Data processing method, and readable storage medium and electronic device

Abstract

The present application relates to the field of communications, and particularly relates to a data processing method, and a readable storage medium and an electronic device. In the method, a first node device in a distributed system receives noise statistical information sent by a decision-making device in the distributed system, receives feature information sent by node devices other than the first node device in the distributed system, and trains a neural network model, so as to obtain a power adaptation sub-model for executing a classification task. It can be understood that since the noise statistical information comprises statistical characteristics of noise in a channel between each node device and the decision-making device, and the feature information comprises feature data of each sample in a sample set of each node device and a classification label of each sample, a power adaptation sub-model, which is obtained after training and corresponds to each node device, can perform adaptation to noise in a channel corresponding to each node device, thereby improving the accuracy of a classification result of a corresponding classification task executed in the distributed system.

Description

Data processing method, readable storage medium and electronic device

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on December 11, 2023, with application number 202311699822.9 and application name “Data processing method, readable storage medium and electronic device”. The entire contents of the above application are incorporated by reference into this application. Technical Field

[0002] The present application relates to the field of communications, and in particular to a data processing method, a readable storage medium, and an electronic device. Background Art

[0003] In some scenarios, a transmitter (the electronic device sending the signal) can transmit a signal related to a classification task (hereinafter referred to as the "task signal") to a receiver (the electronic device receiving the signal) through a channel, thereby achieving information transfer. After receiving the task signal, the receiver can perform the classification task based on the received signal.

[0004] For example, let's assume the transmitting end is a camera and the receiving end is a base station. The classification task is to determine the gender (i.e., male or female) of a person in an image captured by the camera. The camera can capture the person in the surrounding environment, obtain image data, and transmit the signal corresponding to the image data to the base station through a channel. The base station will restore the image captured by the camera based on the received signal and determine whether the person in the camera's surrounding environment is male or female based on the restored image.

[0005] However, the signal corresponding to the task data (eg, image data) is usually interfered with by noise signals during transmission from the transmitting end to the receiving end, and is significantly affected by the noise interference, resulting in an inadequate classification result determined by the receiving end.

[0006] Summary of the Invention

[0007] In order to solve the above problems, embodiments of the present application provide a data processing method, an electronic device, and a readable storage medium for improving the accuracy of classification results of classification tasks.

[0008] In a first aspect, an embodiment of the present application provides a data processing method, comprising: a first node device in a distributed system receives first noise statistical information sent by a decision-making device in the distributed system, and receives first feature information sent respectively by node devices other than the first node device in the distributed system; wherein the first noise statistical information includes statistical characteristics of the noise of the channel between the node device and the decision-making device, and the first feature information includes feature data of each sample in a first sample set of the node device, and a classification label of each sample; and the first noise statistical information and the first feature information are used to determine a first power adaptation sub-model used by each node device to perform a first task.

[0009] It is understandable that the first task can be a classification task and is not limited to any classification scenario. Each node device in the distributed system can perform feature extraction on its own first sample set, thereby obtaining feature data for each sample. Since each sample has a corresponding classification label, the feature data also has a corresponding classification label. When information can be transmitted between the node devices in the distributed system, any node device can receive first feature information sent by other node devices except itself, so that each node device can obtain feature data for samples from all node devices. Since the decision device can observe the statistical characteristics of the noise in the channel corresponding to each node device, the decision device can transmit the observed statistical characteristics of the noise in the channel corresponding to each node device to each node device. Since the node devices have strong computing capabilities, any node device can determine the first power adaptation sub-model required to perform the first task based on the received feature data of samples from all node devices and the statistical characteristics of the noise in the channel corresponding to each node device. The power adaptation sub-model of each node device can be used to adjust the data sent by each node device to the decision device before transmission. For example, the first node device is node 0 below, and the first power adaptation sub-model of each node device can be the power adaptation sub-model corresponding to each node determined by node 0. The characteristic data of each sample in the first sample set of the node device, and the classification label of each sample, can be the label characteristic data below. It can be understood that, since the characteristic data of the first sample set of all node devices and the statistical characteristics of the noise of the channel corresponding to each node device are taken into account when determining the first power adaptation sub-model, the data is adjusted before being sent using the determined first power adaptation sub-model, and the adjusted data can be adapted to the noise of the corresponding channel, and has a stronger ability to protect itself from noise interference. In addition, the correlation of the data of each node device is also taken into account, thereby improving the accuracy of the classification result.

[0010] In a possible implementation of the first aspect above, the first power adapter sub-model used by each node device to perform the first task is determined in the following manner: the first node device trains the first neural network model based on the first noise statistical information and the first feature information corresponding to each node device to obtain the trained first neural network model, wherein the trained first neural network model includes the first power adapter sub-model corresponding to each node device.

[0011] It is understandable that the neural network model training performed on the first node device can be used to train the neural network model for the node 0 described below. Since the neural network model has a strong fitting capability, and the first power adapter model is obtained through the neural network model training, the scientific nature of the first power adapter model obtained for each node device is ensured.

[0012] In a possible implementation of the first aspect above, the statistical characteristics include one or more of the following: mean vector, covariance, arithmetic mean, geometric mean, harmonic mean, weighted mean, square mean, median, median number, mode, mean absolute deviation, and variance.

[0013] In a possible implementation of the first aspect above, the first node device also sends first adaptation processing information to the decision device, wherein the first adaptation processing information includes multiple groups of first input-output data pairs, and each group of first input-output data pairs corresponds to a first power adaptation sub-model of a node device.

[0014] It is understood that the first power adaptation sub-model of each node device can be the power adaptation sub-model corresponding to node 0 and the power adaptation sub-model corresponding to node 1 described below. The input and output data pairs of each group can be the key point information described below. It is understood that sending the input and output data pairs corresponding to the first power adaptation sub-model of each node device to the decision device can reduce the amount of data transmission, thereby saving transmission resources.

[0015] In a possible implementation of the first aspect, the method further includes: the decision-making device obtaining a first power adaptation function corresponding to the first power adaptation sub-model of each node device based on the received first adaptation processing information.

[0016] It can be understood that the decision-making device can restore the first power adaptation sub-model required for decision-making based on the input and output data pairs corresponding to the first power adaptation sub-model of each node device in the first adaptation information received, that is, the first power adaptation function corresponding to the first power adaptation sub-model obtained is the first power adaptation sub-model restored based on the input and output data pairs, thereby facilitating the subsequent execution of the first task.

[0017] In a possible implementation of the first aspect above, the decision-making device obtains a first power adaptation function corresponding to the first power adaptation sub-model of each node device based on the received first adaptation processing information, including: determining the first power adaptation function corresponding to the first power adaptation sub-model of the first node device in the following manner: the decision-making device inputs the first input and output data pair corresponding to the first power adaptation sub-model of the first node device into the first parameter-containing function model to obtain the first solution value of each parameter in the first parameter-containing function model; the decision-making device replaces the parameters in the first parameter-containing function model with the corresponding first solution value to obtain the first power adaptation function.

[0018] It is understandable that by selecting a suitable parameter-containing function model and solving the parameter values ​​of the suitable parameter-containing model through input and output data pairs, the required first power adaptation function corresponding to the first power adaptation sub-model can be quickly restored.

[0019] In a possible implementation of the first aspect above, after the first node device receives the first feature information respectively sent by the node devices other than the first node device in the distributed system, it also includes: the first node device sends first label feature statistical information to the decision device, wherein the first label feature statistical information includes statistical characteristics corresponding to each classification label in a plurality of classification labels, wherein the statistical characteristics corresponding to the first classification label in the plurality of classification labels are the statistical characteristics of the feature data whose classification label is the first classification label in all feature data corresponding to the sample sets of all node devices in the distributed system.

[0020] As can be understood, since the first node device can obtain all feature data corresponding to the first sample set of all node devices, statistics can be performed based on the classification labels to obtain statistical characteristics of feature data with the same classification label within all feature data corresponding to the first sample set of all node devices. The obtained statistical characteristics can then be sent to the decision-making device, which can then simulate the data required for the decision-making process based on the statistical characteristics in the first label feature statistical information. As can be understood, since the first node device only transmits the first label feature statistical information, rather than all feature data, the amount of data transmitted is reduced.

[0021] In a possible implementation of the first aspect above, it also includes: each node device obtains first data to be classified corresponding to the first classification object in the first task; each node device processes the first classification feature data of the first data to be classified into second classification feature data based on its own first power adaptation sub-model, and sends it to the decision device.

[0022] It can be understood that the first data to be classified can be the task data described below; the first classification feature data of the first data to be classified is the feature data obtained after feature extraction of the task data; and the second classification feature data is the adjusted feature data obtained after the feature data obtained after feature extraction is processed using the power adaptation sub-model described below. When each node device adjusts the first classification feature data before transmission according to its respective first power adaptation sub-model to obtain the second classification feature data, since the first power adaptation sub-model is obtained by considering the statistical characteristics of the channel noise, the noise resistance of the second classification feature data during transmission through the channel is improved, thereby improving the accuracy of the data received by the decision-making device.

[0023] In a possible implementation of the first aspect above, the decision-making device also obtains a classification result for the first classification object based on a first power adaptation function corresponding to a first power adaptation sub-model of each node device, third classification feature data corresponding to the second classification feature data sent by each node device, and statistical characteristics corresponding to each classification label in multiple classification labels.

[0024] It can be understood that the third classification feature data is the receiving end data received by the decision center below.

[0025] In a possible implementation of the first aspect above, the decision-making device obtains a classification result for the first classification object based on the first power adaptation function corresponding to the first power adaptation sub-model of each node device, the third classification feature data corresponding to the second classification feature data sent by each node device, and the statistical characteristics corresponding to each classification label in the multiple classification labels, including: the decision-making device determines the likelihood function values ​​corresponding to the first classification object and each classification label respectively based on the first power adaptation function corresponding to each node device, the third classification feature data corresponding to the second classification feature data sent by each node device, and the statistical characteristics corresponding to each classification label in the multiple classification labels; and takes the classification category of the classification label corresponding to the largest likelihood function value among the likelihood function values ​​as the classification result of the first classification object.

[0026] It can be understood that the decision-making device makes a decision based on the received terminal data corresponding to multiple node devices, that is, calculates the likelihood function, so that the classification result takes into account the correlation between multiple node data, thereby improving the accuracy of the decision result.

[0027] In a possible implementation of the first aspect, the first node device further sends second adaptation processing information to the decision device, wherein the second adaptation processing information includes a first power adaptation sub-model for each node device to perform the first task.

[0028] It is understandable that when the first node device sends the first power adaptation sub-model for executing the first task to the decision device, although the transmission volume is large, the decision device obtains a higher accuracy rate based on the first power adaptation sub-model corresponding to each node device.

[0029] In second aspect, an embodiment of the present application provides a data processing method, including: a decision-making device in a distributed system receives second characteristic information sent by each node device in the distributed system; wherein the second characteristic information includes characteristic data of each sample in a second sample set of the node device, and a classification label of each sample; and the second characteristic information is used to determine a second power adaptation sub-model used by each node device to perform a second task.

[0030] It can be understood that the second task can be a classification task and is not limited to any classification scenario. Each node device in the distributed system can perform feature extraction on its own second sample set to obtain feature data for each sample. Since each sample has a corresponding classification label, the feature data will also have a corresponding classification label, for example, the label feature data below. When it is necessary to perform model training on the node devices in the distributed system, each node device can send the second feature information to the decision device. Since the second information includes the feature data of each sample in the second sample set of the corresponding node device and the classification label of each sample, it is convenient for the subsequent decision device to perform model training and obtain the second power adapter model used by each node device to perform subsequent tasks. For example, each node device is node 0 and node 1 below. Node 0 and node 1 respectively send the second feature information to the decision center 00, so that the decision center 00 can determine the power adapter model corresponding to the node.

[0031] In a possible implementation of the second aspect above, the second power adapter sub-model used by each node device to perform the second task is determined in the following manner: the decision-making device trains the second neural network model based on the statistical characteristics of the noise of the channel between each node device and the second feature information corresponding to each node device to obtain a trained second neural network model, wherein the trained second neural network model includes the second power adapter sub-model corresponding to each node device.

[0032] It can be understood that since the decision-making device observes the statistical characteristics of the noise of the channel corresponding to each node device and the corresponding second feature information of each node device to train the neural network model, when determining the second power adaptation sub-model, the feature data of the second sample set of all node devices and the statistical characteristics of the noise of the channel corresponding to each node device are fully considered, so that after the determined second power adaptation sub-model adjusts the data before sending, the adjusted data can be adapted to the noise of the corresponding channel, and the ability to protect itself from noise interference is stronger. At the same time, the correlation of the data of each node device is also considered, thereby improving the accuracy of the classification results.

