Data processing method and apparatus

Abstract

The present application relates to the field of computers. Provided are a data processing method and apparatus, which are used for improving the efficiency of data transmission between an edge device and an edge server when the edge device sends feature data of an AI model to the edge server. The method comprises: acquiring a first bit error rate, wherein the first bit error rate is used for indicating a bit error rate that corresponds to feature data of an AI model when an output result of the AI model meets a preset accuracy rate; on the basis of the first bit error rate, determining from among a plurality of transmission modes a first transmission mode that corresponds to the first bit error rate; and using the first transmission mode to send the feature data of the AI model to an edge server.

Description

Data processing method and apparatus

[0001] The present application claims priority to the Chinese patent application No. 202410472488.1, filed on April 18, 2024, and entitled "Data processing method and apparatus", the content of which is incorporated herein by reference in its entirety. TECHNICAL FIELD

[0002] The present application relates to the field of computer, and in particular, to a data processing method and apparatus. BACKGROUND

[0003] At present, with the increasing computing power of edge devices, more and more artificial intelligence (AI) products begin to try to realize through edge computing. For example, FIG. 1 shows a structural schematic diagram of a communication network of edge computing. In the communication network 10, edge devices and edge servers are included, and edge device 101 and edge server 102 are taken as examples in FIG. 1. The edge device 101 is used to collect raw data and extract features from the raw data to obtain feature data, and the edge device 101 is also used to send the feature data to the edge server 102. The edge server 102 is used to output results using the feature data from the edge device 101.

[0004] Compared with the way of deploying the feature extraction process and the process of outputting results using feature data on the same type of devices (such as cloud devices or data centers) in traditional AI products, on the one hand, through the edge computing mode, the intelligent service can be closer to the data source and the end user, thereby realizing more intelligent service; on the other hand, through the edge computing mode, the dependence on cloud devices or data centers can be reduced, the efficiency and real-time performance of data processing can be provided, and the security and privacy of data can also be improved.

[0005] In the process of realizing AI functions by using edge computing mode, how to improve the data transmission efficiency between edge devices and edge servers is a problem to be solved at present. SUMMARY

[0006] The present application provides a data processing method and apparatus for an edge device to send feature data to an edge server.

[0007] In a first aspect, a data processing method is provided, which is applied to an edge device and includes: obtaining a first bit error rate, the first bit error rate being used to indicate a bit error rate of feature data of an AI model in a case where an output result of the AI model meets a preset accuracy. According to the first bit error rate, a first transmission mode corresponding to the first bit error rate is determined from a plurality of transmission modes. The feature data of the AI model is transmitted to an edge server by using the first transmission mode.

[0008] By the above method provided by the embodiments of the present application, on the one hand, when the bit error rate in the feature data transmission process is high, a transmission mode with higher reliability can be selected to transmit the feature data, so as to reduce the bit error rate in the feature data transmission process (i.e., to ensure that the bit error rate of the feature data received by the edge server is lower than the first bit error rate), so that the output result of the AI model meets the preset accuracy. On the other hand, when the bit error rate in the feature data transmission process is low, a transmission mode with relatively low reliability but low redundancy data ratio can be selected to transmit the feature data, so as to reduce the transmission delay and improve the transmission efficiency on the premise of reducing the bit error rate in the feature data transmission process (i.e., to ensure that the bit error rate of the feature data received by the edge server is lower than the first bit error rate).

[0009] In an implementation manner, different transmission modes in the plurality of transmission modes correspond to different channel coding lengths.

[0010] By the above implementation manner, different transmission modes with different channel coding lengths can be selected to transmit the feature data according to the first bit error rate, so as to reduce the overhead for ensuring transmission reliability on the premise of ensuring the accuracy of the decision result of the AI model, thereby achieving the effects of improving transmission efficiency and reducing communication delay.

[0011] In an implementation manner, different transmission modes in the plurality of transmission modes correspond to different retransmission times.

[0012] By the above implementation manner, different transmission modes with different retransmission times can be selected to transmit the feature data according to the first bit error rate, so as to reduce the overhead for ensuring transmission reliability on the premise of ensuring the accuracy of the decision result of the AI model, thereby achieving the effects of improving transmission efficiency and reducing communication delay.

[0013] In an implementation manner, determining, according to the first bit error rate, a first transmission mode corresponding to the first bit error rate from a plurality of transmission modes includes: determining, according to the first bit error rate, a second bit error rate, the second bit error rate being used to reflect a difference between an actual bit error rate of a transmission channel between the edge device and the edge server and the first bit error rate. A first transmission mode corresponding to the second bit error rate is determined from the plurality of transmission modes.

[0014] By the implementation manner, the transmission manner is determined according to the difference (i.e., the second bit error rate) between the actual bit error rate of the transmission channel and the first bit error rate, so that the overhead for ensuring transmission reliability is reduced while ensuring the accuracy of the decision result of the AI model, thereby achieving the effects of improving transmission efficiency and reducing communication delay.

[0015] In an implementation manner, the first bit error rate is obtained by determining the first bit error rate according to the noise variance corresponding to the feature data of the AI model in a case where the output result of the AI model meets the preset accuracy rate.

[0016] By the implementation manner, the first bit error rate is determined according to the noise variance corresponding to the feature data of the AI model in a case where the output result of the AI model meets the preset accuracy rate, so that the overhead for ensuring transmission reliability is reduced while ensuring the accuracy of the decision result of the AI model, thereby achieving the effects of improving transmission efficiency and reducing communication delay.

