Data distribution method and device, electronic equipment and storage medium

Abstract

The invention discloses a data distribution method and device, electronic equipment and a storage medium, the data distribution method is applied to a distribution server deployed at a network side, and comprises the following steps: obtaining distribution characteristic statistical data of a target data packet to be sent by terminal equipment; inputting the distribution characteristic statistical data of the target data packet into a first data distribution decision network to obtain a first distribution confidence coefficient of the target data packet matched with each transmission path; inputting the distribution characteristic statistical data of the target data packet into a second data distribution decision network to obtain a second distribution confidence coefficient of the target data packet matched with each transmission path; and matching the first distribution confidence coefficient and the second distribution confidence coefficient of each transmission path based on the target data packet, and determining a target distribution strategy of the target data packet. By applying the technical scheme provided by the invention, the distribution strategy of the uplink data packet of the terminal equipment can be accurately and effectively determined, and the data transmission efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer application, in particular to a data shunting method and device, electronic equipment and storage medium. BACKGROUND

[0002] With the rapid development of communication technology and computer technology, a 5th Generation (5G) dual-domain network environment is gradually formed. The 5G dual-domain network environment refers to a network environment that is formed by combining a public operator network (such as a 5G mobile communication network) and a private network (private domain) (such as a Wireless Fidelity (Wi-Fi) network) built by users themselves through technical means such as physical isolation, network slicing, data shunting, and communication encryption. This network environment aims to provide higher levels of security and network performance for enterprises and organizations.

[0003] In the 5G dual-domain network environment, how to accurately and effectively determine the shunting strategy of the uplink data packet of the terminal device and improve the data transmission efficiency is a technical problem that needs to be solved by those skilled in the art. SUMMARY

[0004] The purpose of the present application is to provide a data shunting method, device, electronic equipment and storage medium to accurately and effectively determine the shunting strategy of the uplink data packet of the terminal device and improve the data transmission efficiency.

[0005] To solve the above technical problems, the present application provides the following technical solutions:

[0006] In a first aspect, a data shunting method is provided, which is applied to a shunting server deployed on a network side, and the method comprises:

[0007] Obtaining shunting feature statistical data of a target data packet to be sent by a terminal device;

[0008] Inputting the shunting feature statistical data of the target data packet into a first data shunting decision network to obtain a first shunting confidence of the target data packet matching each transmission path, wherein the first data shunting decision network is obtained by learning network parameters of a pre-constructed convolutional neural network according to a meta-cognition learning algorithm;

[0009] Inputting the shunting feature statistical data of the target data packet into a second data shunting decision network to obtain a second shunting confidence of the target data packet matching each transmission path, wherein the second data shunting decision network is obtained by learning network parameters of a pre-constructed multi-layer perception network according to a biological heuristic algorithm;

[0010] The first confidence degree and the second confidence degree of each transmission path matched with the target data packet are determined based on the target data packet, and a target data shunting strategy of the target data packet is determined.

[0011] In a second aspect, a data shunting device is provided, which is applied to a shunting server deployed at a network side, and the device comprises:

[0012] a feature data obtaining module, configured to obtain shunting feature statistical data of a target data packet to be sent by a terminal device;

[0013] a first confidence degree obtaining module, configured to input the shunting feature statistical data of the target data packet into a first data shunting decision network to obtain a first confidence degree of each transmission path matched with the target data packet, wherein the first data shunting decision network is obtained by performing network parameter learning on a pre-constructed convolutional neural network according to a meta-cognition learning algorithm;

[0014] a second confidence degree obtaining module, configured to input the shunting feature statistical data of the target data packet into a second data shunting decision network to obtain a second confidence degree of each transmission path matched with the target data packet, wherein the second data shunting decision network is obtained by performing network parameter learning on a pre-constructed multi-layer perception network according to a biological heuristic algorithm;

[0015] a shunting strategy determining module, configured to determine a target data shunting strategy of the target data packet based on the first confidence degree and the second confidence degree of each transmission path matched with the target data packet.

[0016] In a third aspect, an electronic device is provided, which comprises:

[0017] a memory, configured to store a computer program;

[0018] a processor, configured to execute the computer program to implement the steps of the data shunting method according to the first aspect.

[0019] In a fourth aspect, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps of the data shunting method according to the first aspect.

[0020] In a fifth aspect, a computer program product is provided, which comprises computer instructions stored in a computer readable storage medium and adapted to be read and executed by a processor to enable a computer device with the processor to execute the steps of the data shunting method according to the first aspect.

[0021] By applying the technical solution provided in the embodiments of the present application, after obtaining the shunting feature statistical data of the target data packet to be sent by the terminal device, the first data shunting decision network and the second data shunting decision network are used for prediction respectively based on the shunting feature statistical data, the first shunting confidence and the second shunting confidence of the target data packet matching each transmission path are obtained, and the target shunting strategy of the target data packet is determined based on the first shunting confidence and the second shunting confidence of the target data packet matching each transmission path. The first data shunting decision network is obtained by learning the network parameters of the pre-constructed convolutional neural network according to the meta-cognition learning algorithm, and the second data shunting decision network is obtained by learning the network parameters of the pre-constructed multi-layer perception network according to the biological heuristic algorithm. By combining the prediction results of the transmission paths by different data shunting decision networks, the shunting strategy of the uplink data packet of the terminal device can be accurately and effectively determined, and the data transmission efficiency is improved.

[0022] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present application. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or the related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0024] Figure 1 The structure schematic diagram of the data transmission system in the embodiments of the present application is shown in FIG. 1.

[0025] Figure 2 The implementation flowchart of the data shunting method in the embodiments of the present application is shown in FIG. 2.

[0026] Figure 3 The structure schematic diagram of the data shunting apparatus in the embodiments of the present application is shown in FIG. 3.

[0027] Figure 4 The structure schematic diagram of the electronic device in the embodiments of the present application is shown in FIG. 4. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art belong to the scope of protection of the present application.

[0029] The terms "first", "second", and the like in this application are used to distinguish similar objects, and are not used to describe a particular order or sequence. It should be understood that the terms used in this way can be interchanged as appropriate, so that the embodiments of the present application can be implemented in an order other than that illustrated or described herein, and the objects distinguished by "first", "second" are generally of a kind and do not limit the number of objects, for example, the first object can be one or more.

[0030] The core of the present application is to provide a data shunting method, which can be applied to the scenario of determining a shunting strategy for terminal equipment to send uplink data packets in a dual-domain network or multi-domain network environment, and is executed by a shunting server deployed on the network side.

[0031] Referring to Figure 1 The structure of the data transmission system in the embodiment of the present application is shown in the figure, which includes a shunting server, a terminal equipment, a mobile communication network, a Wi-Fi network, and a target server. The terminal equipment supports dual-mode access to the mobile communication network and the Wi-Fi network. When the terminal equipment has a sending demand for uplink data, it can send a request to the shunting server through the mobile communication network or the Wi-Fi network to obtain a shunting strategy. The shunting server executes the technical solution provided in the embodiment of the present application to determine a target shunting strategy and indicate the target shunting strategy to the terminal equipment through the mobile communication network or the Wi-Fi network. The terminal equipment sends uplink data packets according to the target shunting strategy indicated by the shunting server, and finally sends the uplink data packets to the target server.

[0032] In the embodiment of the present application, the shunting server is deployed on the network side. The network side system has a more comprehensive global view and centralized control capability in the complex decision-making scenario of cross-domain and cross-network nodes. Specifically, the shunting server deployed on the network side has the following advantages:

[0033] (1) Global view and centralized control: it can obtain multi-domain network state information (such as signal quality, bandwidth utilization, delay, and load condition) in the entire network range, thereby realizing globally optimal shunting decision and resource scheduling;

[0034] (2) Dynamic adaptation and continuous optimization: it can continuously collect network data and perform model retraining and optimization to ensure that the shunting decision continuously adapts to the rapidly changing network environment and business requirements;

[0035] (3) Cross-domain collaboration: it is easier to realize collaborative control between the mobile communication network and the Wi-Fi network, and to ensure smooth switching and efficient transmission of data between different network domains.

[0036] In the related art, a data splitting scheme deployed on the network side is to make a splitting decision through a preconfigured static intranet Internet Protocol (IP) address segment. The basic process is: in the Protocol Data Unit (PDU) session creation process, a user plane function (UPF) for splitting is selected and configured by a session management function (SMF), when a user enters a service area, an access and mobility management function (AMF) obtains user information from a unified data management function (UDM) and initiates a PDU session creation request to the SMF; for an uplink data packet, a static rule judgment is made according to whether the destination IP address matches the preconfigured intranet IP address segment, if it matches, the data packet is forwarded to the intranet, otherwise it is forwarded to the Internet.

[0037] The static matching rule scheme has obvious defects and cannot meet the complex and variable environment requirements of the 5G dual-domain network, which is specifically manifested in:

[0038] Lack of intelligent decision-making ability: only relying on pre-set address segment rules, it cannot make flexible and dynamic splitting decisions according to real-time network state (signal strength, bandwidth utilization, latency), data packet content, transmission protocol, service type, etc.

[0039] Lack of learning and optimization mechanism: no machine learning or artificial intelligence algorithm is introduced for adaptive optimization, resulting in limited accuracy and comprehensiveness of the data splitting scheme, and static rules are difficult to adapt to changes in network environment or business requirements;

[0040] No processing of noise data: when there are abnormal values or error data in the data, there is no corresponding mechanism to identify and eliminate them, which may reduce the decision-making accuracy.

[0041] Therefore, an embodiment of the present application proposes an intelligent adaptive data splitting method deployed on the network side to solve the problems of poor flexibility, low accuracy and inability to adapt to dynamic and complex network environments in the related art.

[0042] It should be noted that the above is an example of explaining that the splitting server is a separate server, but it can be understood that in actual application, the splitting server can also be replaced by a splitting server cluster, or a distributed cluster composed of multiple splitting servers. Correspondingly, in the Figure 1 , the splitting server can also be replaced by a splitting platform composed of multiple splitting servers.

[0043] Referring to Figure 2 As shown in the figure, an implementation flowchart of a data shunting method provided by an embodiment of the present application can include the following steps:

[0044] S210: Obtain shunting feature statistical data of a target data packet to be sent by the terminal device.

[0045] In the embodiment of the present application, the terminal device can be an intelligent camera, an access control, a visitor tablet, a mobile terminal, a wearable device, a robot, an automated guided vehicle (AGV), an Internet of Things (IoT) sensor, etc. The terminal device supports multi-mode networking functions, such as supporting 5G mobile communication network and Wi-Fi network dual-mode networking functions, and has automatic negotiation and switching capabilities.

[0046] When the terminal device needs to send an uplink data packet, it can send a request to the shunting server to inquire which network domain should be selected for data transmission. After receiving the request, the shunting server can obtain the shunting feature statistical data of the target data packet to be sent by the terminal device, which is the basis for intelligent shunting decision.