[0033] In a possible implementation of the second aspect above, the statistical characteristics include one or more of the following: mean vector, covariance, arithmetic mean, geometric mean, harmonic mean, weighted mean, square mean, median, median number, mode, mean absolute deviation, and variance.

[0034] In a possible implementation of the second aspect above, the decision-making device also sends third adaptation processing information to each node device respectively, wherein the third adaptation processing information includes a group of second input-output data pairs, and each group of second input-output data pairs corresponds to the second power adaptation sub-model of the sending node device.

[0035] In a possible implementation of the second aspect, the method further includes: each node device obtaining a second power adaptation function corresponding to the second power adaptation sub-model of the node device based on the received third adaptation processing information.

[0036] In a possible implementation of the second aspect above, each node device obtains a second power adaptation function corresponding to the second power adaptation sub-model of the node device based on the received third adaptation processing information, including: the first node device obtains the second power adaptation function corresponding to the second power adaptation sub-model of the first node device based on the received third adaptation processing information: the first node device inputs the second input and output data pair into the second parameter-containing function model to obtain the second solution value of each parameter in the second parameter-containing function model; the first node device replaces the parameters in the second parameter-containing function model with the corresponding second solution value to obtain the second power adaptation function.

[0037] In a possible implementation of the second aspect above, it also includes: each node device obtains second data to be classified corresponding to the second classification object of the second task; each node device processes the fourth classification feature data of the second data to be classified into fifth classification feature data based on its own second power adaptation function, and sends it to the decision device.

[0038] In a possible implementation of the second aspect above, the decision-making device also obtains a classification result for the second classification object based on the second power adaptation sub-model corresponding to each node device, the sixth classification feature data corresponding to the fifth classification feature data sent by each node device, and the feature data of each sample in the second sample set of all node devices.

[0039] In a possible implementation of the second aspect above, the decision-making device obtains a classification result for the second classification object based on the second power adaptation sub-model corresponding to each node device, the sixth classification feature data corresponding to the fifth classification feature data sent by each node device, and the feature data of each sample in the second sample set of all node devices, including: the decision-making device determines the likelihood function values ​​corresponding to the second classification object and each classification label based on the second power adaptation sub-model corresponding to each node device, the sixth classification feature data corresponding to the fifth classification feature data sent by each node device, and the feature data of each sample in the second sample set of all node devices; and takes the classification category of the classification label corresponding to the largest likelihood function value among the likelihood function values ​​as the classification result of the second classification object.

[0040] In a possible implementation of the second aspect, the decision-making device further sends fourth adaptation processing information to each node device respectively, wherein the fourth adaptation processing information includes a second power adaptation sub-model corresponding to the node device.

[0041] In a third aspect, an embodiment of the present application provides an electronic device comprising: one or more processors; one or more memories; one or more memories storing one or more instructions, which, when one or more instructions are executed by one or more processors, enables the electronic device to execute the above-mentioned first aspect and any possible implementation of the data transmission method provided by the first aspect, or the above-mentioned second aspect and any possible implementation of the data transmission method provided by the second aspect.

[0042] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having instructions stored thereon. When the instructions are executed on a computer, the computer executes the data transmission method provided by the first aspect and any possible implementation of the first aspect, or the data transmission method provided by the second aspect and any possible implementation of the second aspect.

[0043] In a fifth aspect, an embodiment of the present application provides a computer program product comprising instructions, which, when the computer program product is run on a computer, enables the computer to execute the data transmission method provided by the above-mentioned first aspect and any possible implementation of the first aspect, or the data transmission method provided by the above-mentioned second aspect and any possible implementation of the second aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] FIG1A is a schematic diagram showing a scenario of transmission based on the same task in a distributed system according to some embodiments of the present application;

[0045] FIG1B shows a schematic diagram of an image T1 according to some embodiments of the present application;

[0046] FIG1C shows a schematic diagram of an image T2 according to some embodiments of the present application;

[0047] FIG1D shows a schematic diagram of an image T3 according to some embodiments of the present application;

[0048] FIG1E shows a schematic diagram of an image T4 according to some embodiments of the present application;

[0049] FIG2A shows a schematic diagram of a process of linear filtering according to some embodiments of the present application;

[0050] FIG2B shows a schematic diagram of a process of BPSK processing according to some embodiments of the present application;

[0051] FIG2C shows a schematic diagram of a 2-classification interval according to some embodiments of the present application;

[0052] FIG2D shows a schematic diagram of a three-category interval according to some embodiments of the present application;

[0053] FIG2E shows a schematic diagram of interval scaling according to some embodiments of the present application;

[0054] FIG3A shows a schematic diagram of a decentralized distributed system framework 300 according to some embodiments of the present application;

[0055] FIG3B shows a schematic diagram of a process in which nodes 0 and 1 adjust feature data corresponding to task data and send the adjusted feature data to the decision center 00, and the decision center 00 makes a decision, according to some embodiments of the present application;

[0056] FIG4A shows a schematic diagram of a frame system 300 according to some embodiments of the present application;

[0057] FIG4B shows a schematic diagram of a nonlinear structure corresponding to a power adaptation function according to some embodiments of the present application;

[0058] FIG5A is a schematic diagram showing a process in which node 0 receives label feature data from node 1 and noise variance of a decision center to perform neural network model training according to some embodiments of the present application;

[0059] FIG5B shows a schematic diagram of extracting features from sample data of node j according to some embodiments of the present application;

[0060] FIG5C shows a schematic diagram of information interaction between a node j and another node j′ according to some embodiments of the present application;

[0061] FIG5D shows a method in which the decision center 00 calculates the variance σ of the noise according to some embodiments of the present application. 2 Interaction diagram sent to node j;

[0062] FIG5E shows a schematic diagram of a model training process according to some embodiments of the present application;

[0063] FIG6A is a schematic diagram showing a process of transmitting a power adaptation function through key point information according to some embodiments of the present application;

[0064] FIG6B shows a schematic diagram of a curve corresponding to a power adaptation function f0 according to an embodiment of the present application;

[0065] FIG6C shows a schematic diagram of a node j sending key point information of a power adaptation function to a decision center 00 according to some embodiments of the present application;

[0066] FIG7 shows a schematic diagram of a centralized distributed system framework system 700 according to an embodiment of the present application;

[0067] FIG8A shows a schematic diagram of LeNet-5 according to an embodiment of the present application;

[0068] FIG8B shows a schematic diagram of a fully connected neural network according to an embodiment of the present application;

[0069] FIG9 shows a schematic diagram of the hardware structure of an electronic device 10 according to an embodiment of the present application. DETAILED DESCRIPTION

[0070] The illustrative embodiments of the present application include, but are not limited to, a data processing method, a readable storage medium, and an electronic device.

[0071] It is understandable that the data processing method provided in the embodiment of the present application can be applied in any wireless communication scenario, including but not limited to wireless communication systems with perception functions, such as 5G-fifth generation mobile communication, 6G-sixth generation mobile communication and other wireless communication systems, as well as short-range wireless communication systems such as Wi-Fi and ultra-wideband wireless communication technology (UWB), etc., without limitation. Moreover, the data processing method provided in the embodiment of the present application can be widely used in any terminal device or network side device in a wireless communication scenario, without limitation.

[0072] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0073] FIG1A is a schematic diagram showing a scenario of transmission based on the same task in a distributed system according to an embodiment of the present application.

[0074] Understandably, in a distributed system, there will be a group of nodes (i.e., node devices) collaborating on the same classification task, as well as a decision center (i.e., decision device) that can integrate, analyze, and process signals from different nodes. Each node can collect different task data for the classification object (e.g., taking images from different angles of an environment, collecting sounds from different locations in an environment, etc.), and send the collected task data to the decision center, which then classifies the classification object based on the received task data.

[0075] It is understood that the classification objects in the embodiments of the present application may include, but are not limited to, an environment in an image, an object in an image, a speaker in audio, a location where audio is produced, a description object in text, a sentiment in text, and semantics in text. Task data may include, but are not limited to, text data, audio data, image data, and the like.

[0076] It can be understood that the nodes proposed in the embodiments of the present application can be any electronic device with processing capabilities, including but not limited to mobile phones, tablet computers, vehicle-mounted equipment, augmented reality (AR) / virtual reality (VR) devices, ultra-mobile personal computers (UMPCs), netbooks, personal digital assistants (PDAs), servers, cameras, etc. The decision center proposed in the embodiments of the present application can be any electronic device with decision-making capabilities, including but not limited to base stations, routers, computer room hosts, mobile phones, tablet computers, vehicle-mounted equipment, augmented reality / virtual reality devices, ultra-mobile personal computers (UMPCs), netbooks, personal digital assistants, servers, etc., without limitation here.

[0077] The following explanation is based on an example in which the classification object is a person in an image, the node is a camera, and the decision center is a base station.

[0078] As shown in Figure 1A, a distributed system uses multiple cameras, such as cameras A1, A2, ..., and An, as nodes, with base station B as the decision center. Cameras A1-An collaborate with base station B to complete a classification task. The following example illustrates the binary classification task performed by cameras A1-An and base station B.

[0079] Specifically, cameras A1-An capture a classified object G in environment 1. Because cameras A1-An are located at different locations on the road, they capture the classified object G from different angles, resulting in different image data. For example, Figures 1B and 1C show images T1 and T2 of the classified object G, respectively, captured by camera A1. Figures 1D and 1E show images T3 and T4 of the classified object G, respectively, captured by camera A2. As can be appreciated, due to the different shooting angles, the image quality of images T1, T2, T3, and T4 varies. The image of the classified object G captured by image T1 is significantly clearer than that captured by images T2, T3, and T4. Each camera then processes the captured image data into a transmittable task signal and transmits it via a channel to base station B. Base station B, acting as the decision center, makes the final classification decision based on the task signals received from each camera, determining whether the gender of the classified object G in environment 1 is male or female.

[0080] In some embodiments, each camera transmits the image content corresponding to the classification task to base station B, allowing the base station to make decisions based on the image content. For example, camera A1 transmits the content of images T1 and T2 to base station B. Other cameras use the same method and are not described in detail here. Base station B makes decisions based on the task signals received from each camera. Specifically, as shown in Figure 2A, for camera A1, camera A1 can process the captured images T1 and T2 through encoder 01 and linear filter 02 to obtain processed data. The processed data is then converted into a transmittable task signal and transmitted to base station B via a channel. Based on the data received from camera A1, base station B restores it through decoder 03 to obtain data used for decision-making. For other cameras, the interaction process with base station B is essentially the same as that of camera A1 and is not described in detail here. Base station B makes decisions based on the task data received from each corresponding node.

[0081] However, when camera A1 sends the task signal corresponding to image T1 and image T2 to base station B, the task signal will be affected by channel noise (for example, the distribution of channel noise is Z0~N(0, σ 2 )), resulting in the signal received by base station B containing both the task signal and the noise signal. Therefore, in some scenarios, since the signal received by base station B is significantly affected by noise, the classification data obtained by the receiving end from the received signal may be inaccurate, resulting in insufficient accuracy of the task results determined by the receiving end.

[0082] For example, the signal received by base station B includes not only the task signals corresponding to images T1 and T2, but also noise signals. Specifically, before transmission, image T2 could be classified as male based on its task data. However, because image T2 is a back-view of a person, the male features are not clearly visible. During transmission, noise interference occurs. Base station B receives a signal that includes not only the task signal for image T2 but also the noise signal. Based on the signal received corresponding to image T2, base station B may mistakenly identify male features as female, potentially misclassifying male as female.

[0083] In other embodiments, each camera may first classify the currently captured image data to obtain a classification result, such as whether the gender is male or female, and then send the classification result to base station B. Base station B makes a decision based on the classification results sent by each camera to obtain a final decision result. For example, it compares the number of classification results corresponding to "male" and the number of classification results corresponding to "female" in the classification results sent by each camera and selects the classification result with the larger number as the final decision result. Specifically, as shown in Figure 2B, each camera in Figure 1A determines whether the object in the current environment 1 is male or female based on the image-related data it captures (i.e., obtains a classification result). The obtained classification result is then encoded by encoder 11 to obtain task data. For example, the corresponding classification result of "male" can be encoded as 0 and the corresponding classification result of "female" can be encoded as 1. The classification result is then processed by binary phase shift keying (BPSK) 12 and sent to base station B. Base station B restores the classification result through decoder 13 and makes a decision to obtain the decision result.