[0017] In an implementation manner, the method further includes: performing feature extraction on the original data to obtain the feature data.

[0018] In an implementation manner, the edge server is configured to perform decision inference according to the feature data from the edge device and the AI model.

[0019] In a second aspect, a data processing apparatus is provided, which is applied to an edge device and includes: an obtaining unit configured to obtain a first bit error rate, the first bit error rate being used to indicate a bit error rate corresponding to feature data of an AI model in a case where an output result of the AI model meets a preset accuracy rate; a processing unit configured to determine a first transmission manner corresponding to the first bit error rate from a plurality of transmission manners according to the first bit error rate; and a communication unit configured to send the feature data of the AI model to an edge server by using the first transmission manner.

[0020] In an implementation manner, different transmission manners in the plurality of transmission manners correspond to different channel coding lengths.

[0021] In an implementation manner, different transmission manners in the plurality of transmission manners correspond to different retransmission times.

[0022] In an implementation manner, the processing unit is configured to determine the first transmission manner corresponding to the first bit error rate from the plurality of transmission manners according to the first bit error rate, including: the processing unit is configured to determine a second bit error rate according to the first bit error rate, the second bit error rate being used to reflect a difference between an actual bit error rate of a transmission channel between the edge device and the edge server and the first bit error rate; and the processing unit is configured to determine the first transmission manner corresponding to the second bit error rate from the plurality of transmission manners.

[0023] In an implementation manner, the obtaining unit, configured to obtain the first bit error rate, comprises: an obtaining unit configured to determine the first bit error rate according to a noise variance corresponding to feature data of the AI model in a case that an output result of the AI model meets a preset accuracy rate.

[0024] In an implementation manner, the processing unit is further configured to perform feature extraction on the original data to obtain the feature data.

[0025] In an implementation manner, the edge server is configured to perform decision inference according to the feature data from the edge device and the AI model.

[0026] In a third aspect, a data processing apparatus is provided, comprising a processor and an interface, the processor being configured to receive or send data through the interface, and the processor being configured to implement the method according to the first aspect or any implementation manner of the first aspect.

[0027] In a fourth aspect, a computer-readable storage medium is provided, and the computer-readable storage medium stores instructions, when the instructions are executed on a processor, the method according to the first aspect or any implementation manner of the first aspect is implemented.

[0028] In a fifth aspect, a computer program product is provided, and the computer program product comprises instructions, when the instructions are executed on a processor, the method according to the first aspect or any implementation manner of the first aspect is implemented. BRIEF DESCRIPTION OF DRAWINGS

[0029] FIG. 1 is a structural schematic diagram of a communication system according to an embodiment of the present application;

[0030] FIG. 2 is a structural schematic diagram of a communication system according to another embodiment of the present application;

[0031] FIG. 3 is a structural schematic diagram of a communication system according to another embodiment of the present application;

[0032] FIG. 4 is a schematic diagram of inference results of an AI model for dividing sample data into two categories according to an embodiment of the present application;

[0033] FIG. 5 is a schematic diagram of inference results of an AI model for dividing sample data into two categories according to another embodiment of the present application;

[0034] FIG. 6 is a schematic diagram of accuracy of a ResNet model varying with channel noise according to an embodiment of the present application;

[0035] FIG. 7 is a schematic diagram of classification of a Gaussian mixture model according to an embodiment of the present application;

[0036] FIG. 8 is a flowchart of a data processing method according to an embodiment of the present application;

[0037] Fig. 9 is a flow diagram of another embodiment of the data processing method provided by the present application;

[0038] Fig. 10 is a diagram of the relationship between the bit error rate and the channel coding length of the service data according to an embodiment of the present application;

[0039] Fig. 11 is a diagram of the frame structure of different channel coding lengths according to an embodiment of the present application;

[0040] Fig. 12 is a flow diagram of another embodiment of the data processing method provided by the present application;

[0041] Fig. 13 is a diagram of the relationship between the bit error rate and the transmission delay according to an embodiment of the present application;

[0042] Fig. 14 is a diagram of the relationship between the transmission times and the accuracy rate according to an embodiment of the present application;

[0043] Fig. 15 is a flow diagram of another embodiment of the data processing method provided by the present application;

[0044] Fig. 16 is a diagram of the relationship between the bit error rate and the classification accuracy rate according to an embodiment of the present application;

[0045] Fig. 17 is a diagram of the relationship between the bit error rate and the classification accuracy rate according to an embodiment of the present application;

[0046] Fig. 18 is a diagram of the structure of a data processing apparatus according to an embodiment of the present application;

[0047] Fig. 19 is a diagram of the structure of a data processing apparatus according to an embodiment of the present application. DETAILED DESCRIPTION

[0048] The technical solutions provided by the embodiments of the present application will be described below in conjunction with examples.

[0049] Currently, with the increasing computing power of edge devices, more and more artificial intelligence (AI) products begin to try to achieve through edge computing. For example, FIG. 1 shows a structural diagram of a communication network of edge computing. In the communication network 10, edge devices and edge servers are included, and edge device 101 and edge server 102 are taken as examples in FIG. 1. The edge device 101 is used to collect raw data and extract features from the raw data to obtain feature data, and the edge device 101 is also used to send the feature data to the edge server 102. The edge server 102 is used to output results using the feature data from the edge device 101.

