] In some embodiments, the adjacent phase similarity of the phase value and the adjacent phase value, and the adjacent amplitude similarity of the amplitude value and the adjacent amplitude value can be calculated, and the parameters of the classification module can be adjusted according to the adjacent phase similarity and the adjacent amplitude similarity. It can be understood that if the adjacent phase similarity is high, the adjacent amplitude similarity should also be the same as or similar to the adjacent phase similarity, and at this time, the two sample data can be classified into the same type of sample data by adjusting the parameters of the classification module, so as to realize efficient classification of the sample data. It can be understood that if the adjacent phase similarity is low, the adjacent amplitude similarity should also be the same as or similar to the adjacent phase similarity, and at this time, the two sample data can be classified into different types of sample data by adjusting the parameters of the classification module, so as to realize efficient classification of the sample data.

[0176] It can be understood that in order to enable the data recognition model to extract common characteristics of the amplitude entropy and the phase entropy, a first loss function, i.e., a twist loss function, can be set to adjust the similarity of the amplitude entropy and the phase entropy to obtain a first loss value, so that the data recognition model can extract common characteristics of the amplitude entropy and the phase entropy. In some embodiments, the similarity can be adjusted by using a similarity measurement method, such as cosine similarity, Euclidean distance, etc.

[0177] In some embodiments, the first loss function (Twist loss function) corresponding to the first loss value includes a consistency loss value, a sharpness loss value, and a diversity loss value. It can be understood that in the process of training the data recognition model, in order to improve the extraction ability of the data recognition model to the characteristics, the consistency loss value and the sharpness loss value should be as low as possible, and the diversity loss value should be as high as possible.

[0178] In some embodiments, a neighboring phase value of the phase value can be obtained, and a neighboring similarity between the phase value and the neighboring phase value can be calculated. The neighboring similarity can be used as an adjustment target to adjust the amplitude value and the neighboring amplitude value under the first sample data, so as to obtain a second loss value. Alternatively, a neighboring amplitude value of the amplitude value can be obtained, and a neighboring similarity between the amplitude value and the neighboring amplitude value can be calculated. The neighboring similarity can be used as an adjustment target to adjust the phase value and the neighboring phase value under the first sample data, so as to obtain a second loss value.

[0179] In some embodiments, a neighboring similarity between the amplitude probability distribution value and the neighboring amplitude probability distribution value can also be calculated, and the neighboring similarity can be used as an adjustment target to adjust the similarity between the phase probability distribution value and the neighboring phase probability distribution value. Alternatively, a neighboring similarity between the phase probability distribution value and the neighboring phase probability distribution value can be calculated, and the neighboring similarity can be used as an adjustment target to adjust the similarity between the amplitude probability distribution value and the neighboring amplitude probability distribution value. It can be understood that the adjustment based on the similarity between the amplitude values or the phase values of the sample data can make the geometric structure between the amplitude values and the phase values consistent as much as possible, so as to facilitate the classification of the sample data by the data recognition model. Therefore, during the training of the data recognition model by the second sample data, the adjustment between the amplitude values and the phase values can not necessarily require that the sample data must be accompanied by sample labels, so as to limit the data recognition model to a certain scene, or to make the data recognition model spend a large amount of data for training or design a new model each time the new environment is identified. It is only necessary to make the data recognition model have classification ability through similarity adjustment at the beginning.

[0180] It can be understood that, when the first sample data is used to train the data recognition model, the feature extraction capability of the data recognition model for the data is already good enough, so the parameters of the first loss function can not be adjusted and can be directly fixed. Alternatively, the parameters of the first loss function can be fine-tuned and fixed, and the first loss value is recalculated. The embodiments of the present application do not make specific limitations on this.

[0181] It can be understood that, after the first sample data is used to train the data recognition model, the data recognition model already has the recognition capability for the scene or environment corresponding to the first sample data, so the target data can be directly obtained and input into the trained data recognition model, so as to obtain a data recognition result, such as judging that the corresponding person's name is Zhang San, or judging that the corresponding human body action is falling down.