[0084] It can be understood that, similarly, the signal received by base station B includes not only the task signal corresponding to each classification result, but also the noise signal, such as the noise shown in FIG2B (e.g., the distribution of channel noise is Z0~N(0, σ 2 Since only the classification results need to be transmitted, binary phase-shift keying (BPSK) requires only two phases for transmission. The classification results can be distinguished by the phase difference between the two. Therefore, slight changes in the task signal caused by noise will not affect the transmission result.

[0085] However, since base station B can only make its final decision based on the classification results processed by each camera, it cannot consider the correlation between the image data collected by each camera, nor can it determine the reliability of the classification results obtained by each node. Therefore, the decision results are too one-sided and lack high accuracy. For example, if three cameras classify object G as male based on their own images of the object's side and back, while one camera captures the object's face and classifies it as female, the base station will determine that the object is male based on the results received from each camera, resulting in an incorrect classification. Furthermore, each camera needs to process the image data to obtain the classification result, which consumes a lot of resources when there are many cameras.

[0086] In some embodiments, for a classification task involving K classifications of a categorized object, each transmitting end may perform feature extraction on the task data collected by the respective categorized object, obtain feature data corresponding to the task data, and transmit a feature signal corresponding to the feature data to the receiving end. The feature data is used to indicate the classification corresponding to the task data collected by the transmitting end, and the amplitude of the feature signal is positively correlated with the magnitude of the feature data. After receiving the feature signals transmitted by multiple transmitting ends, the receiving end may restore the feature data corresponding to each feature signal based on the amplitude of each received feature signal; the receiving end may then determine the category of the categorized object based on the restored feature data.

[0087] It should be understood that the specific manner in which the receiving end can determine the category of the classification object based on the restored feature data will be introduced below and is not limited here.

[0088] It should be understood that in some embodiments, due to the power limit of the characteristic signal that can be transmitted by the transmitting end, and the amplitude of the characteristic signal is positively correlated with the size of the characteristic data, if the value of the characteristic data is too large, the transmitting end will not have sufficient power to transmit the characteristic signal corresponding to the characteristic data. Therefore, in some embodiments, the characteristic data of the task will be adjusted before transmission. In some embodiments, the characteristic data will be scaled proportionally, and the transmitting end will send the characteristic signal corresponding to the scaled characteristic data to the receiving end. However, because the characteristic signal is significantly affected by noise, the data obtained by the receiving end after restoring the received signal is inaccurate.

[0089] Based on this, the present application proposes a data processing method, in which, after the transmitting end collects the feature data corresponding to the task data, it adjusts the feature data in a nonlinear manner before sending to obtain the adjusted feature data, and sends the adjusted feature data to the receiving end.

[0090] In some embodiments, a power adapter model capable of performing pre-transmission adjustment can be determined through deep learning, and the power adapter model can perform nonlinear adjustment on the characteristic data. Specifically, the power adapter model used by each node device at the departure end for pre-transmission adjustment is determined in advance through a neural network model. In particular, when training the neural network model, it is necessary to consider the statistical characteristics of the noise of the channel of each node and the characteristic data of the samples of all nodes. At this time, since the characteristic data of the samples of all nodes and the noise of the channel of each node are taken into account, the training process not only considers the correlation between the data used to perform the same task between each node, but also considers adaptation to the noise of the channel of each node. As a result, the anti-noise interference ability of the characteristic data obtained after pre-transmission adjustment using the corresponding power adapter model after training is improved, thereby improving the accuracy of the classification results obtained when the subsequent decision center performs classification tasks based on the received data.

[0091] It is understandable that in a distributed system, since both the sending end and the receiving end have strong computing and processing capabilities, when training the neural network model, the neural network model training can be performed at the sending end or at the receiving end.

[0092] In some embodiments, since information can be transferred between nodes in a distributed system, for convenience, neural network model training can be performed at the transmitting end to determine the power adapter model required for each node. Furthermore, after the transmitting end determines the power adapter model required for each node, it is necessary to send the power adapter model required for each node to the receiving end, so that the receiving end can make classification decisions based on the power adapter model corresponding to each node. In other embodiments, since information cannot be transferred between nodes in a distributed system, for convenience, neural network model training can be performed at the receiving end to determine the power adapter model required for each node.

[0093] Similarly, after the receiving end determines the power adaptation sub-model required by each node, it needs to send the power adaptation sub-model required by each node to each node of the transmitting end, so that each node of the transmitting end can perform pre-transmission adjustment according to the corresponding power adaptation sub-model.

[0094] It can be understood that in some embodiments, in order to reduce the amount of data transmission and thus save transmission resources, after the transmitting end obtains the power adaptation sub-model corresponding to each node after training the neural network model, the input and output data pairs corresponding to the power adaptation sub-model corresponding to the node can be sent to the receiving end; the receiving end can restore the required power adaptation sub-model based on the received input and output data pairs corresponding to the power adaptation sub-model corresponding to the node, so as to be used for subsequent decision-making.

[0095] In some implementations, the decision-making device may input input and output data pairs corresponding to the power adaptation sub-model of each node into a parameter-containing function model to obtain the solved values ​​of each parameter in the parameter-containing function model corresponding to each node; replace the parameters in the parameter-containing function model of each node with the corresponding solved values ​​to obtain a power adaptation function, and then use the obtained power adaptation function of each node as the recovered power adaptation sub-model. The parameter-containing function model includes, but is not limited to, an exponential model with parameters, a Tanh function model, an Arctan function model, a power series model, and the like.

[0096] Similarly, after the receiving end obtains the power adaptation sub-model corresponding to each node through neural network model training, it can send the input and output data pairs corresponding to each node's power adaptation sub-model to the corresponding node on the transmitting end. The corresponding node on the transmitting end can also recover the required power adaptation sub-model based on the received input and output data pairs corresponding to its own power adaptation sub-model, thereby performing pre-transmission adjustments. Each node uses the same method to obtain the corresponding power adaptation function and uses the obtained power adaptation function as the recovered power adaptation sub-model.

[0097] It can be understood that in some embodiments, the statistical characteristics of channel noise can be mean vector, covariance, arithmetic mean, geometric mean, harmonic mean, weighted mean, square mean, median, median number, mode, mean absolute deviation, and variance.

[0098] In addition, when the neural network model is trained at the transmitting end, the transmitting end also needs to send the statistical characteristics corresponding to each classification label in the classification labels corresponding to the multiple classification categories to the decision device, wherein the statistical characteristics corresponding to each classification label in the multiple classification labels are the statistical characteristics of the feature data of the corresponding classification label in all feature data corresponding to the sample sets of all nodes in the distributed system. For example, it can be the mean vector and covariance of the feature data corresponding to the classification labels corresponding to each classification category, so as to facilitate the subsequent decision center to make decisions. It can be understood that the statistical characteristics corresponding to each classification label in the classification labels corresponding to the multiple classification categories can include, in addition to the mean vector and covariance, also but not limited to the arithmetic mean, geometric mean, harmonic mean, weighted mean, square mean, median, median number, mode, mean absolute deviation, variance, etc.

[0099] It can be understood that in some embodiments, the neural network model to be trained may include but is not limited to a fully connected neural network, a convolutional neural network, etc.

[0100] In addition, it can be understood that in a classification task, the value range of the numerical values of the feature data can be divided into K feature value intervals, each feature value interval corresponding to a classification category, and the critical value between two adjacent feature value intervals can be referred to as a classification boundary. For example, for a binary classification task, as shown in FIG. 2C, assuming that the value range of the numerical values of the feature data is [A1, B1], the value range of the numerical values of the feature data can be divided into 2 feature value intervals, namely the feature value interval [A1, C1] corresponding to category 1 and the feature value interval (C1, B1] corresponding to category 2. The boundary between the feature value interval [A1, C1] and the feature value interval (C1, B1] is C1, that is, C1 is the classification boundary. Another example, for a three-class classification task, as shown in FIG. 2D, assuming that the value range of the numerical values of the feature data is [A2, B2], the value range of the numerical values of the feature data can be divided into 3 feature value intervals, namely the feature value interval [A2, C2] corresponding to category 1, the feature value interval (C2, D2] for category 2, and the feature value interval (D2, B2] corresponding to category 3, where A2 < C2 < D2 < B1. At this time, the classification boundary between the feature value interval [A2, C2] and the feature value interval (C2, D2] is C2, that is, C2 is the classification boundary 1; the classification boundary between the feature value interval (C2, D2] and the feature value interval (D2, B2] is D2, that is, D2 is the classification boundary 2.

[0101] It should be understood that in some embodiments, due to the limitation of the power of the feature signal that the transmitting end can send, and the amplitude of the feature signal is positively correlated with the size of the feature data. If the numerical value of the feature data is too large, it will cause the transmitting end not to have enough power to send the feature signal corresponding to the feature data. Therefore, in some embodiments, if the value range of the feature data is too large, the transmitting end needs to scale the feature data so that the transmitting end can send all the feature data within the value range of the feature data. For example, assuming that the value range of the feature data is [A1, B1], and the value range of the feature data that the transmitting end can send is [A3, B3], when the transmitting end sends within the range of [A1, B1], it maps it to the range of [A3, B3] (such as scaling in proportion), and then sends the scaled feature signal.

[0102] It should be understood that if the feature data is scaled, the corresponding classification boundary will also be scaled. Referring to FIG. 2E, corresponding to the value range of the feature data being [A1, B1], and the value range of the feature data that the transmitting end can send being [A3, B3], the classification boundary C1 within the value range of [A1, B1] will be scaled to the classification boundary C3 as shown in the figure.

[0103] It is understandable that after the transmitting end collects the feature data corresponding to the task data, it uses the power adapter model obtained by neural network training to perform nonlinear adjustment on the feature data of the task data to obtain the adjusted feature data. In some cases, the scaling ratio of the adjusted feature data to the feature data before adjustment changes with the change of the feature data before adjustment. Assuming that the value range of the feature data before adjustment is a first value range, the value range of the output data of the power adapter model is a second value range, and the data output after the first classification boundary in the first value range (for example, C1 in Figure 2E) is input to the power adapter model is a second classification boundary (for example, C3 in Figure 2E). At this time, the ratio of the second distance of the adjusted feature data minus the second classification boundary to the first distance of the feature data before adjustment minus the first classification boundary increases as the first distance decreases. That is, the feature data close to the classification boundary is far away from the classification boundary after adjustment. In this way, the distance between the feature data close to the classification boundary after the original proportional scaling and the classification boundary can be increased, thereby avoiding the situation where the value of the feature data close to the classification boundary changes from the feature value interval corresponding to one classification category to the feature value interval corresponding to another classification category during the transmission process due to noise signals.

[0104] After receiving the characteristic signals sent by each transmitting end after pre-transmission adjustment using the power adaptation sub-model, the receiving end can receive the signals from each transmitting end (hereinafter referred to as the "receiving end signal" for ease of explanation), restore the receiving end signal to obtain the corresponding data (for ease of explanation, the data corresponding to the "receiving end signal" is referred to as the "receiving end data"), and make decisions based on the receiving end data from each node, such as performing category prediction, to obtain a decision result. It should be understood that due to the nonlinear adjustment, since the adjusted characteristic data still belongs to the same classification category, the receiving end uses the receiving end data affected by noise interference to make decisions, and the decision results obtained are more accurate.

[0105] For example, in the scenario shown in FIG1A , camera A1 performs feature extraction on image T1 and image T2 using a pre-trained LeNet-5 model to obtain one-dimensional feature data J1 and J2 corresponding to image T1 and image T2 , respectively.

[0106] Because image T2 is a back view of the object G, the male characteristics of the person reflected in image T2 are not obvious, and the resulting one-dimensional feature data J2 is close to the classification boundary. Referring to Figure 2C above, assuming that category 1 is male and category 2 is female, the one-dimensional feature data J2 is close to the classification boundary C1. The one-dimensional feature data J2 is numerically adjusted to obtain one-dimensional feature data J2'. At this time, the one-dimensional feature data J2' is farther away from the classification boundary C1 than the one-dimensional feature data J2.