[0050] For example, in the scenario of edge learning, as shown in FIG. 2, on the one hand, model training can be performed on the edge device 101 close to the data source to extract features to obtain feature data, so as to process data locally in real time and update the model; on the other hand, the edge server 102 can use a machine learning model to obtain output results according to the extracted feature data. In this way, data transmission and the load of the edge server 102 can be reduced. This helps to improve privacy, reduce latency, and implement intelligent decision-making under limited computing resources.

[0051] For another example, in the scenario of edge inference, as shown in FIG. 3, a trained model can be deployed on the edge device 101 to perform data inference locally, reduce communication with the central server, and thus improve response speed and efficiency.

[0052] As can be seen, in the process of implementing AI functions through edge computing, the AI capability of the edge device 101 is often used to extract features from raw data, and then the feature data is transmitted to the edge server 102 for processing. Therefore, in the communication process between the edge device 101 and the edge server 102, the feature data is the main content of the communication. Therefore, how to improve the feature data transmission efficiency between the edge device and the edge server is a problem to be solved at present.

[0053] In addition, in many edge communication scenarios, communication latency is a very critical indicator, which plays a decisive role in the performance experience and safety and stability of the overall system. For example, in the automatic driving scenario, the automatic driving vehicle needs real-time and high-precision communication to ensure the safety of driving. If the communication latency is too high, it may affect the reaction speed of the vehicle, thereby affecting the safety of driving. For another example, in the industrial automation scenario, high-speed and real-time communication is needed between machines to ensure the smooth progress of the production process. If the communication latency is too high, it may cause the interruption of the production process, affecting the production efficiency. For another example, in the remote medical scenario, doctors need to receive the physiological data of patients in real time through the network to make diagnosis and treatment. If the communication latency is too high, it may affect the judgment and treatment effect of doctors. For another example, in the virtual reality / augmented reality scenario, users need real-time and delay-free communication to ensure the reality of immersive experience. If the communication latency is too high, it may cause users to feel uncomfortable, affecting the use experience. For another example, in the unmanned aerial vehicle use scenario, the unmanned aerial vehicle needs real-time and high-precision communication to ensure the safety and accuracy of flight. If the communication latency is too high, it may affect the control of the unmanned aerial vehicle, and even may cause the unmanned aerial vehicle to lose control. Therefore, in the process of transmitting feature data between edge devices and edge servers, if the transmission latency can be reduced, the transmission efficiency can be improved.

[0054] To address the above technical problems, in the related art, data transmission can be performed through Ultra Reliable Low-Latency Communication (URLLC) technology. Among them, URLLC is proposed in the first version of 5G standard in 2017. Its main goal is to provide reliable and seamless communication services. With the development of technology and changes in demand, URLLC has become an important feature of 5G and future 6G communication systems, and is one of the key technologies to realize automatic driving, industrial automation, remote medical treatment, etc. On the one hand, URLLC can achieve ultra-high reliability, providing nearly 100% communication reliability and almost no data packet loss. This is very important for applications that require extremely high reliability of data transmission (such as automatic driving, industrial automation, etc.). On the other hand, URLLC can achieve ultra-low latency, with an expected delay of as low as 1 millisecond. This means that data can be sent and received in extreme time, which is very critical for applications that are sensitive to communication delay (such as remote medical treatment, virtual reality / augmented reality, etc.). Key technologies for implementing URLLC include new modulation and demodulation technologies, multiple access technologies, network slicing technologies, etc. At the same time, URLLC also requires high-quality network coverage, as well as efficient network resource management and scheduling, to meet the communication requirements of ultra-reliability and low latency. Currently, there are many methods to implement URLLC in 5G. For example, by using high-frequency resources (such as millimeter wave frequency bands) to implement URLLC. High-frequency resources can provide greater bandwidth and higher data rates, thereby improving communication reliability and reducing latency. However, at the same time, high-frequency signals have a shorter propagation distance and are more susceptible to environmental interference, so more intensive base station deployment is needed to ensure coverage. For another example, new modulation and demodulation technologies can be used to implement URLLC, including: Multiple Input Multiple Output (MIMO) technology, polarization multiplexing technology, beamforming technology, etc., which can improve the quality and stability of signals, thereby improving communication reliability and reducing latency. For another example, network slicing technology can be used to implement URLLC. Through network slicing, network resources can be allocated to specific services or applications, thereby ensuring that they receive high-quality network services and achieve low-latency and high-reliability communication. For another example, edge computing technology can be used to implement URLLC. Specifically, by placing data processing and analysis tasks on devices at the edge of the network, data transmission in the network can be reduced, thereby reducing latency. For another example, efficient resource management and scheduling techniques can be used to ensure the fairness and effectiveness of network resources, thereby improving communication reliability and reducing latency. For another example, device-level optimization, such as using more efficient processors, more advanced antenna technology, and more optimized software algorithms, can reduce device processing delay and improve communication reliability.

[0055] However, the current URLLC design is mainly to ensure the high reliability of data transmission, and thus a large amount of communication overhead can be generated, which makes the implementation process difficult in the following aspects: first, there is a conflict between reliability and low latency. Increasing the length of channel coding can reduce the probability of data packet transmission error, but will increase the data transmission delay. Second, there is a problem of large spectrum resource occupation. In the scene of low delay and high reliability in transmission, more spectrum resources can be required to be allocated, which can affect the spectrum utilization of other communication systems. Third, there is a problem of hardware limitation. Implementing URLLC requires more complex hardware and algorithm support, especially on mobile terminal devices, which can increase the cost and energy consumption.