[0047] Optionally, the shunting feature statistical data of the target data packet can include at least one of the following:

[0048] Data content feature data, such as data type (video stream data, text data, audio data, other types of data, etc.), video resolution (such as 1080p or 4K), frame rate (such as 30 frames per second or 60 frames per second), encoding format (such as H.264 or H.265), etc. These feature data reflect the content attributes of the target data packet itself;

[0049] Transmission protocol feature data, such as Transmission Control Protocol (TCP) or User Datagram Protocol (UDP). These feature data describe the transmission method of the target data packet. If it is a TCP protocol, the relevant protocol parameters such as window size and sequence number can be obtained, and if it is a UDP protocol, the length and sending frequency of the data packet can be focused on;

[0050] Network state feature data, such as signal strength, bandwidth utilization, latency, packet loss rate, etc. of 5G mobile communication network and Wi-Fi network, which reflects the current network environment. For example, the signal strength of 5G mobile communication network is weak, the bandwidth utilization is high, the latency is 50 ms, and the packet loss rate is 5%, while the signal strength of Wi-Fi network is strong, the bandwidth utilization is low, the latency is 20 ms, and the packet loss rate is 2%;

[0051] Dual-domain interaction behavior feature data, such as network switching history data, interaction mode, etc., which reflects the switching situation of data packets between different network domains.

[0052] Among them, the network state feature data can be reported by the terminal device, or obtained by the distribution server through at least one of the following ways:

[0053] Network node reporting: the network nodes such as base stations and access points deployed in the network periodically report their current signal strength, bandwidth utilization and other network state information to the distribution server. These information is an important basis for the distribution server to make data distribution decision;

[0054] Active detection: the distribution server can periodically send detection signals to different areas in the network to obtain the network state information of the area. For example, a packet internet groper (ping) packet or a traceroute command can be sent to detect the delay and packet loss rate between different network domains;

[0055] Third-party data source: the distribution server obtains network state information from a third-party data source. For example, real-time network state data can be obtained from the network management system of the operator, or network state information can be obtained from a third-party network monitoring service provider.

[0056] Through the above ways, the distribution server can obtain the key indicators such as signal strength of mobile communication network or Wi-Fi network in real time, which provides strong support for data distribution decision.

[0057] For example, for a data packet from an online music playing application, its data content feature data may include the code rate and format of the music; the transmission protocol feature data may be UDP protocol, and the sending frequency of the data packet is relatively stable; the network state feature data may be that the current 5G mobile communication network signal is good but the bandwidth utilization is high, and the Wi-Fi network signal is general but the bandwidth is idle; the dual-domain interaction behavior feature data may be that it mainly depends on Wi-Fi network before, and the current 5G mobile communication network is available;

[0058] For a live high-definition video data packet, its data content feature data can include high resolution (such as 4K), high frame rate (such as 60 frames per second), and specific encoding format (such as H.265); the transmission protocol feature data shows TCP protocol, moderate window size, and continuous sequence number; the network state feature data indicates that the 5G mobile communication network signal strength is good but the bandwidth utilization is high, and the Wi-Fi network signal is stable but the bandwidth is limited; the dual-domain interaction behavior feature data shows that it is mainly transmitted through the Wi-Fi network before, and the 5G mobile communication network condition is improved at present.

[0059] S220: input the shunt feature statistical data of the target data packet into the first data shunt decision network to obtain the first shunt confidence of the target data packet matching each transmission path.

[0060] The first data shunt decision network is obtained by learning the network parameters of the pre-constructed convolutional neural network according to the meta-cognition learning algorithm.

[0061] In the embodiments of the present application, a convolutional neural network (CNN) can be pre-constructed, which can include a first input layer, a plurality of convolutional layers, a pooling layer, a fully connected layer and a first output layer connected in sequence.

[0062] The first input layer is used to receive the input data.

[0063] The plurality of convolutional layers are used to extract local features in the input data through convolution kernels. Each convolutional layer applies a plurality of different convolution kernels to capture different types of feature patterns.

[0064] The pooling layer is used to reduce the dimension of the feature map after the plurality of convolutional layers, reduce the calculation amount, and retain important features. The pooling operation includes maximum pooling and average pooling.

[0065] The fully connected layer is used to learn global features after a plurality of convolution and pooling operations. The feature map is flattened and input into the fully connected layer. Each neuron in the fully connected layer is connected to all neurons in the previous layer.

[0066] The first output layer is used to output the prediction result. The form of the first output layer depends on the specific task, for example, a normalization exponential function (such as a softmax function) can be used for multi-classification, or an activation function (such as a sigmoid function) can be used for binary classification.

[0067] According to the meta-cognitive learning algorithm, the network parameters of the convolutional neural network are learned, and a first data offloading decision network can be obtained. Meta-cognitive learning simulates the self-reflection and self-regulation ability of human beings, guides the parameter optimization and model selection in the machine learning process, and improves the accuracy and generalization ability of the model. During the meta-cognitive learning process, the convolutional neural network will continuously adjust the internal weights and parameters to better understand and predict the offloading path of the data packet.

[0068] The offloading feature statistical data of the target data packet is input into the first data offloading decision network, and the neurons in the first data offloading decision network start to calculate and transmit information.

[0069] For example, for data content feature data such as the resolution and frame rate of high-definition video, the first data offloading decision network is processed, and the pattern of similar data content feature data learned before is matched and compared. For transmission protocol feature data such as the window size and sequence number of the TCP protocol, it is used to evaluate the transmission reliability and efficiency of the target data packet under different network conditions. For network state feature data such as the signal strength and bandwidth utilization of 5G cellular networks and Wi-Fi networks, it is used to determine which network domain is more suitable for carrying the target data packet. For dual-domain interaction behavior feature data such as the network switching history data of the target data packet, it is used to predict the possibility of future network switching and its impact on data transmission. After complex calculation and reasoning, the first data offloading decision network can output the first offloading confidence of the target data packet matching each transmission path. For example, the first offloading confidence of the target data packet matching the 5G mobile communication network and the first offloading confidence of the target data packet matching the Wi-Fi network are output. For each transmission path, the higher the first offloading confidence of the target data packet matching the current transmission path, the more likely the first data offloading decision network considers that the data transmission is carried out through the current transmission path. For example, if the first offloading confidence of the target data packet matching the 5G mobile communication network is 0.8, it is considered that the target data packet has an 80% possibility of being suitable for transmission through the 5G mobile communication network.

[0070] S230: Input the offloading feature statistical data of the target data packet into the second data offloading decision network to obtain the second offloading confidence of the target data packet matching each transmission path.

[0071] The second data offloading decision network is obtained by learning the network parameters of the pre-constructed multilayer perceptron network according to the biological heuristic algorithm.

[0072] In the embodiments of the present application, a multilayer perceptron (MLP) network can be pre-constructed, which includes a second input layer, a plurality of hidden layers and a second output layer connected in turn.

[0073] The second input layer is configured to receive input data.

[0074] The plurality of hidden layers each comprise a plurality of neurons that perform non-linear transformations on the input data through activation functions, thereby extracting complex features from the input data.

[0075] The second output layer is configured to output prediction results. The form of the second output layer depends on the specific task, for example, a softmax function can be used for multi-classification, or a sigmoid function can be used for binary classification.

[0076] The biological heuristic algorithm can be used to learn the network parameters of the multi-layer perceptron network to obtain the second data shunting decision network. Optionally, the biological heuristic algorithm can be an ant colony algorithm. The network simulates the intelligent behavior of a biological population to make decisions, such as the behavior of ants searching for the optimal path. Through the ant colony algorithm, the network can adaptively adjust the parameters to find the optimal solution, making the data shunting decision more efficient. At the beginning of training, a basic biological heuristic search cluster can be created, which includes a plurality of multi-layer perceptron networks (or multi-layer perceptron network instances), and the basic weight parameters of each multi-layer perceptron network are configured differently to increase the diversity of the search. During the training process, the evaluation network model can be used to determine the iterative weight distribution of each multi-layer perceptron network according to the estimated confidence generated by each multi-layer perceptron network, which is used for subsequent cluster evolution, thereby obtaining the second data shunting decision network.

[0077] The shunting feature statistical data of the target data packet is input into the second data shunting decision network, and the second data shunting decision network simulates the intelligent behavior of a biological population to make decisions.

[0078] For example, for data content feature data, the second data shunting decision network, like ants searching for the optimal path, explores the path most suitable for processing this type of data according to the previously learned data content patterns. For transmission protocol feature data, it is used to evaluate the efficiency and reliability of information transmission similar to that in a biological population. For network state feature data, it is considered as the environmental conditions in which the biological population is located to determine the optimal processing method of the target data packet in this environment. For dual-domain interaction behavior feature data, similar to the cooperation and competition patterns between individuals in a biological population, it helps to determine the best transmission path of the target data packet in the dual-domain environment.

[0079] The second data offloading decision network can output a second distribution confidence of the target data packet matching each transmission path after operation. For example, the second distribution confidence of the target data packet matching the 5G mobile communication network and the second distribution confidence of the target data packet matching the Wi-Fi network are output. For each transmission path, the higher the second distribution confidence of the target data packet matching the current transmission path, the greater the possibility of data transmission through the current transmission path considered by the second data offloading decision network. For example, if the second distribution confidence of the target data packet matching the 5G mobile communication network is 0.75, it is considered that the target data packet has a 75% possibility of being suitable for transmission through the 5G mobile communication network.

[0080] S240: determining a target offloading strategy of the target data packet based on the first distribution confidence and the second distribution confidence of the target data packet matching each transmission path.

[0081] After obtaining the first distribution confidence of the target data packet matching each transmission path output by the first data offloading decision network and the second distribution confidence of the target data packet matching each transmission path output by the second data offloading decision network, the first distribution confidence and the second distribution confidence are comprehensively considered to determine the target offloading strategy of the target data packet.

[0082] Optionally, the first distribution confidence and the second distribution confidence corresponding to each transmission path can be fused to obtain a distribution confidence corresponding to each transmission path, and the target offloading strategy of the target data packet is determined based on the distribution confidence corresponding to each transmission path.

[0083] Optionally, for each transmission path, the average or weighted average of the first distribution confidence and the second distribution confidence corresponding to the current transmission path can be determined as the distribution confidence corresponding to the current transmission path. The weights corresponding to the first distribution confidence and the second distribution confidence can be set and adjusted according to actual conditions. The current transmission path refers to the transmission path to which the current operation is directed.

[0084] For example, if the first distribution confidence corresponding to the current transmission path is 0.8 and the second distribution confidence corresponding to the current transmission path is 0.75, the distribution confidence corresponding to the current transmission path can be determined as (0.8+0.75) / 2=0.775, or the distribution confidence corresponding to the current transmission path can be determined as 0.8*0.6+0.75*0.4=0.78.

[0085] Optionally, the target offloading strategy can include a target transmission path strategy, and the transmission path with the highest distribution confidence can be determined as the target transmission path of the target data packet.

[0086] Optionally, the transmission path with the highest distribution confidence can be determined as the target transmission path of the target data packet when the highest distribution confidence is greater than or equal to a second threshold. The second threshold can be set and adjusted according to actual conditions, such as being set to 0.7.

[0087] Optionally, the target distribution strategy can include a data caching strategy. When the highest distribution confidence is less than the second threshold, it can be considered that none of the current multiple network domains is suitable for data transmission, and the terminal device can be instructed to cache data and wait for more suitable network conditions. After a suitable transmission path is available, the terminal device can be instructed again.

[0088] Optionally, the target distribution strategy can include a distribution proportion strategy for each transmission path.

[0089] After the target distribution strategy of the target data packet is determined, the target distribution strategy can be further sent to the terminal device, such as being sent to the relevant network nodes and terminal devices through signaling or control messages, so that the terminal device selects a transmission path to send the target data packet based on the target distribution strategy, or sends the target data packet according to the distribution proportion of each transmission path.

[0090] Optionally, the target distribution strategy sent by the distribution server to the terminal device has a time limit, that is, within the effective time of the target distribution strategy, the terminal device can send the uplink data packet according to the target distribution strategy, which helps to reduce the interaction between the terminal device and the distribution server and save network resources.