[0107] Since image T1 is a frontal photo of the classification object G, the male features of the person reflected by image T1 are relatively obvious, and the obtained one-dimensional feature data J1 will be far away from the classification boundary C1. In addition, in some cases, after adjustment through the power adaptation sub-model, feature data far away from the classification boundary can also be mapped to data close to the classification boundary, but the category of the feature data does not change. For example, if the one-dimensional feature data J1 is numerically adjusted to obtain one-dimensional feature data J1', at this time, the one-dimensional feature data J1' will be slightly closer to the classification boundary C1 than the one-dimensional feature data J1. However, at this time, the value of the one-dimensional feature data J1' will still be greater than the one-dimensional feature data J2'. Camera A1 sends the feature signals corresponding to the adjusted one-dimensional feature data J1' and J2' to base station B. Other cameras can achieve the same effect after processing the collected images and sending them to base station B.

[0108] Base station B receives the signals sent by each transmitting end and obtains the receiving end signal, restores the receiving end signal to obtain the corresponding receiving end data, that is, image feature data, and performs gender prediction based on the receiving end data from each node to obtain the gender result.

[0109] It is understandable that after the characteristic data at the classification boundary is adjusted, the anti-interference ability of the corresponding characteristic signal during the transmission process is improved, which can avoid the problem that in some scenarios, the accuracy of the determined task results is affected by the low accuracy of the data that is more critical to the classification due to the greater influence of noise. In addition, in a distributed scenario, the transmitting end (such as each node) sends the characteristic signal corresponding to the task content to the receiving end (such as the decision center). At this time, when making a decision, the receiving end fully considers the correlation of the data of each node, avoiding the decision result being too one-sided, and also avoiding the problem that each node needs to process the collected task data to obtain its own task result, resulting in more resources consumed when the number of nodes is large.

[0110] It can be understood that during the execution of the task, the characteristic data corresponding to the collected task data is adjusted by the corresponding power adapter model before being sent to obtain the adjusted characteristic data. At this time, the signal corresponding to the adjusted characteristic data is sent to the receiving end. After the receiving end receives the characteristic signal sent by each transmitting end, the receiving end can receive the signal from each transmitting end (for the convenience of explanation, hereinafter referred to as the "receiving end signal"), and restore the receiving end signal to obtain the corresponding data (for the convenience of explanation, the data corresponding to the "receiving end signal" is referred to as the "receiving end data"), and make a decision based on the receiving end data from each node, such as performing category prediction to obtain a decision result. It should be understood that in some cases, due to the nonlinear adjustment of the power adapter function, and the adjusted characteristic data still belong to the same classification category, the receiving end uses the receiving end data affected by noise interference to make a decision, and the accuracy of the decision result obtained is relatively high.

[0111] For ease of understanding, the following describes a distributed system consisting of two nodes and a decision center. The system determines the power adaptation sub-model for each node, how each node uses the model to adjust its characteristic data, and how the decision center makes decisions based on the data corresponding to the received signal.

[0112] FIG3A shows a schematic diagram of a framework system 300 in which two nodes transmit data to a decision center 00 based on their respective corresponding power adaptation sub-models, so that the decision center 00 makes a decision, according to some embodiments of the present application.

[0113] As shown in Figure 3A, node 0 includes a feature extraction module 301 and a power adaptation module 302. Similarly, node 1 includes a feature extraction module 311 and a power adaptation module 312. Among them, the feature extraction module 301 is used to obtain task data from node 0 based on a feature extractor (such as a pre-trained model) to perform feature extraction and obtain feature data (such as x0 shown in the figure). The power adaptation module 312 can perform numerical adjustments on different input feature data based on the determined power adaptation sub-model corresponding to node 0 (such as the power adaptation sub-model f0 corresponding to node 0) to obtain adjusted feature data (such as the adjusted feature data x0' shown in the figure). For example, the feature data close to the classification boundary can be adjusted to be farther away from the classification boundary than before, and the classification category will not be changed.

[0114] The feature extraction module 311 and the power adaptation module 312 are substantially the same as the feature extraction module 301 and the power adaptation module 302 , respectively, and are not described again here.

[0115] It can be understood that after the power adaptation module of each node (such as the above-mentioned power adaptation module 302 and power adaptation module 312) makes numerical adjustments to each characteristic data based on the corresponding power adaptation module, the average power of the adjusted data transmitted will not exceed the power allowed by the node (such as the rated power). For different nodes, since the characteristic data distribution and the allowed power (hereinafter referred to as "allowed power") of each node are different, each adaptation module will be processed based on a different power adaptation sub-model, and after the corresponding characteristic data is adjusted, there will be a certain range.

[0116] For node 0, the characteristic signal corresponding to the adjusted characteristic data is transmitted to the decision center 00. At this time, the decision center 00 receives the receiving end signal W0 from node 0. The receiving end signal W0 includes not only the adjusted characteristic signal but also the channel noise Z0 (Z0~N(0, σ) as shown in FIG3A 2 Similarly, for node 1, decision center 00 can receive receiving signal W1 from node 1. Based on the signals received from each node, decision center 00 uses the data corresponding to the receiving signal to make a decision and obtain a decision result. For example, in a binary classification task, decision center 00 uses the data corresponding to the receiving signal to calculate the likelihood function value and determines the corresponding specific category based on the likelihood function value. The following describes the technical solution of this application in conjunction with the framework system 300 shown in Figure 3A.

[0117] FIG3B shows a schematic diagram of a process in which nodes 0 and 1 adjust feature data corresponding to task data and send the adjusted feature data to the decision center 00 , and the decision center 00 makes a decision, according to some embodiments of the present application.

[0118] S101, node 0 collects task data R0.

[0119] In some embodiments, node 0 collects task data R0 related to the target task (ie, as data to be classified). The task data R0 may include multiple task data or one task data.

[0120] For example, in the scenario shown in FIG1A , node 0 collects image data a0 for a binary classification task, where the image data a0 is data corresponding to one image or multiple images, that is, the task data R0 is the image data a0.

[0121] S102, node 0 extracts features from task data R0 to obtain feature data x0.

[0122] In some embodiments, node 0 performs feature extraction on task data R0 through feature extraction module 301 to obtain feature data x0 (ie, as classification feature data).

[0123] In some embodiments, the feature data x0 is one-dimensional data, that is, a single value. For example, when the task data R0 is multiple data, after feature extraction of the task data R0, multiple one-dimensional data will be obtained, that is, the feature data x0 is multiple data.

[0124] For example, node 0 performs image feature extraction on the image data a0 corresponding to multiple images through the LeNet-5 model to obtain multiple one-dimensional image feature data m0, that is, each one-dimensional image feature data m0 corresponds to an image, and the one-dimensional image feature data m0 corresponding to the image data a0 is the feature data x0.

[0125] S103 , node 0 adjusts the characteristic data x0 based on the power adaptation sub-model f0 to obtain adjusted characteristic data x0 ′.

[0126] It can be understood that the power adaptation sub-model f0 can adjust the corresponding input data according to its situation with the classification boundary.

[0127] In some embodiments, the feature data x0 includes multiple feature data. When multiple feature data x0 with different values ​​are input into the power adaptation module 302, the adjusted feature data x0' corresponding to each input feature data will be output. Since the power adaptation sub-model f0 used in the power adaptation module 302 is a nonlinear model, for example, referring to Figure 2C above, assuming that the classification boundary is C1, the feature data x0 with a value of b is on the right side of C1 and is close to the classification boundary C1. At this time, after passing through the power adaptation sub-model f0, it will be mapped to a value farther away from C1. However, at this time, the feature data x0 still belongs to category 2.

[0128] For example, a plurality of one-dimensional image feature data m0 with different values ​​are input into the power adaptation module 302 , and a plurality of adjusted image feature data m0 ′ corresponding to the inputs can be output.

[0129] S104 , the node 0 transmits the characteristic signal V0 corresponding to the adjusted characteristic data x0 ′ to the decision center 00 .

[0130] In some embodiments, the node 0 transmits the characteristic signal V0 corresponding to the adjusted characteristic data x0′ to the decision center 00 through a radio frequency device, and the characteristic signal V0 may be interfered by the channel noise Z0 during the transmission process.

[0131] For example, the node 0 transmits the image feature signal corresponding to the adjusted image feature data m0 ′ to the decision center 00 through a radio frequency device, and the image feature signal may be interfered by the channel noise Z0 during the transmission process.

[0132] S105 , the decision center 00 obtains the receiving end signal W0 corresponding to the adjusted characteristic data x0 ′.

[0133] In some embodiments, the decision center 00 obtains a receiving end signal W0 corresponding to the adjusted characteristic signal x0 ′, where the receiving end signal W0 includes the characteristic signal V0 corresponding to the adjusted characteristic data x0 ′ and the channel noise Z0 .

[0134] For example, the decision center 00 obtains a receiving end signal corresponding to the adjusted image feature data m0 ′, where the receiving end signal includes the adjusted image feature signal and the channel noise Z0 .

[0135] S106, node 1 collects task data R1.

[0136] S107 , node 1 performs feature extraction on task data R1 to obtain feature data x1 .

[0137] S108 , node 1 adjusts the characteristic data x1 based on the power adaptation sub-model f1 to obtain adjusted characteristic data x1 ′.

[0138] S109 , the node 1 transmits the characteristic signal V1 corresponding to the adjusted characteristic data x1 ′ to the decision center 00 .

[0139] S110, the decision center 00 receives a receiving signal W1 corresponding to the characteristic data x1'.

[0140] It can be understood that the above S106-S110 are essentially the same as the above S101-S105, and will not be repeated here.

[0141] S111, the decision center 00 makes a decision based on the receiving end data s0 and s1 corresponding to the receiving end signal W0 and the receiving end signal W1 respectively, and obtains a decision result.

[0142] In some embodiments, the decision center 00 predicts a category based on the received data s0 and s1 corresponding to the received signals W0 and W1, respectively, and obtains a decision result based on the predicted category. It is understood that since the received signals W0 and W1 are electromagnetic waves, the received signals W0 and W1 need to be demodulated to obtain the corresponding received data s0 and s1.

[0143] In some implementations, the receiving end data (s0, s1, ..., s d-1 ) and make category predictions based on the likelihood function value. Calculate the likelihood function value corresponding to each label (i.e., classification category) and take the category with the largest likelihood function value as the decision result.

[0144] It is understandable that when the decision center 00 calculates the likelihood function value of the received end data, it is necessary to consider the statistical characteristics of the noise and the statistical characteristics of the characteristic data transmitted by each node. The following formula (1) shows a classification scenario, based on the received end data (s0, s1, ..., s d-1 ) calculates the expression of the likelihood function value corresponding to each classification category,

[0145] in, is the distribution N(μ given by label k k ,Σ k ) is the feature data generated by the training sample, that is, the simulated feature data obtained according to the distribution corresponding to the statistical characteristics of the training sample, where the label k is the number of the specific classification category, for example, k = 1 corresponds to category 1, which is male, and k = 2 corresponds to category 2, which is female. k and Σ k are the mean vector and covariance of the feature data corresponding to each label obtained based on the training samples of all nodes during the training process, σ 2 is the variance of the channel noise observed by decision center 00, f j (*) is the power adapter sub-model corresponding to each node. The specific explanation will be introduced in the training process below and will not be described in detail here.

[0146] For example, in the above-mentioned binary classification scenario, when the decision center 00 obtains the likelihood function value of the corresponding category of male as 0.4 based on the receiving data s0 and the receiving data s1, and the likelihood function value of the corresponding category of female is 0.7, then the corresponding category is determined to be female, that is, the decision result is female.

[0147] It is understood that the execution order of the above-mentioned steps S101 to S111 is only a schematic. In other embodiments, other execution orders may be adopted, and some steps may be split or merged, which is not limited here. For example, the order in which the above-mentioned node 1 executes the process of S106 to S109 and the node 0 executes the process of S101 to S104 can be changed, that is, node 1 executes S106 to S109 before node 0, or it can be executed in parallel with node 0, and there is no strict order in the execution process. The order in which the decision center 00 receives signals from node 0 and node 1 is not fixed, and can also be received simultaneously.

[0148] As can be understood, processing feature data through the power adaptation sub-model improves its anti-interference capabilities during transmission, thereby improving its transmission quality. This can avoid scenarios where the accuracy of task results determined by the receiving end is affected by poor noise immunity and low accuracy of important feature data received. This saves resources while ensuring the highest possible accuracy. Furthermore, the decision center leverages the correlation between data from each node to help it make better decisions.