[0056] In view of the above technical problems, in the embodiments of the present application, on the one hand, in the scene of implementing AI function by using edge computing mode, the main content of the transmission data between the edge device and the edge server is feature data. After receiving the feature data from the edge device, the edge server inputs the feature data into the AI model to make decision reasoning. On the other hand, the AI model often has a decision boundary. When the input feature data has a certain range of noise or disturbance due to transmission error, as long as it does not exceed this decision boundary, the output result of the AI model can not be affected.

[0057] For example, as shown in FIG. 4, it is an inference result schematic diagram of an AI model for dividing sample data into two categories. In the figure, the thick line represents the decision boundary. When the sample falls on the left upper side of the decision boundary, it means that the sample belongs to type 1; when the sample falls on the right lower side of the decision boundary, it means that the sample belongs to type 2. As can be seen in (a) of FIG. 4, sample a and sample b belong to type 1, and sample c and sample d belong to type 2.

[0058] In addition, when the feature data corresponding to the sample has a certain noise, as long as the noise does not exceed a certain range, it will not affect the inference result of the sample. For example, as shown in (b) of FIG. 4, when the sample data with noise does not exceed the range of the dashed line, the inference result of the sample will not be affected.

[0059] For another example, as shown in FIG. 5, it is another inference result schematic diagram of an AI model for dividing sample data into two categories. Unlike the AI model shown in FIG. 4, the AI model shown in FIG. 5 is a nonlinear model, and the decision boundary corresponding to the model is shown by the thick line in the figure. As can be seen, when the noise does not exceed the size of the margin in the figure, the inference result of the sample will not be affected.

[0060] Taking a residual neural network (ResNet) model as an example, FIG. 6 is a diagram illustrating the accuracy of a ResNet model varying with channel noise. As can be seen, when the channel noise is below a certain threshold (for example, threshold σ in the diagram), the accuracy of the model can be maintained above 90%, that is, the influence of the channel noise on the inference result of the sample can be almost ignored.

[0061] Based on the above, the embodiment of the present application provides a data processing method, which considers that when the edge device sends the feature data of the AI model to the edge server, a suitable transmission mode (hereinafter referred to as the first transmission mode) can be determined according to the bit error rate (hereinafter referred to as the first bit error rate) corresponding to the pre-set accuracy rate of the output result of the AI model, so as to send the feature data from the edge device to the edge server. In this way, unlike the process of transmitting feature data by using the URLLC technology in the related art, a large amount of communication overhead (for example, introducing a longer channel coding length; for example, introducing a multiple retransmission mechanism, etc.) is introduced in order to ensure the high reliability that each bit is not wrong in the data transmission process, the embodiment of the present application can reduce the overhead for ensuring transmission reliability while ensuring the accuracy of the decision result of the AI model, thereby achieving the effect of improving transmission efficiency and reducing communication delay.

[0062] The data processing method provided by the embodiment of the present application will be described in detail below in combination with examples:

[0063] Taking the Gaussian mixture model shown in FIG. 7 as an example, μ0 and μ1 represent two types corresponding to the Gaussian mixture model, respectively. Wherein σ Δ represents the error variance corresponding to the channel noise, σ represents the variance of the Gaussian mixture model, and σ' = σ + σ Δ .

[0064] Suppose that the transmission error of the feature data is a random variable wherein x d represents the feature value received by the edge server, and x d represents the feature value sent by the edge device. Under the discrete channel noise, the distribution probability of Δx d is shown in the following formula (1):

[0065] In addition, suppose that {X d} is a square-integrable independent and identically distributed random variable sequence, wherein E[X] = 0 and E[X 2 ] = 1. Let the sequence {ω d} satisfy the following formula (2):

[0066] In addition, the isotropic Gaussian distribution denotes the channel noise Δ wherein 0 denotes that the mean of the channel noise in each dimension is 0, denotes that the variance of the channel noise in each dimension is For example, in the Gaussian mixture model shown in FIG. 7, the model includes two dimensions (i.e., μ0and μ1), and the channel noise σ Δ The mean and variance in the two dimensions are equal and are 0 and

[0067] It can be further known that the bit error rate P b is positively correlated with , that is,

[0068] Further, in the case of implementing an AI function by using an edge computing manner, the accuracy of the AI model and the bit error rate of the feature data can be fitted into the following formula (3):

[0069] wherein, denotes the feature data including the channel noise, x denotes the original feature data, and s() denotes the output result of the AI model, denotes the probability that the result of is the same as the result of s(x).

[0070] Therefore, in the embodiments of the present application, when the edge device sends the feature data of the AI model to the edge server, a suitable transmission manner (hereinafter referred to as a first transmission manner) can be determined according to the bit error rate (hereinafter referred to as a first bit error rate) corresponding to the preset accuracy of the output result of the AI model, so as to send the feature data from the edge device to the edge server, so that the bit error rate of the feature data received by the edge server is lower than the first transmission manner, thereby ensuring the accuracy of the output result of the AI model.

[0071] Next, taking the process of sending the feature data from the edge device 101 to the edge server 102 in FIG. 1 as an example, the data processing method provided by the embodiments of the present application is introduced. As shown in FIG. 8, the method can include:

[0072] S201, the edge device 101 obtains a first bit error rate P b .

[0073] wherein the first bit error rate P b is used to indicate the bit error rate corresponding to the feature data of the AI model in the case that the output result of the AI model meets the preset accuracy.

[0074] In an implementation manner, S201 can specifically include: determining, by the edge device 101, the noise variance of the feature data of the AI model in a case where the output result of the AI model meets the preset accuracy rate determining the first bit error rate P b .