[0091] By applying the method provided in the embodiments of the present application, after obtaining the distribution characteristic statistical data of the target data packet to be sent by the terminal device, the first data distribution decision network and the second data distribution decision network are used respectively for prediction based on the distribution characteristic statistical data, the first distribution confidence and the second distribution confidence of the target data packet matching each transmission path are obtained, and the target distribution strategy of the target data packet is determined based on the first distribution confidence and the second distribution confidence of the target data packet matching each transmission path. The first data distribution decision network is obtained by learning the network parameters of a pre-constructed convolutional neural network according to a meta-cognition learning algorithm, and the second data distribution decision network is obtained by learning the network parameters of a pre-constructed multi-layer perception network according to a biological heuristic algorithm. By combining the prediction results of different data distribution decision networks for the transmission path, the distribution strategy of the uplink data packet of the terminal device can be accurately and effectively determined, and the data transmission efficiency is improved.

[0092] In some embodiments of the present application, the first data distribution decision network can be obtained by the following steps:

[0093] obtaining a first data set and a second data set, the first data set comprising flow feature statistics and allocation path labels of a plurality of first samples, and the second data set comprising flow feature statistics and allocation path labels of a plurality of second samples;

[0094] inputting the flow feature statistics of each first sample in the first data set into a pre-constructed convolutional neural network, adjusting weights and parameters of the convolutional neural network according to a deviation between an output result of each first sample corresponding to the convolutional neural network and an allocation path label corresponding to the corresponding first sample, and obtaining a meta-flow decision network after meta-cognition learning;

[0095] inputting the flow feature statistics of each second sample in the second data set into the meta-flow decision network, adjusting parameters of the meta-flow decision network according to a deviation between an output result of each second sample corresponding to the meta-flow decision network and an allocation path label corresponding to the corresponding second sample, and obtaining a first data flow decision network.

[0096] For convenience of description, the above steps are combined for description.

[0097] In the embodiments of the present application, a first data set and a second data set can be obtained for training a pre-constructed convolutional neural network. The first data set comprises flow feature statistics and allocation path labels of a plurality of first samples, and the second data set comprises flow feature statistics and allocation path labels of a plurality of second samples. The first data set and the second data set cover various different scenarios and network conditions, for example, there are data in peak network congestion, and there are data in off-peak network smoothness; there are data from video conference applications, and there are data from file download applications. Each first sample and each second sample corresponds to a specific flow decision knowledge point, that is, each first sample and each second sample corresponds to an allocation path label, which indicates which network domain should be preferentially transmitted in the current scenario or network state. For example, for a data packet with high real-time requirement (such as video call), its allocation path label may correspond to a low delay network domain, indicating that it is preferentially transmitted through a low delay network. For example, for a video conference data packet, its allocation path label is 5G mobile communication network, indicating that it is preferentially transmitted through 5G mobile communication network to ensure the smoothness of the video. For example, for a data packet of a previous game, its allocation path label may correspond to a low delay and stable network.

[0098] The first data set can also be referred to as a first sample to be shunted data packet sequence, and the second data set can also be referred to as a second sample to be shunted data packet sequence. The first sample to be shunted data packet sequence includes a plurality of first sample to be shunted data packet subsequences, and first samples in a same first sample to be shunted data packet subsequence have the same allocation path tag. The second sample to be shunted data packet sequence includes a plurality of second sample to be shunted data packet subsequences, and second samples in a same second sample to be shunted data packet subsequence have the same allocation path tag.

[0099] Optionally, the first data set and the second data set can be obtained from a historical data repository. Alternatively, a plurality of first samples and a plurality of second samples are obtained from the historical data repository, and then each first sample and each second sample is analyzed respectively to obtain the shunting feature statistical data and the allocation path tag of each first sample and the shunting feature statistical data and the allocation path tag of each second sample.

[0100] Based on the first data set, meta-cognition learning is performed on a pre-constructed convolutional neural network. Specifically, the shunting feature statistical data of each first sample can be input into the convolutional neural network in sequence, and the weights and parameters of the convolutional neural network are adjusted based on the deviation between the output result of the convolutional neural network corresponding to each first sample and the allocation path tag corresponding to the corresponding first sample, so that the convolutional neural network can better understand and predict the transmission path of the first sample. After multiple iterations and learning, a meta-shunting decision network that completes the meta-cognition learning is obtained. For each first sample, the output result of the convolutional neural network corresponding to the current first sample can be an estimated confidence of the current first sample matching each transmission path. The estimated confidence and the allocation confidence can be considered as different representations of the network output result (prediction result) at different stages.

[0101] Based on the second data set, the meta-shunting decision network is optimized. Specifically, the shunting feature statistical data of each second sample can be input into the meta-shunting decision network in sequence, and the meta-shunting decision network outputs the estimated confidence of each second sample matching each transmission path based on the current network parameters and the learned knowledge. The deviation between the output result of the meta-shunting decision network corresponding to each second sample and the allocation path tag corresponding to the corresponding second sample can be determined.

[0102] For example, if the allocation path tag of the current second sample is a Wi-Fi network, but the estimated confidence predicted by the meta-shunting decision network shows that the 5G mobile communication network should be selected (for example, the estimated confidence of the 5G mobile communication network is 0.9, and the estimated confidence of the Wi-Fi network is 0.1), then there is a large deviation between them.

[0103] Optionally, for each second sample, the deviation between the output result of the meta offloading decision network corresponding to the current second sample and the allocation path label corresponding to the current second sample can be measured using a cross-entropy loss function, which is used to measure the difference between the predicted probability distribution and the true probability distribution. For example, the allocation path label of the current second sample is the Wi-Fi network, and the output result of the meta offloading decision network is: the estimated confidence of the 5G mobile communication network is 0.7, and the estimated confidence of the Wi-Fi network is 0.3. The cross-entropy loss is calculated to measure the deviation, and then the backpropagation algorithm is used to adjust the parameters of the meta offloading decision network, such as adjusting the connection weights related to the Wi-Fi network in the meta offloading decision network, increasing its estimated confidence, while reducing the connection weights related to the 5G mobile communication network, reducing its estimated confidence. After multiple iterations and learning, the meta offloading decision network can gradually recognize that the current second sample should choose the Wi-Fi network under the current network state and data characteristics. Alternatively, mean square error can be used to measure. The specific choice depends on the nature of the task and the output form.

[0104] According to the deviation between the output result of the meta offloading decision network corresponding to each second sample and the allocation path label corresponding to the corresponding second sample, the parameters of the meta offloading decision network are adjusted, including the connection weights and biases between neurons, etc. After multiple such parameter adjustments and learning, a first data offloading decision network with more optimized performance is obtained.

[0105] The purpose of adjustment is to reduce the deviation, so that the prediction result of the meta offloading decision network gradually approaches the corresponding allocation path label. This can be achieved through the backpropagation algorithm, that is, iteratively updating the network parameters according to the gradient of the loss function. The related concepts involved are as follows:

[0106] Backpropagation algorithm: calculate the error of network output according to the loss function, and propagate the error from the output layer to the input layer, and use gradient descent and other optimization algorithms to update the weights and biases in the network;

[0107] Gradient descent: adjust the network parameters in the opposite direction of the gradient of the loss function, so that the loss function value gradually decreases, thereby improving the prediction accuracy of the network;

[0108] Parameter adjustment amplitude: the adjustment amplitude is usually related to the size of the deviation and the learning rate. Larger deviation will lead to larger adjustment amplitude to quickly correct errors; learning rate controls the step size of adjustment to prevent network parameters from oscillating or failing to converge during training;

[0109] Iterative optimization: this process is performed multiple times, and each iteration uses new training data and adjusts the network parameters according to the new deviation until the performance of the network converges to a satisfactory level.

[0110] Based on the first data set and the second data set, the pre-constructed convolutional neural network is learned according to the meta-cognitive learning algorithm, and a first data flow decision network with high prediction accuracy can be obtained. In this way, more accurate prediction results can be obtained when the first data flow decision network is applied.

[0111] During the training process of the first data flow decision network, the parameter adjustment of the meta-flow decision network is a core link. This process iteratively optimizes network performance by comparing the difference between estimated confidence and assigned path labels, thereby improving the accuracy of data flow.

[0112] It should be noted that the second data set and the first data set can be generated in different time periods and different application scenarios to more comprehensively verify and optimize the network model.

[0113] In some embodiments of the present application, before the meta-flow decision network is optimized based on the second data set, the method can further include the following steps:

[0114] According to the flow feature statistical data of each second example, determine the noise feature distribution corresponding to each second example;

[0115] For each second example, if it is determined that the current second example is a noise sample according to the noise feature distribution corresponding to the current second example, delete the current second example in the second data set.

[0116] For convenience of description, the above steps are combined for description.

[0117] In the embodiments of the present application, after obtaining the second data set, before optimizing the meta-flow decision network based on the second data set, the flow feature statistical data of each second example can be analyzed to determine the noise feature distribution corresponding to each second example. Noise can come from errors in the data collection process, sudden abnormalities in the network environment, etc. For example, the network state feature data (such as delay value) of a certain second example has an abnormal large fluctuation, which is obviously inconsistent with the network state feature data of other second examples in the same time period. Such abnormal fluctuation may be a manifestation of noise. Alternatively, the noise feature distribution corresponding to each second example can be determined by at least one of standardization processing, outlier detection, missing value filling, feature correlation analysis, principal component analysis and cluster analysis.

[0118] The standardization processing refers to standardizing each item included in the flow feature statistical data according to its reasonable range and data type. For example, the signal strength value is mapped to the [0, 1] interval, and the delay value is unified in milliseconds, etc.

[0119] Anomaly detection is to detect anomalies in the data using statistical methods and machine learning algorithms. For continuous data (such as delay, bandwidth utilization, etc.), methods based on mean and standard deviation can be used, and values exceeding the mean plus or minus three times the standard deviation are considered as anomalies. For discrete data (such as encoding format, transmission protocol type, etc.), a whitelist of common values can be established, and values not in the whitelist are considered as anomalies.

[0120] Missing value filling is to fill the missing values with mean filling, median filling or using regression model for prediction filling. For a large proportion of missing values, the data acquisition method can be reconsidered or the data can be discarded.

[0121] Feature correlation analysis is to calculate the correlation between each flow feature statistical data. For example, analyze the correlation between network status feature data (such as signal strength, bandwidth utilization, delay, packet loss rate), and the correlation between these feature data and data content feature data (such as video resolution, frame rate) and dual-domain interaction behavior feature data. High correlation feature combination may indicate the presence of noise or redundant information.

[0122] Principal component analysis (PCA) is to apply principal component analysis method to reduce the dimension of standardized flow feature statistical data. Through principal component analysis, multiple related features can be converted into a set of uncorrelated principal components. By observing the distribution of principal components, those second samples that deviate from the main distribution area in the principal component space may be abnormal samples affected by noise.

[0123] Clustering analysis is to use clustering algorithms (such as K-means clustering algorithm (K-Means), hierarchical clustering, etc.) to cluster flow feature statistical data. If some second samples cannot be clearly attributed to any cluster or belong to very small and isolated clusters, these second samples may carry noise features.

[0124] Based on the above feature analysis results, a noise prediction model can be constructed. Machine learning algorithms such as logistic regression, decision tree, random forest, etc. can be selected. Input features can include correlation indicators between features, positions of second samples in principal component space, distances to cluster centers, etc. The output of the model is a probability value indicating that the second sample is a noise sample.