[0149] It can be understood that during the execution of the above-mentioned tasks, if the neural network model training is performed in advance at the transmitting end, and the transmitting end sends the first adaptation processing information to the receiving end, wherein the first adaptation processing information includes the input and output data pairs corresponding to each node, then the receiving end will use the power adaptation function obtained according to the input and output data pairs, and use the power adaptation function as the recovered power adaptation sub-model to make a decision, that is, the power adaptation sub-model used for the likelihood function calculation in the above-mentioned step S111 is the power adaptation function.

[0150] It can be understood that during the execution of the above-mentioned tasks, if the neural network model training is performed in advance at the transmitting end, and the transmitting end sends the second adaptation processing information to the receiving end, where the second adaptation processing information includes the power adaptation sub-model corresponding to each node, the receiving end will directly use the power adaptation sub-model to make decisions, that is, the power adaptation sub-model used for the likelihood function calculation in the above-mentioned step S111 is the received power adaptation sub-model.

[0151] It can be understood that during the execution of the above-mentioned tasks, if the neural network model training is performed at the receiving end in advance, and the receiving end sends the third adaptation processing information to each transmitting end (i.e., each node device) respectively, wherein the third adaptation processing information includes the input and output data pairs of the power adaptation sub-model of the corresponding node, then each transmitting end will obtain the power adaptation function based on the input and output data pairs, and use the power adaptation function as the power adaptation sub-model to adjust the characteristic data before sending, that is, the power adaptation sub-model used in the above-mentioned steps S103 and S108 is the power adaptation function.

[0152] It can be understood that during the execution of the above-mentioned tasks, if the neural network model training is performed in advance at the receiving end, and the receiving end sends the fourth adaptation processing information to each transmitting end (i.e., each node device) respectively, wherein the fourth adaptation processing information includes the power adaptation sub-model of the corresponding node, then each transmitting end will adjust the characteristic data before sending according to the obtained power adaptation sub-model, that is, the power adaptation sub-model used in the above-mentioned steps S103 and S108 is the received power adaptation sub-model.

[0153] The following describes in detail the process of determining the power adaptation sub-model of each node based on deep learning.

[0154] For the sake of convenience, the following article will still use a distributed system including 2 nodes and 1 decision center as an example for detailed introduction.

[0155] The following describes how to determine the power adapter model in conjunction with FIG. 4A to FIG. 6C .

[0156] In some embodiments, when deep learning techniques are employed, a neural network model is used for training, wherein the neural network model includes a power adapter sub-model corresponding to each node. It is understood that after training the neural network model, a trained neural network model can be obtained, thereby determining the power adapter sub-model corresponding to each node.

[0157] It is understandable that, as mentioned above, there are two situations for the training phase:

[0158] (1) Nodes can exchange information with each other. For example, the cameras shown in Figure 1A are close to each other and can exchange signals, such as through information sharing. For ease of explanation, a distributed system that includes a decision center and nodes that can exchange information with each other will be referred to as a "decentralized distributed system."

[0159] (2) Information exchange between nodes is not permitted. For example, information exchange between the cameras shown in Figure 1A is not permitted. For ease of explanation, a distributed system that includes a decision-making center and nodes that cannot exchange information with each other is referred to as a "centralized distributed system."

[0160] (1) Decentralized distributed system

[0161] For the above situation (1), since information can be exchanged between nodes, the node (i.e., the transmitting end) can obtain the characteristic data of all nodes, and thus perform neural network model training at the transmitting end to determine the required power adapter model.

[0162] In conjunction with Figures 4A-6C, neural network model training is first performed under a "decentralized distributed system" to determine the respective power adaptation sub-models. Furthermore, after the power adaptation sub-model is determined at the node (i.e., the transmitting end), the power adaptation sub-model needs to be synchronized at the receiving end so that the receiving end can make decisions during use. Referring to step S111 in Figure 3B, the receiving end (i.e., the decision center 00) needs to use the trained power adaptation sub-models of each node to perform category prediction.

[0163] Figure 4A shows a schematic diagram 400 of a decentralized distributed system framework according to some embodiments of the present application. To facilitate comparison with the schematic diagram of the framework of the actual use process after training shown in Figure 3A, Figure 4A still uses the two nodes and one decision center shown in Figure 3A as an example. The two nodes can interact with each other.

[0164] As shown in FIG4A , node 0 includes a feature extraction module 301 a and a power adaptation module 302 a , and node 1 includes a feature extraction module 311 a and a power adaptation module 312 a .

[0165] In some embodiments, the feature extraction module 301a is used to extract features from each sample data in the sample set of node 0 (i.e., sample 0 in the figure) based on a feature extractor (such as a pre-trained model) to obtain feature data. Since each sample data in the sample set has a classification label, each corresponding feature data has a corresponding classification label. At this time, the label feature data x corresponding to each classification label can be obtained. k；0 , where 0 represents node 0, and subscript k represents the number of the classification label corresponding to the classification category. For example, in a binary classification task, k = 1 and 2, where k = 1 represents male and k = 2 represents female. It can be understood that the feature extractor used here is the same as that used in Figure 3A above.

[0166] The power adaptation module 302a includes a neural network model that needs to be trained. Before training, the parameters in the neural network model can be any value. After training, the parameters in the neural network model are values ​​that make the feature data of node 0 adapt to the noise of the channel. k；0 Perform nonlinear adjustment to obtain the adjusted label feature data x' k；0 For example, as shown in FIG4B , the label feature data x of each node k；j Through nonlinear adjustment of the power adapter model, we can get f j (x k；j ). Where j is the node number, for example, for node 0, it is 0, f j (x k；j ) is the adjusted label feature data x' k；0 .

[0167] It can be understood that the functions of the feature extraction module 311a and the power adaptation module 312a are substantially the same as those of the feature extraction module 301a and the power adaptation module 302a, and are not described in detail here.

[0168] It is understandable that in the case where information exchange can be performed between each node, since data exchange can be performed between each node, each node can receive feature data from other nodes, so each node can obtain feature data of all nodes. At this time, in order to save resources, as long as one node is selected to obtain the feature data of all nodes, and the neural network model is trained according to the feature data, the power adaptation sub-model corresponding to each node can be determined. In addition, in some implementations, each node will transmit the obtained power adaptation sub-model to the corresponding decision center 00, so that the decision center 00 can use it in subsequent decisions. In other implementations, the node performing the training process (assuming it is node 0 in Figure 4A) can also transmit the power adaptation sub-model corresponding to each node to the corresponding decision center 00.

[0169] Specifically, the training process is described using node 0 as an example. FIG. 5A below illustrates a process where node 0 receives label feature data from node 1 and statistical characteristics of noise (e.g., noise variance) from a decision center to train a neural network model, according to some embodiments of the present application. Steps S201-S205 represent the specific training process.

[0170] S201, node 0 extracts features from sample 0 and obtains label feature data x corresponding to each label k；0 .

[0171] As can be understood, each node will receive a large amount of sample data for the same classification task, and each sample number will correspond to a classification label. In addition, each node can use a pre-trained feature extractor (for example, for image tasks, the existing LeNet-5 model can be used) to obtain one-dimensional feature data of the sample data. This one-dimensional feature data has already determined the corresponding classification category, thereby removing redundant information in the sample data. As can be understood, the feature extractor used here is the same as the feature extractor used in actual use after training is completed.

[0172] For example, as shown in FIG5B , each node j will extract features from its own sample data, and each sample data will have a corresponding label, and obtain the label feature data x corresponding to each label. k；j , where the subscript k represents the category of the task label and j represents the node number.

[0173] In some embodiments, node 0 can obtain sample data related to the task (i.e., sample 0), wherein each sample data has a corresponding label, and the feature extraction module 301a is used to extract features from the obtained sample data to obtain feature data, and according to the task label information, the label feature data x corresponding to each task label is obtained. k；0, where subscript k represents the category number of the task label, and subscript 0 represents node 0. Taking the binary classification task as an example, k can be 1 and 2.

[0174] For example, in the scenario shown in Figure 1A, node 0 obtains a large number of image sample data 0, each image sample data corresponds to an image, and each image has a corresponding classification category, for example, male or female. Node 0 performs feature extraction on multiple image samples 0 to obtain multiple one-dimensional feature data, such as one image data corresponds to one one-dimensional feature data. At this time, according to the label corresponding to each image sample 0, the multiple one-dimensional feature data are divided into two categories, and the label feature data x corresponding to the label male (for example, k = 1) is obtained. 0；0 ; and obtain the label feature data x corresponding to the label female (for example, k = 2) 1；0 .

[0175] S202, node 1 extracts features from sample 1 and obtains label feature data x corresponding to each label k；1 .

[0176] It can be understood that for node 1, the process is substantially the same as the process in step S201, and will not be described in detail here.

[0177] S203, node 1 sends first characteristic information to node 0, wherein the first characteristic information includes label characteristic data x k；1 .

[0178] It is understandable that data can be exchanged between nodes 0 and 1. As shown in FIG5C , each node j can exchange label feature data with other nodes j′. Node j sends its own label feature data x k；j Send it to other nodes j', and other nodes j' can use the label feature data x k；j Send to node j.

[0179] For the interaction between node 1 and node 0, as shown in FIG4A , node 0 and node 1 can respectively send first feature information to each other, and the first feature information corresponding to node 0 includes the obtained label feature data x k；0 The first feature information corresponding to node 1 includes the obtained label feature data x k；1 Since node 0 is used for neural network model training as an example, node 1 will pass the label feature data x k；1 Sent to node 0 to enable node 0 to train.

[0180] In some implementations, it is assumed that the interaction between nodes will not be affected by channel noise, for example, other data sharing methods can be used for transmission.

[0181] S204 , the decision center 00 sends first noise statistical information to the node 0 , where the first noise statistical information includes statistical characteristics of noise corresponding to each node.

[0182] It is understandable that since the purpose of training the neural network model is to adapt to the channel noise, the channel noise needs to be taken into account when training the neural network model. Since only the decision center 00 can observe the channel noise, the decision center 00 will calculate the statistical characteristics corresponding to the channel noise, such as the variance σ 2 Send to each node. Referring to Figure 5D, the decision center 00 sends the variance σ of the noise 2 Sent to node j, where node j can be any node.

[0183] For example, taking Gaussian noise as an example, the decision center 00 in FIG4A denotes the variance σ corresponding to the channel noise. 2 Sent to each node.

[0184] It can be understood that in this embodiment, the neural network model training is performed on node 0. At this time, the decision center 00 can only send the noise variance σ to node 0. 2 It is understandable that the decision center 00 can also send the noise variance σ to all nodes. 2 , no specific requirements are made here. Then, node 0 can receive the channel noise variance required for training. It can be understood that the variance of different channel noises can be different values. For the sake of convenience, in some embodiments of this application, it is assumed that the variance value of each channel noise is the same σ 2 Take value as an example to illustrate.

[0185] S205, node 0 performs neural network model training according to a preset loss function and the label feature data of each node to obtain a trained neural network model (i.e., a power adapter sub-model corresponding to each node), wherein the loss function involves the statistical characteristics of the allowed power and noise of each node.

[0186] It is understandable that the simulated receiving data corresponding to each node received by the decision center 00 is determined based on the statistical characteristics of the tag feature data and the channel noise (e.g., variance). Since the receiving signal of the decision center 00 includes the tag feature signal of each node after adjusting the tag feature data using the power adaptation sub-model, as well as the channel noise signal, the simulated receiving data includes the tag feature data processed by the power adaptation sub-model and the data corresponding to the noise signal. The specific method for obtaining the simulated receiving data of the decision center 00 is described in the following formula (3), which will not be repeated here.

[0187] At this point, the maximum mutual information between the simulated receiving data and the classification label can be used as the objective function to train the neural network model and obtain the power adapter sub-model corresponding to each node. For a specific description of the mutual information, refer to the description of the following formula (2), which will not be repeated here.

[0188] Furthermore, since the power of each node is fixed, it is necessary to consider that when the node transmits the signal corresponding to each adjusted characteristic data, the power corresponding to the transmitted characteristic signal cannot exceed the allowed power. Therefore, a penalty is added to the objective function corresponding to the mutual information to constrain the power. For a specific description of the mutual information under power constraints, refer to the description of the following formulas (4) to (6), which will not be repeated here.

[0189] Specifically,

[0190] ① Let the power adaptation sub-model of each node be fj, j is the number of each node, for example, j = 0, ..., d-1, j is a natural number. Let the mutual information between the corresponding task label and the receiving end data of the decision center be, I(H; s0, ..., s d-1 ).