[0182] The data processing method, system, electronic device and medium based on data driving provided in the application can learn sample data features through a data recognition model, fix parameters of a feature extraction module of the model, and train the model based on labeled sample data, so that the data recognition model can adjust parameters of a classification module, and thus has the ability to classify input sample data. Therefore, only a small amount of sample data needs to be input to train the data recognition model, so that the data recognition model has strong feature extraction and data classification capabilities, thereby not being limited to the environment and availability of sample data, making the model suitable for various environments, and greatly improving the efficiency and accuracy of the data recognition model in recognizing target data.

[0183] The specific implementation of the data processing system based on data driving is basically the same as the specific embodiments of the data processing method based on data driving described above, and will not be repeated here. The data processing system based on data driving can also be provided with other functional modules to implement the data processing method based on data driving in the above embodiments, as long as the requirements of the embodiments of the application are met.

[0184] The embodiments of the application also provide an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor implements the data processing method based on data driving when executing the computer program. The electronic device can be any intelligent terminal, such as a tablet computer or a vehicle-mounted computer.

[0185] Please refer to Figure 14 , Figure 14 The hardware structure of the electronic device of another embodiment is illustrated, which includes:

[0186] The processor 1401 can be implemented in the form of a general-purpose CPU (Central Processing Unit), a microprocessor, an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits, and is used to execute related programs to implement the technical solutions provided by the embodiments of the application.

[0187] The memory 1402 can be implemented in the form of a ROM (ReadOnly Memory), a static storage device, a dynamic storage device, or a RAM (Random Access Memory). The memory 1402 can store an operating system and other application programs. When the technical solutions provided by the embodiments of the application are implemented by software or firmware, the related program codes are stored in the memory 1402 and called and executed by the processor 1401 to implement the data processing method based on data driving of the embodiments of the application.

[0188] The input / output interface 1403 is configured to realize information input and output.

[0189] The communication interface 1404 is configured to realize communication interaction between the device and other devices, and the communication can be realized through wired mode (for example, USB, network cable and the like) or wireless mode (for example, mobile network, WIFI, Bluetooth and the like).

[0190] The bus 1405 is configured to transmit information between various components (for example, the processor 1401, the memory 1402, the input / output interface 1403 and the communication interface 1404) of the device.

[0191] The processor 1401, the memory 1402, the input / output interface 1403 and the communication interface 1404 are connected to each other through the bus 1405 to realize communication connection between the device.

[0192] The computer readable medium provided by the embodiment of the present application stores a computer program, and the computer program is executed by the processor to realize the data processing method based on data driving.

[0193] The memory is a non-transient computer readable storage medium, and can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory can include a high-speed random access memory, and can also include a non-transient memory, for example, at least one magnetic disk storage device, a flash memory device or other non-transient solid-state memory device. In some embodiments, the memory can optionally include a memory remotely arranged relative to the processor, and the remote memory can be connected to the processor through a network. Examples of the network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network and a combination thereof.

[0194] The embodiments described in the embodiments of the present application are used to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art can know that, with the evolution of technology and the appearance of new application scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0195] Those skilled in the art can understand that the technical solutions shown in the figure do not constitute a limitation on the embodiments of the present application, and can include more or fewer steps than the figure, or combine certain steps or different steps.

[0196] The apparatus embodiments described above are merely exemplary, and units described as separate units can or can not be physically separate, i.e., can be located in one place, or can be distributed over multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment.

[0197] Those skilled in the art can understand that all or some of the steps in the method disclosed above, the functional modules / units in the system and the device can be implemented as software, firmware, hardware and appropriate combinations thereof.

[0198] The terms "first", "second", "third", "fourth" and the like in the description of the application and in the claims of the foregoing drawings, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of the terms so

[0199] It should be understood that in this application, "at least one" and "several" refer to one or more, and "multiple" refers to two or more. "And / or" is used to describe the relationship between the associated objects, which means that there can be three relationships, for example, "A and / or B" can represent three cases: only A, only B, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally represents that the associated objects before and after are in an "or" relationship. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c, can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0200] In several embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented in other manners. For example, the system embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0201] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0202] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can be physically present separately, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0203] If the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable medium. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that contributes to the technical solutions or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a medium and includes multiple instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of the embodiments of the present application. The aforementioned medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program storage media.