[0125] A possible process of determining the distribution of noise features is as follows:

[0126] 1. Probability threshold setting

[0127] Set a noise probability threshold through cross-validation or based on experience. For example, preliminarily determine the second samples with a probability greater than or equal to 0.8 as possible noise samples.

[0128] 2. Time series analysis

[0129] For the shunt feature statistical data with time series properties (such as the change of network state features over time), time series analysis is performed. Check if the second sample has significant inconsistencies or mutations with other second samples in the previous and subsequent time periods. If there are, and they are determined by the noise prediction model as high noise probability, it is more likely to be a noise sample.

[0130] 3. Multi-dimensional evaluation

[0131] Consider the performance of the second sample in different feature dimensions. For example, a second sample behaves abnormally in network state feature data, while it also has a large difference with other second samples in data content feature data and dual-domain interaction behavior feature data. Then the second sample is more likely to be a noise sample.

[0132] 4. Domain knowledge combination

[0133] Introduce expert knowledge and experience rules in related fields. For example, according to the common sense of network engineers, some specific network state feature combinations are almost impossible to appear in reality, so the second samples that meet these impossible combinations can be considered as noise samples.

[0134] 5. Visualization analysis

[0135] Visualize the shunt feature statistical data, such as drawing two-dimensional or three-dimensional scatter plots, box plots, bar charts, etc. By visually observing the distribution of the data, find the data area that deviates from the normal distribution, and the second samples in these areas may be noise samples.

[0136] A possible verification and adjustment process of the determination method of noise feature distribution is as follows:

[0137] 1. Manual annotation verification

[0138] Select a part of the second samples determined as noise samples, and manually annotate and verify them by domain experts. According to the results of manual annotation, evaluate the accuracy of the noise detection algorithm, and adjust and optimize the algorithm.

[0139] 2. Dynamic adjustment

[0140] As new data is continuously added, continuously monitor the distribution of noise features. If it is found that the pattern of noise features has changed, update the parameters and thresholds of the noise detection model in time to adapt to the dynamic changes of data.

[0141] For example: Suppose in a real 5G dual-domain network environment, the offloading server is obtaining the noise feature distribution of offloading feature statistics.

[0142] For the delay data in the network state feature data, first, standardize the delay values to standard units. Then, through the outlier detection method, it is found that some delay values are far beyond the normal range. Further analysis of these abnormal delay data points finds that they are also abnormal in other feature data (such as signal strength, bandwidth utilization).

[0143] Through principal component analysis, it is found that these abnormal points deviate from the main distribution area in the principal component space. Then, using clustering algorithm to cluster the data, it is found that these abnormal points cannot be well attributed to any normal cluster.

[0144] Based on the above analysis results, input into the constructed noise prediction model. Set the noise probability threshold to 0.8, and preliminarily mark the second samples with probability greater than 0.8 as possible noise samples.

[0145] Time series analysis of these possible noise points finds that they have obvious mutations with delay data in the previous and subsequent time periods. At the same time, combined with domain knowledge, it is found that certain delay values combined with corresponding signal strength and bandwidth utilization are unreasonable in the actual network environment.

[0146] Through visual analysis, scatter plots are drawn with delay and signal strength as coordinate axes, and it is clearly seen that these noise points deviate from the distribution area of normal data.

[0147] Select a part of the second samples marked as noise samples for manual annotation verification, and find that the accuracy of the noise detection algorithm is 85%. According to the feedback of manual annotation, adjust the algorithm, such as adjusting the parameters of the noise prediction model or modifying the number of clusters of the clustering algorithm.

[0148] As new data continues to flow in, continuously monitor the distribution of noise features. When changes in the network environment are found to cause changes in the noise feature pattern, update the model parameters and thresholds in a timely manner to ensure accurate acquisition of the noise feature distribution.

[0149] For example, for the video encoding format in the data content feature data, by establishing a whitelist of common encoding formats, some rare encoding formats not in the whitelist are found. Further analysis of these second samples finds that they are also abnormal in other feature data.

[0150] Through principal component analysis and clustering analysis, it is found that these second samples are significantly different from normal second samples. Input into the noise prediction model, get the corresponding noise probability.

[0151] When time series analysis is performed, it is found that these second examples have a large deviation from historical data in the frequency of occurrence of encoding formats. In combination with visual analysis, in the distribution histogram of encoding formats, these rare encoding format second examples form isolated small columns.

[0152] Through manual annotation verification and dynamic adjustment, the method for obtaining noise feature distribution is continuously optimized, and the accuracy and reliability of noise detection are improved.

[0153] Suppose in another scenario, such as data shunting of intelligent transportation systems.

[0154] For communication data between vehicles and traffic facilities, some abnormal values of packet loss rate appear in network state feature data. After standardization processing, these abnormal packet loss rate second examples are marked by an outlier detection method.

[0155] By analyzing the performance of these second examples in data content feature data (such as the type and size of communication data) and dual-domain interaction behavior feature data (such as the relationship between the moving speed of the vehicle and the network connection), it is found that they are inconsistent with the normal data pattern.

[0156] Using principal component analysis and cluster analysis, the position and cluster of these abnormal packet loss rate second examples in the feature space are determined. Input into the noise prediction model, the probability of being a noise sample is evaluated.

[0157] Time series analysis is performed to observe the distribution of these abnormal packet loss rate second examples at different times, whether there is a sudden increase or decrease. In combination with domain knowledge such as traffic flow changes and vehicle driving rules, it is determined whether these second examples are noise samples.

[0158] Through visual analysis, scatter plots are drawn with packet loss rate and vehicle speed as coordinate axes to observe the distribution characteristics of noise samples.

[0159] After manual annotation verification and dynamic adjustment, the method for obtaining noise feature distribution can adapt to the characteristics and changes of intelligent transportation systems, and accurately identify noise samples.

[0160] For each second example, according to the noise feature distribution corresponding to the current second example, it can be determined whether the current second example is a noise sample. If the current second example is a noise sample (such as the second example of abnormal fluctuation of the delay value described above), the current second example can be deleted from the second data set, and if the current second example is not a noise sample, the current second example is retained in the second data set. The current second example is the second example targeted by the current operation.

[0161] The deletion of noise samples in the second data set helps to improve the accuracy and reliability of the second data set, avoiding the interference and misleading of network model training by noise samples, and thus improving the robustness and generalization ability of the network when optimizing the meta-flow decision network based on the second data set.

[0162] From the above, it can be seen that the convolutional neural network has at least one of the following roles in data flow:

[0163] Feature extraction: the convolutional neural network effectively extracts local and global patterns in the packet flow feature statistics through convolutional layers and pooling layers, which are crucial for understanding the nature of the data packet and the network environment;

[0164] Intelligent decision-making: based on the learned features, the convolutional neural network can predict the confidence of data packet matching different transmission paths, providing an intelligent decision basis for data flow;

[0165] Dynamic adaptation: through meta-cognitive learning and noise processing, the convolutional neural network can dynamically adapt to changes in the network environment and the influence of data noise, ensuring the accuracy and reliability of the flow decision;

[0166] Optimize network resources: through the intelligent flow decision of the convolutional neural network, network resources can be optimized, data transmission efficiency can be improved, and user experience can be improved.

[0167] The meta-flow decision network continuously learns and adjusts during the training process, and finally obtains a first data flow decision network with more optimized performance. The first data flow decision network can accurately predict the transmission path that should be selected according to the flow feature statistics of the target data packet, thereby improving the efficiency and quality of data flow.

[0168] In some embodiments of the present application, the second data flow decision network can be obtained by the following steps:

[0169] Obtain a third data set and a fourth data set, the third data set comprising flow feature statistics and allocation path labels of a plurality of third samples, and the fourth data set comprising flow feature statistics and allocation path labels of a plurality of fourth samples;

[0170] For each third sample in the third data set, input the flow feature statistics of the current third sample into each multi-layer perceptron network of the pre-constructed basic bio-inspired search cluster, respectively, to obtain the estimated confidence of the current third sample matching each transmission path output by each multi-layer perceptron network;

[0171] input the estimated confidence corresponding to each third sample output by each multi-layer perceptron network and the assigned path label corresponding to the respective third sample into a pre-constructed evaluation network model, to obtain an iteration weight distribution of each multi-layer perceptron network;

[0172] evolve the basic bio-inspired search cluster according to the iteration weight distribution of each multi-layer perceptron network;

[0173] In a case where a preset number of evolution rounds is reached, determine the multi-layer perceptron network with an iteration weight distribution greater than or equal to the first threshold in the evolved basic bio-inspired search cluster as a candidate machine learning network.

[0174] learn network parameters of the candidate machine learning network based on the fourth data set, and adjust the parameters of the candidate machine learning network according to the deviation between the output result of the candidate machine learning network corresponding to each fourth sample and the assigned path label corresponding to the respective fourth sample, to obtain a second data offloading decision network.

[0175] For convenience of description, the above steps are combined for description.

[0176] In the embodiments of the present application, the third data set and the fourth data set can be obtained for training the pre-constructed multi-layer perceptron network. The third data set includes the offloading feature statistical data and the assigned path label of a plurality of third samples, and the fourth data set includes the offloading feature statistical data and the assigned path label of a plurality of fourth samples. The third data set and the fourth data set cover various different scenarios and network conditions. For example, there are a large number of high-definition video stream data packets generated during a large-scale sports event live broadcast, at this time, the 5G mobile communication network and the Wi-Fi network both face high load; there are also a small amount of low-priority data packets generated at the early morning when the number of users is small, at this time, the network resources are relatively abundant. Each third sample and each fourth sample correspond to an assigned path label, which indicates which network domain the corresponding sample should preferentially transmit through in the current scenario or network state. Taking the data packets of the sports event live broadcast as an example, it may be labeled as preferentially transmitted through the Wi-Fi network to avoid excessive pressure on the 5G mobile communication network when the network is congested; and it may be labeled as transmitted through the 5G network to obtain faster transmission speed when the network is relatively idle.

[0177] The third data set can also be referred to as a third sample to-be-offloaded data packet sequence, and the fourth data set can also be referred to as a fourth sample to-be-offloaded data packet sequence. The fourth data set and the third data set can come from different application scenarios and time periods, and have more extensive representativeness. For example, the fourth data set includes real-time interactive data in online education courses, image transmission data in telemedicine, vehicle communication data in intelligent transportation systems, etc.

[0178] Optionally, the third data set and the fourth data set can be obtained from the historical data repository, or a plurality of third samples and a plurality of fourth samples are obtained from the historical data repository, and then each third sample and each fourth sample is analyzed respectively to obtain the shunt feature statistical data and the assigned path label of each third sample and the shunt feature statistical data and the assigned path label of each fourth sample.

[0179] A basic bio-inspired search cluster is pre-constructed, the basic bio-inspired search cluster comprising a plurality of multi-layer perceptron networks (or MLP network instances), and the basic weight parameters of each multi-layer perceptron network are configured differently.

[0180] Based on the third data set, each multi-layer perceptron network included in the pre-constructed basic bio-inspired search cluster is trained. Specifically, for each third sample in the third data set, the shunt feature statistical data of the current third sample can be input into each multi-layer perceptron network, and each multi-layer perceptron network can make a prediction based on the shunt feature statistical data of the current third sample to output an estimated confidence of the current third sample matching each transmission path, indicating the likelihood of the current third sample matching each transmission path. The current third sample refers to the third sample to which the current operation is directed.