[0075] For example, taking a classifier model as the AI model, as shown in FIG. 9, the edge device 101 can determine the classification interval of the AI model according to the AI model used, and determine the requirement of classification accuracy according to the specific classification task. Further, the edge device 101 can determine the noise variance allowed in the feature data according to the classification interval of the AI model and the requirement of classification accuracy In other words, the noise variance may be understood as the noise variance of the feature data of the AI model in a case where the output result of the AI model meets the preset accuracy rate. Then, based on the corresponding relationship between the noise variance and the bit error rate, the first bit error rate P b .

[0076] In addition, in actual application, the edge server 102 can also determine the first bit error rate P b according to the process shown in FIG. 9, and send the first bit error rate P b to the edge device 101, so that the edge device 101 obtains the first bit error rate P b . The specific process of obtaining the first bit error rate P b by the edge device 101 is not limited in the embodiments of the present application.

[0077] S202, the edge device 101 determines a first transmission mode corresponding to the first bit error rate P b from a plurality of transmission modes according to the first bit error rate P b .

[0078] For example, in a case where the parameters of the channel noise of the transmission channel between the edge device 101 and the edge server are the same, on the one hand, when the first bit error rate P b is low, a transmission mode with higher reliability can be selected to transmit the feature data, so as to ensure that the bit error rate of the feature data received by the edge server 102 is lower than the first bit error rate P b , so that the output result of the AI model meets the preset accuracy rate. On the other hand, when the first bit error rate P bUnder the premise of reducing transmission delay and improving transmission efficiency.

[0079] In an implementation manner, different channel coding lengths are selected for transmission of the feature data according to the difference between the first bit error rate P b .

[0080] For example, in the case of transmitting service data through the same channel, the relationship between the bit error rate of the service data and the channel coding length is shown in FIG. 10. It can be seen that the longer the channel coding length is, the lower the bit error rate of the service data is. Therefore, different channel coding lengths are selected for transmission of the feature data according to the difference between the first bit error rate P b .

[0081] For another example, FIG. 10 shows frame structure diagrams corresponding to two channel coding lengths provided by the embodiments of the present application. As shown in FIG. 10, the frame structures shown in (a) and (b) of FIG. 11 both include 8 data blocks carrying feature data, but the frame structure of (a) of FIG. 11 includes 12 data blocks carrying redundant data, and the frame structure shown in (b) of FIG. 11 includes 3 data blocks carrying redundant data. In this case, the reliability of (a) of FIG. 11 is generally higher than that of (b) of FIG. 11, in other words, the bit error rate of the transmission mode corresponding to (a) of FIG. 11 is generally lower than that of the transmission mode corresponding to (b) of FIG. 11. Therefore, when the first bit error rate P b is low, the transmission mode corresponding to (a) of FIG. 11 is selected for transmission of the feature data; when the first bit error rate P b is high, the transmission mode corresponding to (b) of FIG. 11 is selected for transmission of the feature data.

[0082] Specifically, S202 can be implemented through the following processes of S202a1-S202a2:

[0083] S202a1, the edge device 101 determines the second bit error rate P b according to the first bit error rate P δ , the second bit error rate P δ reflects the difference between the actual bit error rate P c of the transmission channel between the edge device 101 and the edge server 102 and the first bit error rate P b , that is, P δ = P c -P b .

[0084] In other words, the second bit error rate P δ can be understood as the bit error rate that needs to be resisted by encoding in the feature data transmission process.

[0085] S202a2, determining a second bit error rate P from the plurality of transmission modes δ The corresponding first transmission mode.

[0086] Among the plurality of transmission modes, there are transmission modes with different channel coding lengths.

[0087] For example, when the second bit error rate P δ is higher, a transmission mode with a shorter channel coding length can be selected as the first transmission mode to transmit the feature data, so as to minimize the transmission delay while ensuring that the bit error rate of the feature data after transmission through the channel is less than the first bit error rate P b . δ When the second bit error rate P b is lower, a transmission mode with a longer channel coding length can be selected as the first transmission mode to transmit the feature data, so as to ensure that the bit error rate of the feature data after transmission through the channel is less than the first bit error rate P b .

[0088] In addition, in a possible design, when the actual bit error rate P c is greater than the first bit error rate P b , the first transmission mode can be a transmission mode without encoding.

[0089] For example, as shown in FIG. 12, after obtaining the actual bit error rate P c and the first bit error rate P b , the edge device 101 can first compare the actual bit error rate P c and the first bit error rate P b . If P c >P b , it indicates that without encoding, the bit error rate of the feature data received by the edge server 102 will not affect the inference result of the AI model, and therefore the feature data can be sent to the edge server 102 without encoding. In addition, if P c ≤P b , it indicates that the bit error rate of the feature data received by the edge server 102 needs to be lower than the first bit error rate P b by encoding, and at this time, an encoding mode with a corresponding channel coding length can be selected according to the second bit error rate P δ , and the feature data is sent to the edge server 102 by using the encoding mode.

[0090] For example, assuming that the first bit error rate P b is 0.4, according to the method described above in the embodiments of the present application, as long as the actual bit error rate P cless than 0.4, the feature data can not be additionally encoded. At this time, as shown in FIG. 13, when the actual bit error rate P c When the actual bit error rate P c is less than 0.4, the transmission delay can be kept at a low level because the feature data is not additionally encoded. In the transmission mode of the related art, the feature data is transmitted more accurately because the robustness of the AI model is not considered in the related art. At this time, even if the actual bit error rate P c of the transmission channel is less than 0.4, the transmission delay will continue to increase as the actual bit error rate P c increases.