[0204] The preferred embodiments of the embodiments of the present application are described above with reference to the accompanying drawings, but this does not limit the scope of the rights of the embodiments of the present application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall be within the scope of the rights of the embodiments of the present application.

Claims

1. A data processing method based on data driving, characterized in that, The method comprises: acquiring labeled first sample data and channel state signals corresponding to the first sample data, and inputting the channel state signals into a pre-set data recognition model; wherein the channel state signals are WiFi signals; the data recognition model is obtained by feature learning and training of a plurality of unlabeled second sample data; the number of the first sample data is less than the number of the second sample data; the data recognition model is provided with a feature extraction module and a classification module; calculating the amplitude value and the phase value of the channel state signals by the data recognition model, and calculating the similarity of the amplitude value and the phase value; adjusting the parameters of the classification module according to the similarity of the amplitude value and the phase value to obtain a first loss value, the first loss value being a distortion loss value; acquiring adjacent phase values of the phase value, and calculating the adjacent phase similarity of the phase value and the adjacent phase value; acquiring adjacent amplitude values of the amplitude value, and calculating the adjacent amplitude similarity of the amplitude value and the adjacent amplitude value; adjusting the parameters of the classification module according to the adjacent phase similarity and the adjacent amplitude similarity to obtain a second loss value, the second loss value being a geometric consistency loss value; fixing the parameters of the feature extraction module, and fixing the parameters of the classification module according to the adjusted first loss value and the second loss value to obtain a target data recognition model; acquiring target data, inputting the target data into the target data recognition model, and obtaining a data recognition result.

2. The data processing method based on data driving according to claim 1, characterized in that, After acquiring the labeled first sample data and the channel state signals corresponding to the first sample data, the method further comprises: establishing a local Gaussian noise model, and performing noise analysis on the channel state signals by the local Gaussian noise model to obtain a noise analysis result; identifying noise signals according to the noise analysis result; mapping the noise signals to a high-dimensional feature space, and fitting the noise signals using a linear regression model to obtain a fitting result; removing the noise signals in the channel state signals according to the fitting result.

3. The data-driven based data processing method of claim 1, wherein, The data recognition model is obtained by the following steps: acquiring second sample data, and inputting the second sample data into the data recognition model; performing feature detection on the second sample data to obtain probability distribution values of the second sample data; analyzing channel state signals of the second sample data to obtain amplitude values and phase values corresponding to each of the second sample data; calculating the probability distribution of the amplitude values and the phase values in the probability distribution values to obtain amplitude probability distribution values and phase probability distribution values; calculating the amplitude entropy corresponding to the amplitude probability distribution values and the phase entropy corresponding to the phase probability distribution values according to the amplitude probability distribution values and the phase probability distribution values; calculating first loss values and second loss values according to the amplitude entropy and the phase entropy; calculating a total loss value according to the first loss values and the second loss values, and fixing the parameters of the data recognition model according to the total loss value to obtain the trained data recognition model.

4. The data processing method based on data driving according to claim 3, characterized in that, The data recognition model further comprises a time attention module; the obtaining of the second sample data and the inputting of the second sample data into the data recognition model comprise: extracting an amplitude entropy image and a phase entropy image from each channel state signal of the second sample data; extracting a subcarrier of the channel state signal and a feature of the channel state signal in a time dimension according to the amplitude entropy image and the phase entropy image, to obtain an attention feature map of the subcarrier in the time dimension; multiplying the amplitude entropy image and the phase entropy image with the attention feature map respectively, to obtain an amplitude entropy feature map and a phase entropy feature map; inputting the amplitude entropy feature map and the phase entropy feature map into the data recognition model.