[0181] For example, one of the multi-layer perceptron networks predicts that the estimated confidence of a certain third sample transmitted through the 5G mobile communication network is 0.8 and the estimated confidence of the third sample transmitted through the Wi-Fi network is 0.2 based on the input shunt feature statistical data of the third sample; another multi-layer perceptron network can give different prediction results, such as the estimated confidence of the third sample transmitted through the 5G mobile communication network is 0.6 and the estimated confidence of the third sample transmitted through the Wi-Fi network is 0.4.

[0182] The estimated confidence of each third sample output by each multi-layer perceptron network and the assigned path label corresponding to the corresponding third sample are input into a pre-constructed evaluation network model, the performance of each multi-layer perceptron network is evaluated through the evaluation network model, and the iteration weight distribution of each multi-layer perceptron network can be determined according to the output result of the evaluation network model. For any multi-layer perceptron network, if the prediction result of the multi-layer perceptron network is more consistent with the actual situation, the iteration weight distribution assigned to the multi-layer perceptron network is larger, and vice versa, if the prediction result of the multi-layer perceptron network deviates more from the actual situation, the iteration weight distribution assigned to the multi-layer perceptron network is smaller. That is, the larger the iteration weight distribution of the multi-layer perceptron network, the better the performance of the multi-layer perceptron network in the current training round and the more credible the prediction result.

[0183] According to the iteration weight distribution of each multi-layer perceptron network, fusion operations, evolution operations, update operations, etc. can be performed to realize the evolution of the basic bio-inspired search cluster.

[0184] In the case of reaching the preset evolution round, the iteration weight distribution of each multi-layer perceptron network in the evolved basic bio-inspired search cluster can be determined, and according to the iteration weight distribution of each multi-layer perceptron network, the multi-layer perceptron networks are screened, and the multi-layer perceptron networks with iteration weight distribution greater than or equal to the first threshold value are screened out and determined as candidate machine learning networks, ready to enter the next round of learning and optimization. The first threshold value can be set and adjusted according to actual conditions, such as setting it to 0.7.

[0185] Based on the fourth data set, the candidate machine learning network can be learned for network parameters, and according to the deviation between the output result of each fourth example corresponding to the candidate machine learning network and the corresponding fourth example corresponding to the assigned path label, the parameters of the candidate machine learning network are constantly adjusted. After multiple iterations and learning, a second data flow decision network is obtained. For each fourth example, the output result of the candidate machine learning network corresponding to the current fourth example can be the estimated confidence of matching each transmission path of the current fourth example.

[0186] Among them, the related operations involved in the cluster evolution process in the training process include:

[0187] 1) Fusion operation: multi-layer perceptron networks with higher iteration weight distribution play a greater role in the fusion operation, and their parameters and structures are more borrowed and integrated into new network structures. Similar to the genes of excellent individuals in biological populations being more preserved and inherited.

[0188] Specifically, the fusion operation includes:

[0189] 1. Weight sorting and grouping

[0190] First, sort the iteration weight distribution of all multi-layer perceptron networks in descending order. Then, according to the iteration weight distribution, these multi-layer perceptron networks are divided into several groups. For example, the top 20% of multi-layer perceptron networks with the highest iteration weight distribution are divided into a group, the next 30% of multi-layer perceptron networks are divided into a group, and so on.

[0191] 2. Feature extraction and sharing

[0192] For each group, extract the key features and parameters of each multi-layer perceptron network. These features can include neuron connection mode, weight value distribution, selection of activation function, etc. Then, share and exchange features within the group, so that each multi-layer perceptron network can learn the superior features of other multi-layer perceptron networks in the same group.

[0193] 3. Knowledge fusion

[0194] Based on the shared features, each multi-layer perceptron network fuses the knowledge of other multi-layer perceptron networks with certain rules and proportions. For example, for the features of multi-layer perceptron networks with higher iteration weight distribution, a higher fusion proportion is given. The specific fusion method can be weighted averaging of weight values, or selective merging of neuron connections.

[0195] 4. New network generation

[0196] Through the fusion operation, a batch of new multi-layer perceptron networks are generated. These multi-layer perceptron networks combine the advantages of multiple multi-layer perceptron networks in the group, and have more comprehensive and optimized structure and parameters.

[0197] 2) Evolution operation: according to the iteration weight distribution, the structure and parameters of the multi-layer perceptron network are adjusted and mutated adaptively to explore better solutions. Similar to the mutation and natural selection in biological evolution process.

[0198] Specifically, the evolution operation includes:

[0199] 1. Mutation operation

[0200] For the new multi-layer perceptron networks generated by the fusion operation, random mutation operation is introduced. This can include randomly adjusting the connection weights of some neurons, changing the type of activation function, or adding / deleting some neuron connections. The amplitude and probability of mutation can be dynamically adjusted according to the weights and performance of the multi-layer perceptron network. The iteration weight distribution of the multi-layer perceptron network is higher, the mutation amplitude is smaller, to maintain its stability; while the iteration weight distribution of the network is lower, the mutation amplitude is larger, to increase the possibility of exploring new structure.

[0201] 2. Cross operation

[0202] Some new multi-layer perceptron networks are selected for cross operation. Specifically, part of the structure and parameters of two multi-layer perceptron networks are randomly selected to exchange and combine. For example, exchange the connection mode and weight value of some hidden layers of two multi-layer perceptron networks, or combine the output layer of one multi-layer perceptron network with the hidden layer of another multi-layer perceptron network. Cross operation helps to produce completely new network structure and parameter combination.

[0203] 3. Adaptive adjustment

[0204] During the evolution process, the structure and parameters of the multi-layer perceptron network are adjusted adaptively according to the current network state feature data and data content feature data. For example, if the current data content is mainly high-resolution image data, the number and parameters of the layers related to image processing in the multi-layer perceptron network can be increased; if the network state shows high delay, the structure of the multi-layer perceptron network can be adjusted to reduce the amount of calculation and transmission time.

[0205] 4. Performance evaluation and selection

[0206] The multi-layer perceptron networks that have undergone mutation and crossover operations are quickly evaluated for performance. The evaluation can be based on some simple sample data and indicators such as accuracy, recall rate, etc. According to the evaluation results, the multi-layer perceptron networks with better performance are selected to enter the next round of evolution operation, while the multi-layer perceptron networks with significantly poor performance are eliminated.

[0207] 3) Update operation: eliminate multi-layer perceptron networks with poor performance, and introduce new network structures and parameter configurations with potential. Similar to the survival of the fittest mechanism in biological populations.

[0208] Specifically, the update operation includes:

[0209] 1. Weight recalculation

[0210] For the multi-layer perceptron networks after the evolution operation, the iterative weight distribution is recalculated according to their performance in the new evaluation. The calculation of the iterative weight distribution considers multiple indicators such as accuracy, recall rate, F1 value, etc., and combines the complexity of the multi-layer perceptron network and the consumption of computing resources and other factors for comprehensive evaluation.

[0211] 2. Network structure optimization

[0212] The structure and parameters of the multi-layer perceptron network are analyzed to find out possible redundancies and inefficiencies. For example, some connections with always small iterative weight distribution values can be deleted to simplify the network structure; some overly complex layers can be compressed and optimized.

[0213] 3. Data augmentation and retraining

[0214] In order to further improve the performance and generalization ability of the multi-layer perceptron network, data augmentation techniques are used to expand and transform the training data. For example, image data is flipped, rotated, scaled, etc., and text data is replaced with synonyms, randomly deleted, etc. Then, the enhanced data set is used to retrain the multi-layer perceptron network, and the parameters of the multi-layer perceptron network are updated.

[0215] 4. Model fusion and selection

[0216] The multiple updated multi-layer perceptron networks are fused to form a comprehensive model. The fusion method can be weighted averaging of the prediction results of the multiple multi-layer perceptron networks, or using an ensemble learning method such as random forest, adaptive boosting (Adaboost), etc. Then, according to the final performance evaluation, the optimal multi-layer perceptron network is selected as the updated cluster member.

[0217] 5. Feedback and adjustment

[0218] During the entire updating operation process, the parameters and strategies of each operation are adjusted according to the feedback information by continuously collecting and analyzing data. For example, if it is found that a certain mutation operation or crossover method often leads to performance improvement, the frequency of its application can be increased; if a certain data enhancement method is not obvious, the method can be reduced or replaced.

[0219] Through the continuous circulation and optimization of the above fusion operation, evolution operation and updating operation, the continuous evolution of the basic bio-inspired search cluster including multiple multi-layer perceptron networks can be realized, so that it can better adapt to the complex and variable data distribution requirements in the multi-domain network environment, and improve the accuracy and efficiency of data distribution decision-making.

[0220] Suppose that in the scenario of online video playing, the above cluster evolution operation is being performed.

[0221] In the fusion operation, several multi-layer perceptron networks with larger iteration weight distribution all use larger convolution kernels and deeper network structures when processing high-definition video stream data. Therefore, in the generation of new multi-layer perceptron networks, more of these features are fused, so that the new multi-layer perceptron networks have better performance when processing high-definition video data.

[0222] In the mutation operation of the evolution operation, for some multi-layer perceptron networks with innovative structure but low iteration weight distribution, the neuron connections are randomly adjusted to a large extent. For example, a part of neurons originally used to process audio features are reconnected to process video frame rate features, which may improve the accuracy of video smoothness judgment.

[0223] In the crossover operation, part of the hidden layers of a multi-layer perceptron network that is good at processing 4K video are combined with a multi-layer perceptron network that performs well in processing low-resolution video, and the new multi-layer perceptron network generated can have good adaptability in processing video data of different resolutions.

[0224] In the update operation, the multi-layer perceptron network trained through certain data enhancement methods (such as random cropping and scaling of video frames) is more accurate in shunting decisions when dealing with network bandwidth fluctuations. Therefore, in subsequent updates, more data enhancement methods are applied, and the parameters of the enhancement are continuously adjusted according to the real-time performance feedback of the multi-layer perceptron network.

[0225] For example, in a remote medical scenario, a large amount of medical image and real-time monitoring data transmission is involved.

[0226] In the fusion operation, the multi-layer perceptron network with higher iteration weight distribution has similar feature extraction methods when processing specific medical image formats (such as Digital Imaging and Communications in Medicine (DICOM)). Therefore, these effective feature extraction modules are strengthened in the new multi-layer perceptron network.

[0227] In the evolution operation, due to the high requirements of medical data on accuracy and security, the mutation operation is relatively cautious, mainly focusing on fine-tuning the parameters of the multi-layer perceptron network and optimizing the connection structure. The crossover operation focuses on exchanging part of the structure of multi-layer perceptron networks that process different types of medical data (such as X-ray images and electrocardiogram data) to improve the comprehensive processing ability of multi-layer perceptron networks for various medical data.

[0228] In the update operation, according to the characteristics and regulatory requirements of medical data, the security and privacy protection mechanisms of the multi-layer perceptron network are optimized and updated, and the structure and parameters of the network are continuously adjusted to adapt to the data shunting needs of different medical institutions and network environments.

[0229] Through continuous practice and adjustment in various practical scenarios, the basic bio-inspired search cluster including multiple multi-layer perceptron networks can continuously evolve and improve, providing more intelligent and efficient decision support for data shunting in multi-domain network environments, ensuring the data transmission quality and user experience of medical data, video playback, and other applications.

[0230] Suppose in a smart traffic system scenario, a large amount of communication data between vehicles and traffic facilities needs to be processed.

[0231] In the fusion operation, for vehicle control instruction data with extremely high real-time requirements, multi-layer perceptron networks with higher iteration weight distribution all use fast forward propagation algorithms and simplified network structures. The newly generated multi-layer perceptron network inherits more of these efficient algorithms and structural characteristics.