[0091] In another implementation, considering that the first bit error rate P b may be different, a transmission mode with different retransmission times is selected to transmit the feature data.

[0092] It can be understood that when the service data is transmitted through the same channel, the transmission mode corresponding to the more retransmission times corresponds to a lower bit error rate of the service data. Therefore, considering that the first bit error rate P b may be different, a transmission mode with different retransmission times is selected to transmit the feature data. When the first bit error rate P b is lower, a transmission mode with more retransmission times can be selected to transmit the feature data, so as to ensure that the bit error rate of the feature data is low (lower than the first bit error rate P b ), thereby avoiding errors in the inference result of the AI model. When the first bit error rate P b is higher, a transmission mode with less retransmission times can be selected to transmit the feature data, so as to minimize the transmission delay while ensuring that the bit error rate of the feature data after being transmitted through the channel is less than the first bit error rate P b .

[0093] Specifically, S202 can be implemented through the following S202b1-S202b2 processes:

[0094] S202b1, the edge device 101 determines a second bit error rate P b based on the first bit error rate P δ . The second bit error rate P δ reflects the difference between the actual bit error rate P c of the transmission channel between the edge device 101 and the edge server 102 and the first bit error rate P b , that is, P δ = P c -P b .

[0095] As S202a1, the second bit error rate P δIt can be understood that in the feature data transmission process, the error bit rate resistant to encoding needs to be determined.

[0096] S202b2, determining a second error bit rate P from a plurality of transmission modes δ The corresponding first transmission mode.

[0097] Among the plurality of transmission modes, transmission modes with different retransmission times are included.

[0098] For example, as shown in FIG. 14, a schematic diagram of the accuracy rate changing with the transmission times of an AI model in the case of three different feature data error bit rates (indicated by BEP in the figure). Further, in the case of applying the above-mentioned method provided by the embodiments of the present application, since the retransmission times are determined according to the error bit rate resistant to encoding (i.e. the second error bit rate P δ ) in the embodiments of the present application, the retransmission times of the transmission process can be reduced compared to the related art scheme. For example, when the actual error bit rate P c is 0.4, the first error bit rate P b (i.e. the error bit rate of the feature data of the AI model corresponding to the case where the output result of the AI model meets the preset accuracy rate) is 0.2, if the related art scheme is followed, without considering the first error bit rate P b , in order to ensure that the accuracy rate of the AI model is above 0.9, the retransmission times need to be set to 30 times or more according to the curve of BEP=0.4 in FIG. 14. In the above-mentioned method provided by the embodiments of the present application, since the first error bit rate P b is considered, it can be known that the error bit rate resistant to encoding (i.e. the second error bit rate P δ ) is 0.4-0.2, so the retransmission times can be set to 5 times according to the curve of BEP=0.2 in FIG. 14.

[0099] In addition, as S202a2, in one possible design, when the actual error bit rate P c is greater than the first error bit rate P b , the first transmission mode can be a transmission mode without encoding.

[0100] For example, as shown in FIG. 15, after the edge device 101 obtains the actual error bit rate P c and the first error bit rate P b , it can first compare the actual error bit rate P c and the first error bit rate P b . If P c >P b, it indicates that the error bit rate of the feature data received by the edge server 102 without encoding does not affect the inference result of the AI model, and therefore the feature data can be sent to the edge server 102 without encoding. In addition, if P c b , it indicates that the error bit rate of the feature data received by the edge server 102 needs to be reduced to below the first error bit rate P b , the second error bit rate P δ , and the feature data is sent to the edge server 102 by using the encoding mode with the corresponding retransmission number.

[0101] S203, the edge device 101 sends the feature data of the AI model to the edge server 102 by using the first transmission mode.

[0102] In addition, in an implementation manner, as shown in FIG. 8, before the edge device 101 sends the feature data of the AI model to the edge server 102, the method can further include:

[0103] S204, the edge device 101 performs feature extraction on the original data to obtain the feature data.

[0104] For example, when the AI model in the embodiment of the present application is an image processing model, the edge device 101 can perform feature extraction on the input image data (i.e., the original data) to obtain the feature data of the image data. Then, the edge device 101 sends the feature data to the edge server 102 through the process of S203, so that the edge server performs decision inference according to the feature data from the edge device and the AI model.

[0105] Through the above method provided by the embodiment of the present application, on the one hand, when the error bit rate in the feature data transmission process is high, a transmission mode with higher reliability can be selected to transmit the feature data, so as to reduce the error bit rate in the feature data transmission process (i.e., to ensure that the error bit rate of the feature data received by the edge server 102 is lower than the first error bit rate P b ), so that the output result of the AI model meets the preset accuracy rate. On the other hand, when the error bit rate in the feature data transmission process is low, a transmission mode with relatively low reliability but small redundancy data ratio can be selected to transmit the feature data, so as to reduce the transmission delay and improve the transmission efficiency on the premise of reducing the error bit rate in the feature data transmission process (i.e., to ensure that the error bit rate of the feature data received by the edge server 102 is lower than the first error bit rate P b ).