5. The data processing method based on data driving according to claim 3, characterized in that, The first loss value is a distortion loss value, the distortion loss value comprises a consistency loss value, a sharpness loss value and a diversity loss value; the distortion loss value is calculated by the following steps: calculating an amplitude distribution divergence value and a phase distribution divergence value according to the amplitude probability distribution value and the phase probability distribution value; accumulating the amplitude distribution divergence value and the phase distribution divergence value corresponding to each sample data to obtain a divergence sum, and dividing the divergence sum by the number of samples to obtain an average divergence value; dividing the average divergence value by 2 to obtain a consistency loss value; accumulating the amplitude entropy of each sample data and dividing the amplitude entropy by the number of samples to obtain an amplitude entropy average value; accumulating the phase entropy of each sample data and dividing the phase entropy by the number of samples to obtain a phase entropy average value; adding the amplitude entropy average value and the phase entropy average value and dividing the sum by 2 to obtain a first entropy average value of each sample data; multiplying the first entropy average value by a first weight hyperparameter to obtain a sharpness loss value; adding the amplitude probability distribution values corresponding to each sample data to obtain an amplitude probability distribution total value; adding the phase probability distribution values corresponding to each sample data to obtain a phase probability distribution total value; dividing the amplitude probability distribution total value by the number of samples to obtain an amplitude probability distribution average value; dividing the phase probability distribution total value by the number of samples to obtain a phase probability distribution average value; calculating a first amplitude entropy of the amplitude probability distribution average value and a first phase entropy of the phase probability distribution average value; accumulating the first amplitude entropy and the first phase entropy and dividing the sum by 2 to obtain a second entropy average value corresponding to each sample data; multiplying the second entropy average value by a second weight hyperparameter to obtain a diversity loss value; adding the consistency loss value, the sharpness loss value and the diversity loss value to obtain a distortion loss value.

6. The data processing method based on data driving according to claim 3, characterized in that, The second loss value is a geometric consistency loss value, and the geometric consistency loss value is calculated by the following steps: calculating a similarity of the amplitude probability distribution value and the phase probability distribution value; constructing a similarity matrix according to the similarity, and mapping the similarity matrix to a low-dimensional space to obtain a geometric embedding result; wherein the geometric embedding result is used to represent the position of each amplitude probability distribution value or phase probability distribution value in a geometric space. construct a second loss function according to the geometric embedding result and the similarity; calculate a second loss value according to the second loss function.

7. The data processing method based on data driving according to claim 6, characterized in that, The similarity of the amplitude probability distribution value and the phase probability distribution value comprises: obtaining a phase value of the second sample data and adjacent phase values of the phase value, an amplitude value and adjacent amplitude values of the amplitude value; wherein the adjacent phase values and the adjacent amplitude values are multiple; calculating a first similarity of the phase value and one of the adjacent phase values according to a positive value similarity function, and calculating a second similarity of the amplitude value and the multiple adjacent amplitude values; obtaining a normalized similarity by dividing the first similarity by the cumulative value of the second similarity.

8. A data processing system based on data driving, characterized by, The system comprises: a first sample data acquisition module configured to acquire labeled first sample data and a channel state signal corresponding to the first sample data, and input the channel state signal into a pre-set data recognition model; wherein the channel state signal is a WiFi signal; the data recognition model is obtained by feature learning and training of multiple unlabeled second sample data; the number of the first sample data is less than the number of the second sample data; the data recognition model is provided with a feature extraction module and a classification module; a similarity calculation module configured to calculate an amplitude value and a phase value of the channel state signal by the data recognition model, and calculate a similarity of the amplitude value and the phase value; a first loss value acquisition module configured to adjust parameters of the classification module according to the similarity of the amplitude value and the phase value, to obtain a first loss value, the first loss value being a distortion loss value; an adjacent phase similarity acquisition module configured to obtain adjacent phase values of the phase value, and calculate an adjacent phase similarity of the phase value and the adjacent phase values; an adjacent amplitude similarity acquisition module configured to obtain adjacent amplitude values of the amplitude value, and calculate an adjacent amplitude similarity of the amplitude value and the adjacent amplitude values; a second loss value acquisition module configured to adjust the parameters of the classification module according to the adjacent phase similarity and the adjacent amplitude similarity, to obtain a second loss value, the second loss value being a geometric consistency loss value; a target data recognition model acquisition module configured to fix the parameters of the feature extraction module, and fix the parameters of the classification module according to the adjusted first loss value and the second loss value, to obtain a target data recognition model; a data recognition result acquisition module configured to acquire target data, input the target data into the target data recognition model, and obtain a data recognition result.

9. An electronic device, comprising: The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the data-driven data processing method of any one of claims 1 to 7 when executing the computer program.

10. A computer readable medium, said medium storing a computer program, characterized in that, The computer program is executed by the processor to implement the data-driven data processing method of any one of claims 1 to 7.

Images