[0232] In the mutation operation of the evolution operation, for some multi-layer perceptron networks that perform poorly in complex traffic scenarios (such as intersections and highway congestion sections), the activation thresholds of their neurons are randomly adjusted, which may better cope with data shunting strategies for sudden traffic conditions.

[0233] In the crossover operation, the multi-layer perceptron network that performs well in urban road environments is partially structurally crossed with the multi-layer perceptron network that performs well in highway environments, and the resulting new multi-layer perceptron network can make more reasonable data shunting decisions in different types of road scenarios.

[0234] In the update operation, the learning rate and weight decay coefficient of the multi-layer perceptron network are dynamically adjusted according to the periodic changes of traffic flow and the network load in different time periods. At the same time, new traffic data samples are introduced to retrain the multi-layer perceptron network, constantly improving the adaptability and accuracy of the multi-layer perceptron network in intelligent transportation systems.

[0235] For example, in an industrial Internet of Things scenario, sensor data and control instructions from various industrial devices need to be processed.

[0236] In the fusion operation, for industrial control data with extremely high stability requirements, focus on fusing network features that are stable and have low misjudgment rates in long-term operation.

[0237] In the mutation operation of the evolution operation, considering the complexity and diversity of industrial environments, the input layer of the multi-layer perceptron network is diversified to adapt to different types and specifications of sensor data.

[0238] In the crossover operation, the multi-layer perceptron network that processes device data in high-temperature environments is crossed with the multi-layer perceptron network that processes device data in low-temperature environments, so that the new multi-layer perceptron network can cope with a wider range of industrial environmental conditions.

[0239] In the update operation, according to the maintenance cycle of the device and the adjustment of the production plan, the training data and network parameters are updated in a timely manner to ensure that the data shunting decisions of the multi-layer perceptron network in the industrial Internet of Things environment remain efficient and accurate.

[0240] That is, by using the above technical solutions, the evolution of the basic bio-inspired search cluster can be continuously promoted, and it can play an excellent data shunting decision-making capability in various multi-domain network application scenarios, providing high-quality network services and data transmission support for different industries and fields.

[0241] From the above, the multi-layer perceptron network at least has the following role in data shunting:

[0242] Complex feature learning: Multi-layer perceptron networks can learn complex patterns in packet splitting feature statistics through multiple hidden layers and activation functions, which are crucial for understanding the characteristics of packets and network environments.

[0243] Simulating biological intelligence: Multi-layer perceptron networks are optimized by ant colony algorithms to simulate the intelligent behavior of biological groups to make decisions, such as the behavior of ants finding the optimal path, making data splitting decisions more efficient.

[0244] Dynamic parameter adjustment: Through the ant colony algorithm, the multi-layer perceptron network can adaptively adjust the parameters to find the optimal solution to adapt to the changing network environment.

[0245] Cluster evolution: Through fusion, evolution, and update operations, the cluster of multiple multi-layer perceptron networks continuously evolves, improving the accuracy and efficiency of data splitting decisions.

[0246] Based on the third and fourth data sets, the multi-layer perceptron network is iterated and parameter-adjusted multiple times based on biological heuristic algorithms, and the cluster evolves, effectively handling complex data splitting problems and obtaining a second data splitting decision network with high prediction accuracy. This way, when applying the second data splitting decision network, more accurate prediction results can be obtained, ensuring the efficiency and quality of data transmission, and helping to improve the user's network experience.

[0247] For ease of understanding, the following will take the actual deployment application scenario of a smart park as an example to explain the technical solutions provided by the embodiments of the present application.

[0248] In a smart park, the network infrastructure is usually composed of public 5G base stations and park-owned high-density Wi-Fi systems. 5G base stations achieve full coverage of indoor and outdoor areas in the park, providing high bandwidth and low latency for various scenarios such as park security, personnel office, and visitor reception. At the same time, the Wi-Fi system is deployed with distributed access points (APs) in conference rooms, office areas, and important production areas to ensure local high-speed data transmission and flexible access. The park core routing and management server is responsible for access authentication, terminal behavior monitoring, splitting policy issuance, and real-time traffic scheduling, and the management node with edge computing capabilities can assist in decision-making and response speed.

[0249] Various intelligent terminal devices are distributed in the park, for example, intelligent cameras, access control, visitor tablets, employee mobile phones, wearable devices, robots, AGVs and a large number of IoT sensors. These terminal devices generally support 5G and Wi-Fi dual-mode networking functions, with automatic negotiation and switching capabilities. Taking a newly deployed intelligent camera as an example, when it first goes online, it will first complete identity authentication (such as Subscriber Identity Module (SIM) authentication and enterprise Wi-Fi 802.1X or Media Access Control (MAC) authentication) through 5G and Wi-Fi respectively. After authentication, the camera will automatically obtain two independent IP addresses, and report its unique identifier (such as the combination of International Mobile Equipment Identity (IMEI) and MAC) to the park core management server, and the system will establish a multi-network state tracking file for the terminal device in the background accordingly.

[0250] In actual operation, all access terminal devices are under real-time monitoring of the management server and the edge node. The system not only actively collects network and terminal device state data regularly (such as every second), but also triggers feature collection immediately when certain events occur. Specifically, the system pays attention to very comprehensive data features, including but not limited to the signal strength of 5G and Wi-Fi (such as Received Signal Strength Indicator (RSSI), Reference Signal Received Power (RSRP)), real-time bandwidth and remaining bandwidth of wireless link, round-trip time (RTT) of each path, packet loss rate, and motion state of terminal device (such as device stationary, moving, acceleration change). At the same time, the system also collects business layer features, such as data packet type, service priority (such as high-definition video, IoT regular reporting, emergency voice, etc.), port and protocol type (Dots Per Inch (DPI)), and device five-minute history switching behavior, network exception record, etc. The park management system also synchronously collects the current load of each AP and base station, network congestion state, to provide global reference for subsequent shunting decision. All these feature data will be automatically normalized and cleaned to remove outliers, noise and false positives, to ensure the stability and accuracy of subsequent intelligent analysis.

[0251] In a regular case, the offloading decision mechanism runs periodically, such as once per second. In addition, once the system detects a mutation of a key feature, for example, a sudden drop in signal strength, a sudden increase in packet loss rate, high-speed movement of terminal equipment, a surge in traffic data volume, etc., it will trigger the reevaluation and adjustment of the offloading strategy in real time. The normalized multi-dimensional features are organized into a feature vector, which is then input into two adaptive decision networks in the system (the first data offloading decision network and the second data offloading decision network).

[0252] The first data offloading decision network adopts a convolutional neural network (CNN) combined with a meta-cognition learning algorithm, which can automatically capture local changes and short-term mutations between features, and is suitable for making offloading decisions quickly when the camera encounters sudden scenes (such as high-flow security alarms, link jitter). The meta-cognition part continuously monitors the deviation between network prediction and actual business performance, automatically adjusts its weights and model parameters, and thus improves the response capability to uncertainty and sudden conditions. The second data offloading decision network is based on a multi-layer perceptron (MLP) network combined with an ant colony optimization algorithm, focusing on global optimization in long-term operation, such as comprehensive analysis of traffic distribution trends under different businesses, different terminal types, and different time periods, dynamic adjustment of network parameters, and guarantee of overall optimization and adaptive ability in daily operation. These two networks respectively output the offloading confidence of each transmission path (such as 5G, Wi-Fi). For example, the first data offloading decision network may determine that the offloading confidence of selecting 5G is 0.65 and the offloading confidence of selecting Wi-Fi is 0.35, while the second data offloading decision network outputs the offloading confidence of selecting 5G may be 0.48 and the offloading confidence of selecting Wi-Fi may be 0.52. The offloading confidence is not a set of static values, but will be further refined according to different business types, such as high-definition video, voice, IoT regular reporting, low-latency business, and ordinary file download. Each business will have an independent set of offloading confidence, which facilitates more detailed and reasonable offloading strategies.

[0253] The system will comprehensively analyze the distribution confidence through multi-model fusion methods, such as weighted average, priority weighting, etc., to dynamically generate the final distribution recommendation. For example, the management server can fine-tune the weights of the two networks according to the current business priority and historical performance, and finally determine that the distribution confidence of the main stream of the camera selecting 5G is 70%, the distribution confidence of selecting Wi-Fi is 30%, and the distribution confidence of the auxiliary stream selecting Wi-Fi is 100%. The generated distribution strategy is automatically distributed to the terminal device, and the terminal device's distribution client (Agent) or embedded software development kit (Software Development Kit, SDK) adjusts the data upload method accordingly - the main stream real-time 5G to ensure the high reliability and low delay of security data, and the auxiliary stream or backup stream fully utilizes the Wi-Fi bandwidth or supplements when the Wi-Fi link is restored, greatly improving the stability of the main business and the utilization rate of network resources.

[0254] During the whole process, both the terminal side and the network side will monitor the link performance of the actual business in real time, including the frame loss rate, end-to-end delay, real-time packet loss, and other core indicators of the main stream and auxiliary stream. If any link exception occurs (such as 5G temporary disconnection, Wi-Fi sudden congestion), the system will automatically switch all traffic back to the available path, record the exception event and trigger the management background warning. The terminal device such as camera is generally configured with local cache capability, which automatically caches video clips during short-term network jitter, and automatically supplements after the link is restored, ensuring the integrity of video and data stream. All distribution decisions, network adjustments and effect monitoring data will be periodically returned to the management background and enter the big data analysis platform, providing real samples for the continuous optimization and training of the model, forming a complete closed loop of distribution-monitoring-feedback-self-learning.

[0255] In actual park pilots, through the above method, not only can the key business such as high-definition video stream, intelligent security, production scheduling, etc. always have the best network experience, but also greatly improves the utilization rate and management efficiency of the whole network. According to statistics, the main stream frame loss rate has decreased by more than 20%, the business delay has been reduced by about 25%, and the dynamic balance of Wi-Fi and 5G resource utilization has significantly reduced the unit business cost. The system background visual interface also supports administrators to remotely monitor, intervene and adjust the strategy, further enhancing the flexibility of operation and maintenance and the reliability of business.

[0256] In actual business deployment, how to reasonably fuse the prediction results of the two sets of data shunting decision networks is a key link for the landing of the intelligent shunting system. At each time of shunting decision, the shunting confidence of 5G and Wi-Fi output by the two sets of data shunting decision networks is combined with the business type, terminal device historical performance and real-time network state, and a dynamic weighted fusion strategy is used for fusion. For example, when facing high-definition video streaming and other services with extremely high requirements for bandwidth and stability, the system usually gives higher weight to the judgment of the first data shunting decision network, and for the IoT scene of periodic small data packet reporting, the overall optimal judgment of the second data shunting decision network is preferentially referred to. The fusion method can be weighted average, and the weight of both sides will be dynamically adjusted according to the scene. In addition, for some special services with strong real-time performance and high priority, the system also supports priority weighting, allowing to forcibly specify that the business traffic always preferentially goes through a certain path, ensuring the reliability of business in emergency situations. For extremely rare scenes, if the judgments of the two sets of data shunting decision networks differ too much, the system will trigger the confidence interval mechanism, temporarily switching to manual intervention or automatic retraining strategy, to ensure that the fusion judgment always fits the actual effect.

[0257] The shunting decision result after fusion is often not simply selecting one path, but giving the most appropriate shunting proportion for each business flow, each session and even each data packet. For example, for high-definition video conference, the system may recommend that 70% of the data packets preferentially pass through the 5G channel for transmission, and the remaining 30% go through Wi-Fi. For services supporting multi-link, shunting will be accurate to the level of packet or multi-session, such as multi-path transmission control protocol (MPTCP), multi-access point name (APN), virtual extended local area network (VxLAN) and other multi-session mechanisms, which can intelligently allocate each piece of business traffic to different paths, realizing true network cooperation and resource balance.