[0106] ​Taking an AI model for classification as an example, the robustness of AI models with different discrimination gains is also different. For example, as shown in FIG. 16, a comparison diagram of the robustness of two binary AI models (i.e., AI models that classify data into two categories) is shown. As can be seen, the robustness of the AI model with a discrimination gain of 1 unit is better than that of the AI model with a discrimination gain of 2 units, that is, the greater the discrimination gain of the AI model, the better the robustness. Therefore, in the case of applying the method provided in the embodiments of the present application, the appropriate transmission mode can be selected to transmit the feature data according to the difference in the robustness of the AI model, thereby improving the transmission efficiency (specifically, the higher the robustness of the AI model, the higher the corresponding first bit error rate, and then a transmission mode with less redundancy can be selected to transmit the feature data, and vice versa). For another example, as shown in FIG. 17, a comparison diagram of the robustness of two ten-class AI models (i.e., AI models that classify data into ten categories) is shown. Similarly, as can be seen in FIG. 17, the greater the discrimination gain of the AI model, the better the robustness. Therefore, in the case of applying the method provided in the embodiments of the present application, the appropriate transmission mode can also be selected to transmit the feature data according to the difference in the robustness of the AI model, thereby improving the transmission efficiency.

[0107] Based on the above method embodiments, the device provided in the embodiments of the present application is described below. As shown in FIG. 18, a structural diagram of a data processing device provided in the embodiments of the present application is shown. Specifically, the data processing device 30 can be used for the functions of the edge device 101 described above.

[0108] Referring to FIG. 18, the data processing device 30 includes part or all of an acquisition unit 301, a processing unit 302, and a communication unit 303. Specifically:

[0109] The acquisition unit 301 is configured to acquire a first bit error rate, the first bit error rate being used to indicate a bit error rate of feature data of an AI model corresponding to a case where an output result of the AI model meets a preset accuracy rate.

[0110] The processing unit 302 is configured to determine, according to the first bit error rate, a first transmission mode corresponding to the first bit error rate from a plurality of transmission modes.

[0111] The communication unit 303 is configured to send the feature data of the AI model to an edge server by using the first transmission mode.

[0112] In an implementation manner, different transmission modes in the plurality of transmission modes correspond to different channel coding lengths.

[0113] In an implementation manner, different transmission modes in the plurality of transmission modes correspond to different retransmission times.

[0114] In an implementation manner, the processing unit 302 is configured to determine, according to the first bit error rate, a first transmission mode corresponding to the first bit error rate from a plurality of transmission modes, including: the processing unit 302 is configured to determine, according to the first bit error rate, a second bit error rate, the second bit error rate is used to reflect a difference between an actual bit error rate of a transmission channel between the edge device and the edge server and the first bit error rate. The processing unit 302 is configured to determine, from the plurality of transmission modes, a first transmission mode corresponding to the second bit error rate.

[0115] In an implementation manner, the obtaining unit 301 is configured to obtain the first bit error rate, including: the obtaining unit 301 is configured to determine the first bit error rate according to a noise variance corresponding to feature data of an artificial intelligence (AI) model in a case that an output result of the AI model meets a preset accuracy rate.

[0116] In an implementation manner, the processing unit 302 is further configured to perform feature extraction on the original data to obtain the feature data.

[0117] In an implementation manner, the edge server is configured to perform decision inference according to the feature data from the edge device and the AI model.

[0118] FIG. 19 is a structural schematic diagram of another data processing apparatus provided by the embodiment. The data processing apparatus 40 can be a chip or a system on chip. Specifically, the data processing apparatus 40 can include all or part of hardware of a desktop computer, a tablet computer, a desktop, a laptop, a handheld computer, a notebook computer, an ultra-mobile personal computer, a netbook, and an electronic device such as a cellular phone, a personal digital assistant, and an augmented reality / virtual reality device.

[0119] The data processing apparatus 40 can include some or all of the following components: a processor 401, a communication circuit 402, a memory 403, and at least one communication interface 404.

[0120] The processor 401 is configured to perform all or part of the steps performed by the edge device 101 in the embodiment.

[0121] In particular, the processor 401 can include a general-purpose central processing unit (CPU), and can also include a microprocessor, a microcontroller, a field programmable gate array (FPGA), a digital signal processor (DSP), or an application-specific integrated circuit (ASIC), or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or the like.

[0122] In a particular implementation, as one example, the processor 401 can include one or more CPUs, such as CPU0 and CPU1 in FIG. 19.

[0123] In a particular implementation, as one example, the apparatus 40 can include multiple processors, such as the processor 401 and the processor 408 in FIG. 19. Each of these processors can be a single-CPU processor or a multi-CPU processor. A processor here can refer to one or more devices, circuits, and / or processing cores for processing data, such as computer program instructions.

[0124] In addition, the memory 403 can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically EPROM (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM) used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous dynamic RAM (SDRAM), double data rate synchronous dynamic RAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DR RAM). The memory 403 can exist independently, and be connected to the processor 401 through the communication line 402. The memory 403 can also be integrated with the processor 401.

[0125] The memory 403 stores computer instructions. The processor 401 can execute the computer instructions stored in the memory 403 to perform all or part of the steps of the method provided in the embodiments.

[0126] Optionally, the computer execution instructions in the embodiments can also be referred to as application program codes, which are not specifically limited in the embodiments.

[0127] In addition, the communication interface 404 uses any transceiver-like device to communicate with other devices or communication networks, such as an Ethernet, a radio access network (RAN), a wireless local area network (WLAN), etc.

[0128] In addition, the communication line 402 is used to connect the components in the data processing apparatus 40. Specifically, the communication line 402 can include a data bus, a power supply bus, a control bus, and a state signal bus, etc. However, for the purpose of clear illustration, all kinds of buses are marked as the communication line 402 in the figure.