[0191] The following formula (2) shows an expression for maximizing mutual information:

[0192] Where I(*) represents mutual information, s0,…,s d-1 Indicates the receiving data of each node received by the decision center, f j The neural network model that maximizes I is the power adaptation sub-model corresponding to each node.

[0193] ② According to the label feature data and the statistical characteristics of the channel noise (for example, the variance of the noise), the theoretical receiving end data (that is, the data corresponding to the signal of each node received by the decision center) can be obtained as follows: s0,…,s d-1 . Among them, each s j (j=0,......,d-1) includes the receiving end data s corresponding to each tag k；j .

[0194] For example, the following formula (3) shows the receiving end data s corresponding to each tag: k；j The expression: s k；j =f j (x k；j )+z j ,z j ~N(0,σ 2 )Formula (3)

[0195] Among them, z j Represents noise data, k is the label category, that is, sk；j It represents the label feature data of each node obtained after being affected by the power adaptation sub-model and the interference of channel noise.

[0196] ③ Based on the introduction in ① and ② above, we can get a loss function with power constraints on each node as shown in the following formula (4), where the power is constrained by adding a penalty term to the loss function. The loss function is:

[0197] in, That is the above I(H;s0,…,s d-1 )’s specific expression formula. represents the power used by the transmitting end (i.e., each node) when transmitting the signal corresponding to the tag feature data adjusted by the power adaptation sub-model, and p j Represents the allowed power of each node, that is, the power allowed to pass through the channel. Weight λ j is a hyperparameter that needs to be tuned, and the weight λ j Bottom, training set Not greater than p j .

[0198] The explanations of each item in the loss function shown in the above formula (4) are as follows:

[0199] (1) For the mutual information term Introduction to the mutual information term Reflects the mutual information between the label and the receiving end data, through the label feature data x of each label k k；j And the corresponding label feature data x k；j The receiving end data s k；j Calculated.

[0200] Assume that the corresponding label feature data x of each node k；j and receiving end data s k；j Included N f,k Under label k, it represents the different data in the label feature data of each node. s,k Under label k, it represents different data in the receiving end data received by each node.

[0201] The following formula (5) is the above formula (4) The specific expansion formula is:

[0202] Among them, The sequence includes and Among them, k is the category number of the label, K is the total number of labels, and is the number of the last label.

[0203] (2) For power constraint Introduction. Indicates the power used when Less than the allowable power p j , when this item is 0, no power constraint is performed, and the objective function is mutual information.

[0204] When the power used If the power is greater than the allowed power pj and the power constraint term is not 0, a penalty is required. In this case, the loss function includes the power constraint term.

[0205] Specifically, With the upper limit p j Do the wrong thing, if The loss is positive, otherwise it is zero.

[0206] The following formula (6) shows The expression for the power used is,

[0207] FIG5E is a schematic diagram illustrating a process of training a neural network model including power adaptation sub-models according to some embodiments of the present application. The process is described using node 0 as an example execution subject.

[0208] S205A, using the label feature data as input data of the neural network model, obtaining the model output result, and calculating the loss function.

[0209] In some embodiments, a neural network model including each power adaptation sub-model is initialized, and the label feature data is input into the neural network model to obtain a model output result. A loss function is calculated based on the model output result and the label feature data.

[0210] For example, based on the model output results and label feature data, the loss function value corresponding to the above formula (5) is calculated.

[0211] S205B, iterating the neural network model for a preset number of rounds, and stopping the training after reaching the preset number of training rounds.

[0212] In some embodiments, stochastic gradient can be used to iterate the parameters in the model, and the training is stopped after reaching a preset number of training rounds (e.g., 50 times).

[0213] S205C: The neural network model with the smallest loss function value in each round is used as the trained neural network model, wherein the trained neural network model includes the required power adapter sub-models.

[0214] In some embodiments, the neural network model corresponding to the round with the smallest loss function value is used as the trained neural network model, wherein the trained neural network model includes the required power adapter sub-models.

[0215] As can be understood, the loss function is calculated based on the label feature data of each node's corresponding label, the noise variance, and the allowed power of each node. The neural network model corresponding to the round with the minimum loss function value is used as the trained neural network model, thereby obtaining the rate adaptation sub-model corresponding to each node. As can be understood, when the loss function is minimized, the mutual information is as large as possible.

[0216] It can be understood that the execution order of the above steps S201 to S205 is only an illustration. In other embodiments, other execution orders may be adopted, and some steps may be split or combined, which is not limited here.

[0217] It is understandable that during training, the characteristic data of the samples of each node are involved in the training process, and the power adapter model corresponding to each node is jointly trained. Therefore, the correlation between the signals of each node is taken into account in the training process, making the rate adapter model of each node more scientific and effective.

[0218] In addition, when node 0 obtains the label feature data of all nodes, it can obtain the joint distribution of the label feature data corresponding to all nodes. Specifically, it can obtain the statistical characteristics of the label feature data corresponding to all nodes, for example, the mean vector μ k , covariance∑ k Where k is the number corresponding to the label category. For example, in the binary classification, when k = 1, the mean vector μ0 and covariance Σ0 of the label feature data corresponding to k = 1 for all nodes are obtained. For k = 2, the essence is the same and will not be repeated here.

[0219] When node 0 obtains the statistical characteristics of the label feature data corresponding to all nodes, for example, the mean vector μ k , covariance Σ k After that, the first label feature statistical information can be sent to the decision center, and the first label feature statistical information includes the statistical characteristics of the label feature data of all nodes. Referring to FIG4A above, for node 0, the statistical characteristics of the label feature data, such as the mean vector μ of the label feature data corresponding to all nodes, are converted into k , covariance Σ k Send it to the decision center, so that the decision center 00 can calculate the mean vector μ k , covariance Σ kDuring use, simulated data is generated for decision making, for example, refer to the introduction in step S111 above. It is understandable that since data between nodes can be transmitted, other nodes, such as node 1 shown in the figure, can also send the mean vector μ of the label feature data corresponding to all nodes to the same node. k , covariance Σ k Sent to decision center 00, no requirement here.

[0220] The following describes the power adapter model at both the receiving and transmitting ends.

[0221] After the power adaptation sub-model is determined at a node (ie, a transmitting end), the power adaptation sub-model needs to be synchronized at a receiving end so that the receiving end can make a decision during use.

[0222] In some embodiments, the node may send a second adaptation processing information to the decision center 00, and the second adaptation processing information includes each power adaptation sub-model, so that the decision center 00 can obtain the required power adaptation sub-model. It is understandable that after node 0 determines the power adaptation sub-model corresponding to each node, it will send the corresponding power adaptation sub-model to other nodes (such as node 1). In addition, in some implementations, node 0 can send the power adaptation sub-model corresponding to all nodes to the decision center 00 so that the decision center 00 will use it in the decision-making process. In other implementations, each node can send the received power adaptation sub-model corresponding to each node to the decision center 00 so that the decision center 00 will use it in the decision-making process. The specific method is not limited here.

[0223] In other embodiments, the amount of data corresponding to each power adaptation sub-model is large. In order to save transmission resources, multiple groups of input-output data pairs (i.e., key points) related to the neural network model (i.e., the determined power adaptation sub-model) can be selected and transmitted to the decision center 00. The decision center 00 reconstructs the required power adaptation function based on the key points. Specifically, the node can send first adaptation processing information to the decision center 00. The first adaptation processing information includes multiple groups of input-output data pairs, each group of input-output data pairs corresponding to the power adaptation sub-model of a node, so that the decision center 00 can recover the required power adaptation sub-model.

[0224] Figure 6A shows a schematic diagram of a process of transmitting a power adaptation sub-model through key point information according to some embodiments of the present application, wherein, in the decentralized distribution system shown in Figure 4A, node 0 transmits the key point data of the power adaptation sub-model corresponding to each node to the decision center 00, and then the decision center 00 restores each power adaptation sub-model according to the received key point information to obtain the power adaptation sub-model corresponding to each node for introduction as an example.

[0225] S301, Node 0 determines the key point information corresponding to the input-output of the power adaptation sub-model of each node.

[0226] In some embodiments, it is necessary to determine the number of key points to be transmitted. According to the preset number of key points and the trained power adaptation sub-model, the key point information of the preset number is obtained.

[0227] Specifically, preset the number of key points to be transmitted, obtain the curves corresponding to the power adaptation sub-models of each node, and according to the curves corresponding to the trained power adaptation sub-models, obtain the key point coordinates (x i , y i ) of the corresponding input-output of the preset number of each node, and transmit them as key point information.

[0228] It can be understood that the following formula (7) shows a general model of a power adaptation sub-model as: y = f j (x; θ) Formula (7)

[0229] Among them, y is the output data after the action of the power adaptation sub-model f j , and x is the input data to be adapted. j represents the node number corresponding to the current power adaptation sub-model, and the parameter θ is the parameter of the neural network model. At this time, select the preset number of key points to obtain the key point coordinates (x i , y i ) of the preset number. i represents the subscript of the key point coordinates, and send them to the decision center 00.

[0230] For example, for the power adaptation sub-model of Node 0 as f0, the power adaptation sub-model f0 is processed by a non-linear function, and the power adaptation sub-model will increase the data far from the classification boundary and decrease the data close to the classification boundary, and the size order of each data after adjustment is the same as that before adjustment.

[0231] FIG. 6B shows a schematic diagram of a curve corresponding to the power adaptation sub-model f0 according to an embodiment of the present application. The horizontal asymptote ordinates of the input-output curve of the current power adaptation sub-model f0 are m and M (m < M) respectively, that is, m is the minimum value of the output value, and M is the maximum value of the output value.

[0232] Taking the preset number of key points as 15 as an example, 15 key point coordinates corresponding to the input-output data of the power adaptation sub-model f0 can be obtained as (x1, y1), (x2, y2), (x3, y3),......, (x 15 , y 15 ). It can be understood that 15 key points can be selected by any existing method, such as equidistant sampling of the abscissa, equidistant sampling of the ordinate, etc., which are not required here.

[0233] In some implementations, the selection can be performed in the following manner. For example, 15 key points are selected at equal intervals within the range [m, M] on the vertical axis. FIG6B shows the 15 key points (x1, y1), (x2, y2), (x3, y3), ..., (x 15 ,y 15 Specifically, since the differences between the output values ​​are the same, the following formula (8) shows the expression of the vertical coordinate yi corresponding to the 15 key points,

[0234] Among them, i = 1, 2, ..., 15, corresponding to the number of each key point.

[0235] The following formula (9) shows the horizontal coordinates x1,…,x1 of each of the 15 key points obtained according to the above formula (8): 15 The expression:

[0236] S302: Node 0 sends first adaptation processing information to decision center 00, where the first adaptation processing information includes key point information of the power adaptation sub-model corresponding to each node.

[0237] In some embodiments, node 0 sends multiple sets of key point coordinates to decision center 00, each set corresponding to a power adaptation sub-model, and each set includes coordinates of 15 key points. For example, as shown in Figure 6C, node j sends key point information of each power adaptation sub-model to decision center 00. When the node is 0, it means that node 0 sends key point information of each node's power adaptation sub-model to decision center 00.

[0238] S303: The decision center 00 reconstructs the power adaptation sub-model according to the key point information to obtain restored power adaptation sub-models.

[0239] In some embodiments, for each power adaptation sub-model to be restored, a parameter-containing model is selected, for example, a parameter-containing model is any one of an exponential model, a Tanh function model, an Arctan function model, and a power series model, and each parameter-containing model is solved according to key point information to obtain a power adaptation function, and the solved power adaptation function is used as the restored power adaptation sub-model.

[0240] (1) Obtain a parameter-containing model.

[0241] The following is an introduction to some specific parameter-containing models.

[0242] Specifically, the following formulas (10) to (13) respectively show the expressions of the exponential model, the Tanh function model, the Arctan function model, and the power series model:

[0243] ①. Exponential Model

[0244] ②.Tanh function model

[0245] ③.Arctan function model

[0246] ④. Power Series Model

[0247] Among them, a in the above formulas (10) to (13) is l ,b l ,c l are all parameters to be determined, and l is the parameter subscript, the value of m can represent the number of parameters, and m is a positive integer.

[0248] For example, an exponential model can be selected as a parameter-containing model for subsequent calculations.

[0249] (2) According to the received preset number of key point information and the preset parameter model, a power adaptation function is obtained, and the obtained power adaptation function is used as the restored power adaptation sub-model.