[0258] After the offloading decision is determined, the system generates specific offloading and switching instructions at the core server or edge computing node, which explicitly indicate path priority, offloading ratio, policy validity period, special service identification, and other parameters. The instructions are issued to the offloading client of the terminal device through a secure encrypted channel, or sent to the access control device on the network side, ensuring accurate policy execution. After receiving the new policy, the terminal device can automatically rebuild the network session, such as switching the main connection, adjusting the packet routing, or opening or closing sub-streams under multi-path transmission. During execution, the terminal device also regularly reports the execution status, and the system can monitor the offloading effect in real time. If there are unexpected situations such as switching failure, abnormal increase in packet loss rate, etc., the terminal side and the network side can automatically trigger policy rollback or re-optimization adjustment, maximizing the continuous and stable operation of the business.

[0259] The entire system architecture can be seamlessly integrated with various network devices such as 5G core networks, enterprise Wi-Fi controllers, software-defined networks (SDN) / switches, etc. of related networks, supporting mainstream interface protocols such as Network Configuration Protocol (NETCONF) and OpenFlow, and facilitating connection with existing network management platforms of operators.

[0260] In specific business scenarios, such as an employee initiating a high-definition video conference, the system first collects the 5G and Wi-Fi link characteristics of the terminal device in real time, and inputs the characteristic vector into two sets of data offloading decision networks. The first data offloading decision network is more sensitive to sudden changes in bandwidth, and the second data offloading decision network is better at assessing stability and congestion risk. The final fusion result shows that 5G is more optimal, so 70% of the main stream data flow is allocated to 5G, and the remaining part is offloaded through Wi-Fi in parallel. If the Wi-Fi link stability improves significantly, the offloading ratio will be automatically and dynamically adjusted to maximize the saving of 5G resources; conversely, if the 5G signal drops suddenly, the main stream will automatically switch to Wi-Fi, ensuring uninterrupted video conference.

[0261] For example, in the scenario of high-speed movement of AGV robots, the system can predict in advance that there may be a Wi-Fi signal blind area based on historical data, and then preferentially switch the key control signal to the 5G path. The entire switching process is transparent to the business, ensuring uninterrupted AGV movement and continuous and efficient operation of the production line.

[0262] For massive IoT periodic data reporting, the system monitors the real-time bandwidth and terminal behavior of the current network, and usually preferentially offloads all low-latency, low-bandwidth services to Wi-Fi, only invoking 5G resources for a short time when important alarms occur, reducing the long-term pressure on 5G bandwidth and improving the utilization rate of Wi-Fi.

[0263] All offloading decisions and service execution situations are returned to the core server in real time and enter the big data analysis platform as samples for continuous training and weight adjustment of the model. The system supports automatic attribution and closed-loop optimization. For example, once abnormal switching, increased packet loss, service interruption, and other phenomena are detected, the system can automatically analyze the causes and guide the continuous evolution of the model and strategy. This mechanism not only ensures that each offloading decision is scientific and reasonable, but also allows the system to continuously "evolve" with changes in business and network environment, always maintaining the best state.

[0264] Taking real-time uploading of high-definition video streams from intelligent cameras as an example, the entire business process embodies the full-link intelligence of the system. The camera offloading client regularly collects information such as local 5G, Wi-Fi signals, delay, bandwidth, and synchronizes AP and base station load conditions with the management center. After standardizing these features, they are sent to the dual-data offloading decision network. The two data offloading decision networks respectively output which path is optimal for the main code stream and how to allocate the auxiliary code stream / redundant stream. The offloading strategy derived from model fusion is issued to the terminal device client with clear parameters. The camera main code stream prefers 5G, and the auxiliary code stream prefers Wi-Fi. Quality of Service (QoS) and Forward Error Correction (FEC) are simultaneously turned on. The offloading strategy also dynamically adjusts according to actual network load and service performance. If 5G is congested, it automatically increases the Wi-Fi proportion. When Wi-Fi improves, some traffic is migrated back. The entire process does not require human intervention.

[0265] Whether it is AGV robot cross-zone switching or employee mobile office, IoT terminal periodic reporting, the system can combine the real-time needs and historical performance of different businesses to perceive risks in advance, reasonably offload, and ensure that critical businesses do not interrupt. Actual pilot data shows that terminal automatic offloading and switching have triggered tens of thousands of times, core business end-to-end latency has been significantly reduced, packet loss rate has also decreased significantly, user satisfaction has continuously improved, and business complaints have significantly decreased. All of this is inseparable from the system's underlying dynamic self-learning mechanism and multi-tenant support, which not only adapts to various park network environments, but also flexibly responds to different customers' business strategies, achieving truly intelligent offloading and efficient operation and maintenance.

[0266] The whole system is also very convenient to deploy and maintain. All core servers integrate shunting decision modules, and terminal sides are pre-installed with shunting clients, all of which support remote upgrading. Administrators can master the shunting state, network health, and abnormal early warning of all terminals in real time through the background, and can manually intervene or adjust the strategy. All shunting and abnormal events have automatic log retention, providing a reliable data basis for subsequent operation and audit traceability. After a period of actual operation, the system not only stably supports the access and business operation of large-scale intelligent cameras, employee terminals, and IoT devices, but also has no obvious business interruption in switching and shunting, and the overall network utilization is improved, and the user experience is improved.

[0267] Overall, the embodiments of the present application realize efficient and intelligent processing of target data packets in a multi-domain network environment by constructing a first data shunting decision network and a second data shunting decision network. Specifically:

[0268] 1. By extracting data content feature data, transmission protocol feature data, network state feature data, and dual-domain interaction behavior feature data, etc. Shunting feature statistical data of target data packets provides comprehensive decision basis for data shunting. These data features cover the content, transmission method, current network environment and behavior pattern of data packets in the multi-domain network, ensuring the accuracy and comprehensiveness of the shunting decision;

[0269] 2. The meta-cognition learning algorithm is used to learn the parameters of the convolutional neural network to obtain the first data shunting decision network, and the biological heuristic algorithm is used to learn the parameters of the multi-layer perceptron network to obtain the second data shunting decision network. The first data shunting decision network continuously optimizes its internal parameters through meta-cognition learning, improving the intelligence and accuracy of the decision; the second data shunting decision network simulates the intelligent behavior of biological groups, adjusts the parameters through the biological heuristic algorithm, and enhances the adaptability and robustness of the network. The combination of the two data shunting decision networks improves the diversity and reliability of data shunting decision;

[0270] 3. The influence of noise features is fully considered in the training process, and noise data is identified and removed through feature standardization, outlier detection, principal component analysis, clustering analysis, etc. Ensuring the purity of the training data and the accuracy of the network model. At the same time, by dynamically adjusting the parameters and thresholds of the noise detection model, it adapts to the needs of data dynamic changes, further improves the generalization ability and stability of the network;

[0271] 4. Through multiple rounds of evolution and optimization, a high-performance candidate machine learning network is gradually screened out, and a high-efficiency and accurate second data shunting decision network is finally generated. The network can intelligently match the best intelligent processing path according to the characteristics of different data packets and network environment, ensuring the efficiency and quality of data transmission and improving the user's network use experience.

[0272] That is, through the construction of an adaptive data shunting decision network, the extraction of comprehensive shunting feature statistical data, the application of advanced machine learning algorithms, the identification and elimination of noise data, and multiple rounds of optimization and evolution, efficient and intelligent processing of target data packets in a multi-domain network environment is realized. The accuracy and efficiency of data shunting are improved, and the adaptability and robustness of the network model are enhanced, providing users with a more high-quality network service experience.

[0273] Corresponding to the above method embodiments, the embodiments of the present application also provide a data shunting device 300 applied to a shunting server deployed on the network side. The data shunting device described below can be referred to in conjunction with the data shunting method described above.

[0274] Referring to Figure 3 As shown in the figure, the device includes the following modules:

[0275] The feature data acquisition module 310 is configured to acquire shunting feature statistical data of a target data packet to be sent by a terminal device.

[0276] The first confidence obtaining module 320 is configured to input the shunting feature statistical data of the target data packet into a first data shunting decision network to obtain a first shunting confidence of the target data packet matching each transmission path. The first data shunting decision network is obtained by learning network parameters of a pre-constructed convolutional neural network according to a meta-cognition learning algorithm.

[0277] The second confidence obtaining module 330 is configured to input the shunting feature statistical data of the target data packet into a second data shunting decision network to obtain a second shunting confidence of the target data packet matching each transmission path. The second data shunting decision network is obtained by learning network parameters of a pre-constructed multi-layer perception network according to a biological heuristic algorithm.

[0278] The shunting strategy determination module 340 is configured to determine a target shunting strategy of the target data packet based on the first shunting confidence and the second shunting confidence of the target data packet matching each transmission path.

[0279] The device provided in the embodiments of the present application is used to obtain the shunt feature statistical data of the target data packet to be sent by the terminal device, and based on the shunt feature statistical data, the first data shunt decision network and the second data shunt decision network are used for prediction respectively, the first shunt configuration probability and the second shunt configuration probability of the target data packet matching each transmission path are obtained, and based on the first shunt configuration probability and the second shunt configuration probability of the target data packet matching each transmission path, the target shunt strategy of the target data packet is determined. The first data shunt decision network is obtained by learning the network parameters of the pre-constructed convolutional neural network according to the meta-cognition learning algorithm, and the second data shunt decision network is obtained by learning the network parameters of the pre-constructed multi-layer perception network according to the biological heuristic algorithm. The prediction results of the transmission paths by different data shunt decision networks can accurately and effectively determine the shunt strategy of the uplink data packet of the terminal device, and improve the data transmission efficiency.

[0280] In some embodiments of the present application, the convolutional neural network comprises a first input layer, a plurality of convolutional layers, a pooling layer, a fully connected layer and a first output layer connected in sequence.

[0281] In some embodiments of the present application, the first network obtaining module is further used to obtain the first data shunt decision network by the following steps:

[0282] The first data set and the second data set are obtained, the first data set comprises the shunt feature statistical data and the allocation path label of a plurality of first samples, and the second data set comprises the shunt feature statistical data and the allocation path label of a plurality of second samples;

[0283] The shunt feature statistical data of each first sample in the first data set is input into the pre-constructed convolutional neural network, the weights and parameters of the convolutional neural network are adjusted according to the deviation between the output result of the convolutional neural network corresponding to each first sample and the allocation path label corresponding to the corresponding first sample, and a meta-shunt decision network completing meta-cognition learning is obtained;

[0284] The shunt feature statistical data of each second sample in the second data set is input into the meta-shunt decision network, the parameters of the meta-shunt decision network are adjusted according to the deviation between the output result of the meta-shunt decision network corresponding to each second sample and the allocation path label corresponding to the corresponding second sample, and the first data shunt decision network is obtained.

[0285] In some embodiments of the present application, the first obtaining module is further used to:

[0286] Before the meta-shunt decision network is optimized based on the second data set, the noise feature distribution corresponding to each second sample is determined according to the shunt feature statistical data of each second sample;

[0287] For each second sample, in a case that the current second sample is determined as a noise sample according to the noise feature distribution corresponding to the current second sample, deleting the current second sample from the second data set.

[0288] In some embodiments of the present application, the multi-layer perceptron network comprises a second input layer, a plurality of hidden layers and a second output layer connected in sequence.