[0129] In a specific implementation, as an embodiment, the data processing apparatus 40 can further include an output device 407 and an input device 406. The output device 407 can communicate with the processor 401 and can display information in various ways. For example, the output device 407 can be a liquid crystal display (LCD), a light emitting diode (LED) display device, a cathode ray tube (CRT) display device, or a projector, etc. The input device 406 can communicate with the processor 401 and can receive user input in various ways. For example, the input device 406 can be a mouse, a keyboard, a touch screen device, or a sensor device, etc.

[0130] In addition, the data processing apparatus 40 can further include a storage medium 405. The storage medium 405 is used to store computer instructions and various data for implementing the technical solutions of the present embodiment. In order for the data processing apparatus 40 to execute the above-mentioned method of the present embodiment, the computer instructions and various data stored in the storage medium 405 are loaded into the memory 403, so that the processor 401 can execute the computer instructions stored in the memory 403 to execute the method provided by the present embodiment.

[0131] The method steps in the present embodiment can be implemented by hardware or by the processor executing software instructions. The software instructions can be composed of corresponding software modules, which can be stored in a RAM, a flash memory, a ROM, a PROM, an EPROM, an EEPROM, a register, a hard disk, a mobile hard disk, a CD-ROM, or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor, so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and the storage medium can be located in an ASIC. In addition, the ASIC can be located in the data processing apparatus. Of course, the processor and the storage medium can also exist as discrete components in the data processing apparatus.

[0132] In the above embodiments, all or part can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part can be implemented in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer programs or instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments are performed. The computer can be a general purpose computer, a special purpose computer, a computer network, a communication device, a user equipment or other programmable device. The computer programs or instructions can be stored in a computer readable storage medium or transferred from one computer readable storage medium to another computer readable storage medium, for example, the computer programs or instructions can be transferred from one website site, computer, server or data center to another website site, computer, server or data center through wired or wireless manner. The computer readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center and the like integrated with one or more available media. The available media can be a magnetic medium, such as a floppy disk, a hard disk, a magnetic tape; an optical medium, such as a digital video disc (digital video disc, DVD); and a semiconductor medium, such as an SSD.

[0133] In the embodiments, the terms and / or descriptions among different implementation manners are consistent and can be referred to each other if there is no special description and logical conflict, and the technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationship.

[0134] In the embodiments, "at least one" means one or more, "multiple" means two or more, and other quantifiers are similar. The association relationship of the associated objects is described by "and / or", which means that there can be three relationships, for example, A and / or B can represent the following three cases: A exists alone, A and B exist together, and B exists alone. In addition, for the elements (element) appearing in the singular form "a", "an" and "the", unless the context clearly specifies otherwise, it does not mean "one or only one", but means "one or more than one". For example, "a device" means one or more such devices. Furthermore, "at least one of" means one or any combination of the subsequent associated objects, for example, "at least one of A, B and C" includes A, B, C, AB, AC, BC or ABC. In the literal description of the embodiments, the character " / ", generally represents that the preceding and following associated objects are in an "or" relationship; in the formulas of the embodiments, the character " / ", represents that the preceding and following associated objects are in a "division" relationship.

Claims

1. A data processing method, characterized by, The method is applied to an edge device, and comprises: obtaining a first bit error rate, the first bit error rate being used to indicate a bit error rate corresponding to feature data of an AI model in a case where an output result of the AI model meets a preset accuracy rate; determining, according to the first bit error rate, a first transmission mode corresponding to the first bit error rate from a plurality of transmission modes; and transmitting, by using the first transmission mode, the feature data of the AI model to an edge server.

2. The method of claim 1, wherein, Different transmission modes in the plurality of transmission modes correspond to different channel coding lengths.

3. The method of claim 1, wherein, Different transmission modes in the plurality of transmission modes correspond to different retransmission times.

4. The method according to any one of claims 1 to 3, characterized in that, The determining, according to the first bit error rate, of a first transmission mode corresponding to the first bit error rate from a plurality of transmission modes comprises: determining, according to the first bit error rate, a second bit error rate, the second bit error rate being used to reflect a difference between an actual bit error rate of a transmission channel between the edge device and the edge server and the first bit error rate; and determining, from the plurality of transmission modes, a first transmission mode corresponding to the second bit error rate.

5. The method according to any one of claims 1 to 4, characterized in that, The obtaining of the first bit error rate comprises: determining the first bit error rate according to a noise variance corresponding to the feature data of the AI model in a case where the output result of the AI model meets the preset accuracy rate.

6. The method according to any one of claims 1 to 5, characterized in that, The method further comprises performing feature extraction on original data to obtain the feature data.

7. The method according to any one of claims 1 to 6, characterized in that, The edge server is configured to perform decision inference according to the feature data from the edge device and the AI model.

8. A data processing apparatus, characterized by, The method is applied to an edge device, and comprises: an obtaining unit configured to obtain a first bit error rate, the first bit error rate being used to indicate a bit error rate corresponding to feature data of an AI model in a case where an output result of the AI model meets a preset accuracy rate; a processing unit configured to determine, according to the first bit error rate, a first transmission mode corresponding to the first bit error rate from a plurality of transmission modes; and a communication unit configured to transmit, by using the first transmission mode, the feature data of the AI model to an edge server.

9. A data processing apparatus, characterized by, The computer program product comprises instructions, and the instructions, when executed on a processor, implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program product comprises instructions, and the instructions, when executed on a processor, implement the method according to any one of claims 1 to 7.

11. A computer program product, characterised in that, The computer program product comprises instructions, and the instructions, when executed on a processor, implement the method according to any one of claims 1 to 7.

Images