[0250] The received key point value (x i ,y i ),i=1,…,e (the number e depends on the number of parameters of the parameter-containing model) is substituted into the parameter-containing model, and the equation group is solved to obtain the parameters to be determined in the parameter-containing model.

[0251] For example, taking the exponential model as an example, the following formula (14) shows an exponential parameter model including 15 parameters to be determined,

[0252] At this time, according to the coordinates of 15 key points (x1, y1), (x2, y2), (x3, y3), ..., (x 15 ,y 15 ) into the above formula (14), that is, referring to the following formula (15), we can get a set of a l ,b l ,c l The feasible solution of is obtained, thereby obtaining the restored power adaptation function.

[0253] In addition, combined with the above formula (8), formula (15) can be transformed into the following formula (16). In this case, it is only necessary to change the coordinates x of the 15 key points to i Substitute into the following formula (16):

[0254] The parameter a can be obtained l ,b l ,c l A set of feasible solutions. Substitute this set of feasible solutions into the parameterized model f(x; a l ,b l ,c l ) is the power adaptation function of the receiving end.

[0255] It can be understood that the execution order of the above steps S301 to S303 is only an illustration. In other embodiments, other execution orders may be adopted, and some steps may be split or combined, which is not limited here.

[0256] It can be understood that in the embodiments of the present application, when the power adaptation sub-model needs to be synchronized between the node and the decision center, only a small amount of data needs to be transmitted to achieve the synchronization of the power adaptation sub-model. Especially in the situation where the number of nodes is large and the amount of data of the power adaptation sub-model is large, the power adaptation function sub-model can be effectively synchronized by transmitting relatively few curve key points, thereby overcoming the problem of large and complex amount of data to be transmitted, and also improving transmission efficiency and saving transmission resources.

[0257] (2) Centralized Distributed System

[0258] In the case where data exchange between nodes is not possible, only the decision center can obtain the characteristic data of all nodes. Neural network model training is required on the decision center to determine the required power adapter model.

[0259] Figure 7 shows a schematic diagram 700 of a centralized distributed system framework according to some embodiments of the present application. Figure 7 still uses the two nodes and one decision center shown in Figure 3A as an example to illustrate a schematic diagram of the framework related to the training process, thereby facilitating comparison with the schematic diagram of the framework of the actual use process after training shown in Figure 3A.

[0260] As shown in Figure 7 , node 0 includes a feature extraction module 301b and a power adaptation module 302b. Node 1 includes a feature extraction module 311b and a power adaptation module 312b. It is understood that the features extraction module 301b and the power adaptation module 302b, as well as the features extraction module 311b and the power adaptation module 312b, have essentially the same functions as the features extraction module 301a and the power adaptation module 302a, as well as the features extraction module 311a and the power adaptation module 312a shown in Figure 4A , and are not further described here.

[0261] Specifically, during the training process, for node 0, the label feature data x is obtained at node 0 and node 1 through the feature extraction module 301b and the feature extraction module 311b. k；0 and label feature data x k；1 After that, the second feature information will be sent to the decision center 00 respectively. The second feature information includes the feature data of each sample in the sample set of each node and the classification label of each sample. It can be understood that this method is the default method of sending without using channel transmission.

[0262] Then, the decision center 00 trains the preset neural network model based on the second feature information received from all nodes, for example, the label feature data in the second feature information, and the statistical characteristics of the channel noise observed between each node and the decision center 00, such as the noise variance. The training method is the same as the specific training process of step S205 shown in Figure 5A above, which will not be repeated here. Thus, the neural network model corresponding to each node after training is obtained, thereby obtaining the power adaptation sub-model corresponding to each node. The power adaptation sub-model corresponding to each node is then synchronized to each node. The sending process is essentially the same as the process of synchronizing the power adaptation sub-model from the corresponding node 0 to the decision center 00 under the above-mentioned decentralized system. The specific process can be referred to the description of Figure 6A, which will not be repeated here.

[0263] It can be understood that since the decision center 00 can obtain the label feature data of all nodes, the decision center can obtain the statistical characteristics of the label feature data corresponding to all nodes, for example, the mean vector μ k , covariance∑ k , in order to make decisions in the actual use process. It can be understood that in the above step S111, for the centralized system, since the decision center 00 can obtain the label feature data of all nodes, the label feature data is directly used when calculating the likelihood function value, and there is no need to use the mean vector μ k , covariance Σ k The simulated data makes the decision-making more effective.

[0264] The following describes experimental data demonstrating the effectiveness of the data transmission method used in the embodiments of this application for the classification of images passing through a noisy channel. The results obtained using the modified National Institute of Standards and Technology (MNIST) dataset (referred to as the MNIST dataset) and the Canadian Institute for Advanced Research-10 (CIFAR-10) dataset (referred to as the CIFAR-10 dataset) are presented.

[0265] (1) Experimental introduction using the MNIST dataset:

[0266] The MNIST dataset primarily includes handwritten digit recognition. It contains 10 classes (i.e., 10 labels) for the digits 0 to 9. For example, consider the digits 0 and 1, which have two labels. The dataset contains 50,000 28x28 black-and-white images, used for offline training of feature extractors and power adaptation functions, as well as for evaluating receiver performance in actual compression and transmission tests.

[0267] 2. LeNet-5 is used as a feature extractor to extract one-dimensional feature data from the image. The neural network structure of LeNet-5 is shown in Figure 8A. The basic structure of LeNet-5 consists of a seven-layer network structure (excluding the input layer), including two convolutional layers, two downsampling layers (pooling layers), two fully connected layers, and an output layer.

[0268] 3. Linear enhancement using a single linear layer (as a comparison baseline); nonlinear power adaptation using a fully connected neural network with an activation layer. The neural network structure is shown in Figure 8B , where the input is one-dimensional data, the output is one-dimensional data, and it includes at least one hidden layer.

[0269] 4. Based on the mean and variance of the extracted features, generate data that obeys the normal distribution and train the power adapter model.

[0270] The performance of the receiving end binary classification decision-making using the three methods is compared in scenarios with three different channel rated power combinations:

[0271] 1) Using a nonlinear power adaptation sub-model to combat noise (i.e., the method proposed in the embodiments of the present application);

[0272] 2) Using linear enhancement to combat noise;

[0273] 3) A method of performing classification at the transmitting end and transmitting the result using BPSK.

[0274] The experimental results are shown in Table 1 below (the vertical columns in the table correspond from top to bottom to the classification accuracy results obtained using linear enhancement, BPSK and nonlinear power adaptation methods; the horizontal columns are the upper limits of the transmission power from the two nodes to the decision center; the numbers are classification accuracy).

[0275] Referring to the classification results corresponding to Table 1, the first and second rows are the control group, which respectively use the traditional linear filtering method and the method of judging the label at the transmitting end and performing BPSK transmission. The third and fourth rows are the experimental groups for the method proposed in the embodiment of the present application, corresponding to the decentralized system and the centralized system respectively. The last row is the high-traffic control group. Assuming that there are sufficient communication conditions, the power adapter model is directly transmitted to the receiving end / transmitting end, which can prove the reliability of saving the communication cost of the adaptation function. Through the classification accuracy results shown in Table 1, it can be seen that under different rated powers, the classification accuracy results obtained by the embodiment scheme of the present application (such as the decentralized system, centralized system and power adapter model lossless transmission in Table 1) are better than the accuracy of the scheme using BPSK and linear filtering. When the power adapter model is synchronized at both ends, if the power adapter model is directly synchronized, the classification result obtained at this time (the power adapter model lossless transmission in the 5th row of Table 1) will be better than the classification result obtained by transmitting key point information (the decentralized system in the 3rd row of Table 1 and the centralized system in the 4th row).

[0276] Table 1

[0277] (2) Using the CIFAR-10 dataset

[0278] 1. The CIFAR-10 dataset is primarily an object recognition dataset. For example, consider the two categories ("airplane" and "bird"), with two labels. Each category contains 1,000 32x32 color (three-channel) images, which are used for offline training of feature extractors and power adapter models, as well as for evaluating receiver performance in actual compression and transmission tests.

[0279] 2. Use LeNet-5 as the feature extractor to extract one-dimensional feature data of the image (the network structure is the same as before).

[0280] 3. Use a single linear layer for linear enhancement (as a comparison baseline); use a fully connected neural network with an activation layer for nonlinear power adaptation.

[0281] 4. Based on the mean and variance of the extracted feature data, generate data that obeys the normal distribution and train the power adaptation sub-model.

[0282] The performance of the receiving end binary classification decision-making using the three methods is compared in scenarios with three different channel rated power combinations:

[0283] 1) Using a nonlinear power adaptation sub-model to combat noise (i.e., the method proposed in the embodiments of the present application);

[0284] 2) Using linear enhancement to combat noise;

[0285] 3) A method of performing classification at the transmitting end and transmitting the result using BPSK.

[0286] The experimental results are shown in the following table (the vertical columns in the table correspond to the use of linear enhancement, BPSK and nonlinear power adaptation from top to bottom; the horizontal columns are the upper limits of the transmission power from the two nodes to the decision center; the numbers are classification accuracy). The experimental results are shown in Table 2. Specifically, the first and second rows are the control groups, which respectively use the traditional linear filtering method and the method of judging the label at the transmitting end and performing BPSK transmission. The third and fourth rows are the experimental groups, corresponding to the decentralized system and the centralized system respectively. The last row is the high-traffic control group. Assuming that there are sufficient communication conditions, the power adaptation sub-model is directly passed to the receiving end / transmitting end, which can prove the reliability of saving communication costs. From the classification accuracy results shown in Table 2, it can be seen that under different rated powers, the classification accuracy results obtained corresponding to the embodiment scheme of the present application (such as the decentralized system, centralized system and lossless transmission of the adaptation function in Table 2) are better than the accuracy of the scheme using BPSK and linear filtering. When synchronizing the power adapter model at both ends, if the power adapter model is synchronized directly, the classification result obtained at this time (lossless transmission of the power adapter model in the 5th row of Table 2) will be better than the classification result obtained by transmitting key point information (the decentralized system in the 3rd row of Table 2 and the centralized system in the 4th row).

[0287] Table 2

[0288] Figure 9 is a hardware block diagram of an electronic device according to an embodiment of the present application. As shown in Figure 9, the electronic device 10 includes a processor 110, a communication module 120, a screen 130, an interface module 140, a memory 150, a power module 160, an audio module 170, a camera 180, and a sensor module 190, wherein the audio module 170 includes a speaker 171 and a microphone 172.

[0289] Wherein: the processor 110 may include one or more processing units, for example, a processing module or processing circuit such as a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a microprocessor (MCU), an artificial intelligence (AI) processor or a programmable logic device (field programmable gate array, FPGA). Wherein, different processing units may be independent devices or integrated into one or more processors. In some embodiments, the processor 110 may extract features from the data, adjust the feature data before sending it, or perform decision processing by executing a program related to the data processing method of the present application.

[0290] The communication module 120 may include various wired or wireless communication modules, such as a Bluetooth module (BT), a wireless local area network (WLAN) module, etc., for providing wireless fidelity (Wi-Fi), Bluetooth (BT), a global navigation satellite system (GNSS), frequency modulation (FM), near field communication (NFC), infrared technology (IR), a wide area network (WAN), and other wired or wireless communication solutions. In some embodiments, the processor 110 may execute a program related to the data processing method of the present application.

[0291] The screen 130 includes a display panel. The display panel can be a liquid crystal display (LCD), an organic light-emitting diode (OLED), an active-matrix organic light-emitting diode (AMOLED), a flexible light-emitting diode (FLED), a Mini LED, a Micro LED, a Micro OLED, or a quantum dot light-emitting diode (QLED).

[0292] The interface module 140 may include various forms of input or output interfaces. The electronic device 10 may transmit video and / or audio data to other electronic devices through the output interface, and receive video and / or audio data from other electronic devices through the input interface. In some embodiments, the input and output interfaces may include: a Sony / Philips digital interface format (S / PDIF) interface, a high-definition multimedia interface (HDMI), a local area network (LAN) interface, a universal serial bus (USB) interface, an audio-video (AV) interface, etc.

[0293] The memory 150 can be used to store data, software programs, and modules, and can be a volatile memory (volatile memory), such as random-access memory (RAM); or a non-volatile memory (non-volatile memory), such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD); or a combination of the above types of memory, or it can be a removable storage medium, such as a secure digital (SD) memory card. Specifically, in some embodiments of the present application, the memory can be used to store relevant instructions for executing the data processing method of the present application.

[0294] The power module 160 may include a power button, an IR receiver, etc., for turning on or off the power of the electronic device 10 according to user operations.

[0295] The audio module 170 can convert digital audio signals into analog audio signals for output, or convert analog audio input into digital audio signals, and can also transmit digital audio signals and / or analog audio signals to other electronic devices through the interface module 140. In some embodiments, the audio module 170 may include a speaker 171 and a microphone 172.

[0296] Camera 180 is used to capture still images or videos. The optical image of a scene, generated by the lens, is projected onto the image sensor surface. This image is then converted into an electrical signal, which undergoes analog-to-digital conversion (A / D) to become a digital image signal. This signal is then processed by a digital signal processing chip.

[0297] The sensor module 190 may include a magnetic sensor, an acceleration sensor, a temperature sensor, a voice sensor, and the like.

[0298] It can be understood that the structure of the electronic device 10 shown in Figure 3 is only an example. In other embodiments, the electronic device 10 may also include more or fewer modules, and some modules may be combined or split, which is not limited in the embodiments of the present application.

[0299] It is understandable that the electronic device 10 can be a specific hardware structure diagram of a node, or a specific hardware structure diagram of a decision center, which is not required here. When the electronic device is a node, the operations performed by the node in the data processing method proposed in the embodiment of the present application can be executed. When the electronic device is a decision center, the operations performed by the decision center in the data processing method proposed in the embodiment of the present application can be executed. It is understandable that the structure illustrated in the embodiment of the present invention does not constitute a specific limitation on the electronic device 10.

[0300] The various embodiments of the mechanisms disclosed in this application can be implemented in hardware, software, firmware, or a combination of these implementation methods. The embodiments of the present application can be implemented as a computer program or program code executed on a programmable system, which includes at least one processor, a storage system (including volatile and non-volatile memory and / or storage elements), at least one input device, and at least one output device.

[0301] Program code can be applied to input instructions to perform the functions described herein and generate output information. The output information can be applied to one or more output devices in a known manner. For purposes of this application, a processing system includes any system having a processor such as, for example, a digital signal processor, a microcontroller, an application-specific integrated circuit (ASIC), or a microprocessor.

[0302] Program code can be implemented with a high-level programming language or an object-oriented programming language to communicate with the processing system. Where necessary, program code can also be implemented in assembly language or machine language. In fact, the mechanism described in this application is not limited to the scope of any particular programming language. In either case, the language can be a compiled language or an interpreted language.

[0303] In some cases, the disclosed embodiments can be implemented in hardware, firmware, software or any combination thereof. The disclosed embodiments can also be implemented as instructions carried or stored on one or more temporary or non-temporary machine-readable (e.g., computer-readable) storage media, which can be read and executed by one or more processors. For example, instructions can be distributed over a network or through other computer-readable media. Therefore, machine-readable media may include any mechanism for storing or transmitting information in a machine (e.g., computer) readable form, including but not limited to, floppy disks, optical disks, optical disks, read-only memories (compact disc read-only memories, CD-ROMs), magneto-optical disks, read-only memories (ROMs), random access memories (RAMs), erasable programmable read-only memories (EPROMs), electrically erasable programmable read-only memories (EEPROMs), magnetic or optical cards, flash memory, or tangible machine-readable memories for transmitting information (e.g., carrier waves, infrared signal digital signals, etc.) using the Internet in electrical, optical, acoustic or other forms of propagation signals. Accordingly, machine-readable media includes any type of machine-readable media suitable for storing or transmitting electronic instructions or information in a form readable by a machine (eg, a computer).

[0304] In the accompanying drawings, some structural or method features may be shown in a particular arrangement and / or order. However, it should be understood that such a particular arrangement and / or order may not be required. Rather, in some embodiments, these features may be arranged in a manner and / or order different from that shown in the illustrative drawings. In addition, the inclusion of a structural or method feature in a particular figure does not imply that such feature is required in all embodiments, and in some embodiments, such features may not be included or may be combined with other features.

[0305] It should be noted that the units / modules mentioned in the various device embodiments of the present application are all logical units / modules. Physically, a logical unit / module can be a physical unit / module, or a part of a physical unit / module, or can be implemented as a combination of multiple physical units / modules. The physical implementation of these logical units / modules themselves is not the most important. The combination of functions implemented by these logical units / modules is the key to solving the technical problems raised by this application. In addition, in order to highlight the innovative part of this application, the above-mentioned device embodiments of this application do not introduce units / modules that are not closely related to solving the technical problems raised by this application. This does not mean that other units / modules do not exist in the above-mentioned device embodiments.

[0306] It should be noted that in the examples and description of this patent, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "including a" does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0307] Although the present application has been shown and described with reference to certain preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the application.

Claims

1. A data processing method, characterized in that: include: A first node device in a distributed system receives first noise statistical information sent by a decision device in the distributed system, and receives first feature information respectively sent by node devices other than the first node device in the distributed system; wherein the first noise statistical information includes statistical characteristics of noise in a channel between the node device and the decision device, and the first feature information includes feature data of each sample in a first sample set of the node device, and a classification label of each sample, and The first noise statistical information and the first characteristic information are used to determine a first power adaptation sub-model used by each node device to perform a first task.

2. The method according to claim 1, characterized in that: The first power adaptation sub-model used by each node device to perform the first task is determined in the following manner: The first node device trains the first neural network model based on the first noise statistical information and the first feature information corresponding to each node device to obtain a trained first neural network model, wherein the trained first neural network model includes the first power adapter sub-model corresponding to each of the node devices.

3. The method according to claim 1, characterized in that The statistical characteristics include one or more of the following: Mean vector, covariance, arithmetic mean, geometric mean, harmonic mean, weighted mean, square mean, median, median number, mode, mean absolute deviation, variance.

4. The method according to any one of claims 1 to 3, characterized in that It also includes that the first node device sends first adaptation processing information to the decision device, wherein the first adaptation processing information includes multiple groups of first input-output data pairs, and each group of first input-output data pairs corresponds to the first power adaptation sub-model of a node device.

5. The method according to claim 4, characterized in that Also includes: The decision device obtains a first power adaptation function corresponding to the first power adaptation sub-model of each node device based on the received first adaptation processing information.

6. The method according to claim 5, characterized in that The decision device obtains, based on the received first adaptation processing information, a first power adaptation function corresponding to the first power adaptation sub-model of each node device, including: Determine a first power adaptation function corresponding to the first power adaptation sub-model of the first node device in the following manner: The decision device inputs the first power adaptation sub-model of the first node device into the first parameter-containing function model corresponding to the first input-output data pair, and obtains a first solution value of each parameter in the first parameter-containing function model; The decision-making device replaces the parameters in the first parameter-containing function model with the corresponding first solution values ​​to obtain the first power adaptation function.

7. The method according to claim 6, characterized in that After the first node device receives the first feature information respectively sent by the node devices other than the first node device in the distributed system, the method further includes: The first node device sends first label feature statistical information to the decision device, wherein the first label feature statistical information includes statistical characteristics corresponding to each classification label in a plurality of classification labels, wherein the statistical characteristics corresponding to a first classification label in the plurality of classification labels are statistical characteristics of feature data whose classification label is the first classification label in all feature data corresponding to the sample set of all node devices in the distributed system.

8. The method according to claim 7, characterized in that Also includes: Each of the node devices acquires first data to be classified corresponding to the first classification object in the first task; Each of the node devices processes the first classification feature data of the first data to be classified into second classification feature data based on its own first power adaptation sub-model, and sends the second classification feature data to the decision-making device.

9. The method according to claim 8, characterized in that It also includes the decision device obtaining a classification result for the first classification object based on the first power adaptation function corresponding to the first power adaptation sub-model of each node device, the third classification feature data corresponding to the second classification feature data sent by each node device, and the statistical characteristics corresponding to each classification label in multiple classification labels.

10. The method according to claim 9, characterized in that The decision device obtains a classification result for the first classification object based on the first power adaptation function corresponding to the first power adaptation sub-model of each node device, the third classification feature data corresponding to the second classification feature data sent by each node device, and the statistical characteristics corresponding to each classification label in the multiple classification labels, including: The decision device determines the power of each node device based on the first power adaptation function corresponding to each node device and the second classification feature sent by each node device. The likelihood function values ​​corresponding to the first classification object and each classification label are determined based on the third classification feature data corresponding to the data and the statistical characteristics corresponding to each classification label in the plurality of classification labels; The classification category of the classification label corresponding to the maximum likelihood function value among the likelihood function values ​​is used as the classification result of the first classification object.

11. The method according to any one of claims 1 to 3, characterized in that: It also includes that the first node device sends second adaptation processing information to the decision device, wherein the second adaptation processing information includes a first power adaptation sub-model for each node device to perform a first task.

12. A data processing method, characterized in that: include: A decision-making device in a distributed system receives second characteristic information sent by each node device in the distributed system; The second feature information includes feature data of each sample in the second sample set of the node device and a classification label of each sample; and The second characteristic information is used to determine a second power adaptation sub-model used by each node device to perform a second task.

13. The method according to claim 12, characterized in that The second power adaptation sub-model used by each node device to perform the second task is determined in the following manner: The decision device trains the second neural network model based on the statistical characteristics of the noise of the channel between each of the node devices and the second feature information corresponding to each node device to obtain a trained second neural network model, wherein the trained second neural network model includes the second power adaptation sub-model corresponding to each node device.

14. The method according to any one of claims 13, characterized in that The statistical characteristics include one or more of the following: Mean vector, covariance, arithmetic mean, geometric mean, harmonic mean, weighted mean, square mean, median, median number, mode, mean absolute deviation, variance.

15. The method according to any one of claims 12 to 14, characterized in that: It also includes that the decision device sends third adaptation processing information to each of the node devices respectively, wherein the third adaptation processing information includes a group of second input-output data pairs, and each group of second input-output data pairs corresponds to the second power adaptation sub-model of the sending node device.

16. The method according to claim 15, characterized in that Also includes: Each node device obtains a second power adaptation function corresponding to the second power adaptation sub-model of the node device based on the received third adaptation processing information.

17. The method according to claim 16, characterized in that Each node device obtains a second power adaptation function corresponding to the second power adaptation sub-model of the node device based on the received third adaptation processing information, including: The first node device obtains a second power adaptation function corresponding to the second power adaptation sub-model of the first node device based on the received third adaptation processing information: The first node device inputs the second input-output data pair into a second parameter-containing function model to obtain a second solution value of each parameter in the second parameter-containing function model; The first node device replaces the parameters in the second parameter-containing function model with the corresponding second solved values ​​to obtain the second power adaptation function.

18. The method according to claim 17, characterized in that Also includes: Each of the node devices acquires second data to be classified corresponding to the second classification object of the second task; Each of the node devices processes the fourth classification feature data of the second data to be classified into fifth classification feature data based on its own second power adaptation function, and sends the fifth classification feature data to the decision device.

19. The method according to claim 18, characterized in that It also includes that the decision device obtains a classification result for the second classification object based on the second power adaptation sub-model corresponding to each node device, the sixth classification feature data corresponding to the fifth classification feature data sent by each node device, and the feature data of each sample in the second sample set of all the node devices.

20. The method according to claim 19, characterized in that The decision device obtains a classification result for the second classification object based on the second power adaptation sub-model corresponding to each node device, the sixth classification feature data corresponding to the fifth classification feature data sent by each node device, and the feature data of each sample in the second sample set of all the node devices, including: The decision device determines the likelihood function values ​​corresponding to the second classification object and each classification label respectively based on the second power adaptation sub-model corresponding to each node device, the sixth classification feature data corresponding to the fifth classification feature data sent by each node device, and the feature data of each sample in the second sample set of all the node devices; The classification category of the classification label corresponding to the maximum likelihood function value among the likelihood function values ​​is used as the classification result of the second classification object.

21. The method according to any one of claims 12 to 14, characterized in that The method further includes: the decision device sends fourth adaptation processing information to each of the node devices respectively, wherein the fourth adaptation processing information includes a second power adaptation sub-model corresponding to the node device.

22. An electronic device, characterized in that: include: one or more processors; One or more memories; the one or more memories store one or more instructions, and when the one or more instructions are executed by the one or more processors, the electronic device executes the data processing method described in any one of claims 1 to 21.

23. A computer-readable storage medium, characterized in that: The storage medium stores instructions, which, when executed on a computer, cause the computer to execute the data processing method according to any one of claims 1 to 21.

Images