[0289] In some embodiments of the present application, the second network obtaining module is further configured to obtain the second data flow decision network by the following steps:

[0290] obtaining a third data set and a fourth data set, the third data set comprising flow feature statistics and allocation path labels of a plurality of third samples, and the fourth data set comprising flow feature statistics and allocation path labels of a plurality of fourth samples;

[0291] For each third sample in the third data set, inputting the flow feature statistics of the current third sample into each multi-layer perceptron network of the pre-constructed basic bio-inspired search cluster respectively, to obtain an estimated confidence of the current third sample matching each transmission path output by each multi-layer perceptron network;

[0292] inputting the estimated confidence of each third sample corresponding to each multi-layer perceptron network output into the pre-constructed evaluation network model, and inputting the allocation path label corresponding to the corresponding third sample into the pre-constructed evaluation network model, to obtain an iteration weight distribution of each multi-layer perceptron network;

[0293] evolving the basic bio-inspired search cluster according to the iteration weight distribution of each multi-layer perceptron network;

[0294] in a case that a preset evolution round is reached, determining the multi-layer perceptron network with an iteration weight distribution greater than or equal to a first threshold in the evolved basic bio-inspired search cluster as a candidate machine learning network;

[0295] based on the fourth data set, performing network parameter learning on the candidate machine learning network, adjusting parameters of the candidate machine learning network according to a deviation between an output result of the candidate machine learning network corresponding to each fourth sample and the allocation path label corresponding to the corresponding fourth sample, and obtaining the second data flow decision network.

[0296] In some embodiments of the present application, the flow strategy determination module 340 is specifically configured to:

[0297] fusing the first distribution confidence and the second distribution confidence corresponding to each transmission path to obtain a distribution confidence corresponding to each transmission path;

[0298] determining a target flow strategy of the target data packet based on the distribution confidence corresponding to each transmission path.

[0299] In some embodiments of this application, the traffic splitting strategy determination module 340 is specifically used for:

[0300] For each transmission path, the average of the first and second sub-configuration confidence scores corresponding to the current transmission path, or the weighted average of the first and second sub-configuration confidence scores corresponding to the current transmission path, is determined as the sub-configuration confidence score corresponding to the current transmission path. The current transmission path is the transmission path targeted by the current operation.

[0301] In some embodiments of this application, a traffic splitting strategy distribution module is also included, used for:

[0302] The target traffic splitting strategy is sent to the terminal device so that the terminal device can select a transmission path to send the target data packet based on the target traffic splitting strategy, or send the target data packet according to the splitting ratio of each transmission path.

[0303] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0304] Corresponding to the above method embodiments, this application also provides an electronic device, including:

[0305] Memory, used to store computer programs;

[0306] A processor is used to implement the above-described data splitting method when executing a computer program.

[0307] like Figure 4 The diagram shows the structural composition of an electronic device, which may include a processor 10, a memory 11, a communication interface 12, and a communication bus 13. The processor 10, memory 11, and communication interface 12 all communicate with each other through the communication bus 13.

[0308] In this embodiment, the processor 10 may be a central processing unit (CPU), an application-specific integrated circuit, a digital signal processor, a field-programmable gate array, or other programmable logic devices.

[0309] The processor 10 can call the program stored in the memory 11. Specifically, the processor 10 can execute the operations in the embodiment of the data splitting method.

[0310] The memory 11 is used to store one or more programs. The programs may include program code, which includes computer operation instructions. In this embodiment, the memory 11 stores at least a program for implementing the following functions:

[0311] obtain the shunt characteristic statistical data of the target data packet to be sent by the terminal device;

[0312] input the shunt characteristic statistical data of the target data packet into a first data shunt decision network to obtain a first shunt confidence of the target data packet matching each transmission path, the first data shunt decision network being obtained by learning network parameters of a pre-constructed convolutional neural network according to a meta-cognition learning algorithm;

[0313] input the shunt characteristic statistical data of the target data packet into a second data shunt decision network to obtain a second shunt confidence of the target data packet matching each transmission path, the second data shunt decision network being obtained by learning network parameters of a pre-constructed multi-layer perception network according to a biological heuristic algorithm;

[0314] determine a target shunt strategy of the target data packet based on the first shunt confidence and the second shunt confidence of the target data packet matching each transmission path.

[0315] In a possible implementation, the memory 11 can include a program storage area and a data storage area, where the program storage area can store an operating system and application programs required by at least one function, etc.; and the data storage area can store data created during use.

[0316] In addition, the memory 11 can include a high-speed random access memory, and can also include a non-volatile memory, for example, at least one magnetic disk storage device or other volatile solid-state memory device.

[0317] The communication interface 12 can be an interface of a communication module, configured to connect with other devices or systems.

[0318] Of course, it needs to be explained that, Figure 4 the structures shown do not constitute a limitation on the electronic device in the embodiments of the present application, and the electronic device can include more or fewer components than those shown, or combine certain components in actual applications. Figure 4 the structures shown do not constitute a limitation on the electronic device in the embodiments of the present application, and the electronic device can include more or fewer components than those shown, or combine certain components in actual applications.

[0319] Corresponding to the above method embodiments, the embodiments of the present application also provide a computer readable storage medium, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the above data shunt method.

[0320] In addition, it should be noted that the embodiments of the present application also provide a computer program product or computer program, which can include computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor can execute the computer instructions to make the computer device execute the description of the data shunting method in the foregoing embodiments, and thus, here will not be described in detail. In addition, the beneficial effects of using the same method will not be described in detail. For technical details not disclosed in the computer program product or computer program embodiments of the present application, please refer to the description of the method embodiments of the present application.

[0321] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0322] It should be noted that in this paper, the term "include", "contain" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or device. Without more limitations, the element defined by the sentence "including a…" does not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiments of the present application is not limited to the order of functions shown or discussed, but can also include functions performed in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method can be performed in an order different from the described order, and various steps can also be added, omitted or combined. In addition, the features described with reference to some examples can be combined in other examples.

[0323] From the above description of the embodiments, those skilled in the art can also clearly understand that the units and algorithm steps of the examples described in combination with the embodiments disclosed in the present paper can be realized by electronic hardware, computer software or combination of both. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been described in the above description. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0324] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in Random-Access Memory (RAM), flash memory, Read-Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), registers, hard disk, a removable disk, a compact disc read-only memory (CD-ROM), or any other form of storage medium known in the art including a storage medium in a server or a computer cloud available via the Internet or extranet, including a series of instructions for causing a processor to execute a series of instructions. An exemplary storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium. In the alternative, hard disk drive can store data which is accessed by a processor through a storage controller coupled to the processor. While a removable memory is shown for purposes of example only, and is not intended to be limiting, as any form of storage medium known in the art can be utilized including a storage medium in a server or a computer cloud available via the Internet or extranet.

[0325] The embodiments of the present application described above are only used to help understand the technical solutions and core ideas of the present application. It should be pointed out that the present application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative, not restrictive. For those skilled in the art, many forms of embodiments can be made without departing from the scope of the present application, and the present application can be improved and modified, which all belong to the protection scope of the present application.

Claims

1. A method of data offloading, the method comprising: The method is applied to a distribution server deployed at a network side, and the method comprises: obtaining distribution characteristic statistical data of a target data packet to be sent by a terminal device; inputting the distribution characteristic statistical data of the target data packet into a first data distribution decision network to obtain a first distribution confidence of the target data packet matching each transmission path, the first data distribution decision network being obtained by learning network parameters of a pre-constructed convolutional neural network according to a meta-cognition learning algorithm; inputting the distribution characteristic statistical data of the target data packet into a second data distribution decision network to obtain a second distribution confidence of the target data packet matching each transmission path, the second data distribution decision network being obtained by learning network parameters of a pre-constructed multi-layer perception network according to a biological heuristic algorithm; determining a target distribution strategy of the target data packet based on the first distribution confidence and the second distribution confidence of the target data packet matching each transmission path.

2. The method of claim 1, wherein, The first data distribution decision network is obtained by the following steps: obtaining a first data set and a second data set, the first data set comprising distribution characteristic statistical data and distribution path labels of a plurality of first samples, and the second data set comprising distribution characteristic statistical data and distribution path labels of a plurality of second samples; inputting the distribution characteristic statistical data of each first sample in the first data set into a pre-constructed convolutional neural network, adjusting weights and parameters of the convolutional neural network according to a deviation between an output result of the convolutional neural network corresponding to each first sample and a distribution path label corresponding to the corresponding first sample, and obtaining a meta-distribution decision network after meta-cognition learning; inputting the distribution characteristic statistical data of each second sample in the second data set into the meta-distribution decision network, adjusting parameters of the meta-distribution decision network according to a deviation between an output result of the meta-distribution decision network corresponding to each second sample and a distribution path label corresponding to the corresponding second sample, and obtaining the first data distribution decision network.

3. The method of claim 2, wherein, Before the meta-distribution decision network is optimized based on the second data set, the method further comprises: determining a noise feature distribution corresponding to each second sample according to the distribution characteristic statistical data of each second sample; for each second sample, in a case where the current second sample is determined to be a noise sample according to the noise feature distribution corresponding to the current second sample, deleting the current second sample in the second data set.

4. The method of claim 1, wherein, The second data distribution decision network is obtained by the following steps: obtaining a third data set and a fourth data set, the third data set comprising distribution characteristic statistical data and distribution path labels of a plurality of third samples, and the fourth data set comprising distribution characteristic statistical data and distribution path labels of a plurality of fourth samples; for each third sample in the third data set, inputting distribution characteristic statistical data of the current third sample into each multi-layer perception network of a pre-constructed basic biological heuristic search cluster respectively to obtain an estimated confidence of the current third sample matching each transmission path output by each multi-layer perception network; inputting the estimated confidence corresponding to each third sample output by each multi-layer perception network and the assigned path label corresponding to the respective third sample into a pre-constructed evaluation network model to obtain an iteration weight distribution of each multi-layer perception network; evolving the basic bio-inspired search cluster according to the iteration weight distribution of each multi-layer perception network; in a case where a preset evolution round number is reached, determining a multi-layer perception network with an iteration weight distribution greater than or equal to a first threshold in the evolved basic bio-inspired search cluster as a candidate machine learning network; learning network parameters of the candidate machine learning network based on the fourth data set, adjusting the parameters of the candidate machine learning network according to a deviation between an output result of the candidate machine learning network corresponding to each fourth sample and the assigned path label corresponding to the respective fourth sample, and obtaining the second data offloading decision network.

5. The method of claim 1, wherein, The method further includes: fusing the first offloading confidence and the second offloading confidence corresponding to each transmission path to obtain an offloading confidence corresponding to each transmission path; determining the target offloading strategy of the target data packet based on the offloading confidence corresponding to each transmission path.

6. The method of claim 5, wherein, The method further includes: for each transmission path, determining, as an offloading confidence corresponding to the current transmission path, an average value of the first offloading confidence and the second offloading confidence corresponding to the current transmission path, or a weighted average value of the first offloading confidence and the second offloading confidence corresponding to the current transmission path, the current transmission path being a transmission path targeted by a current operation.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: sending the target offloading strategy to the terminal device, so that the terminal device selects a transmission path to send the target data packet based on the target offloading strategy, or sends the target data packet according to an offloading proportion of each transmission path.

8. A data offloading apparatus, characterized by, The apparatus is applied to a network-side deployed offloading server, and includes: a feature data acquisition module configured to acquire offloading feature statistical data of a target data packet to be sent by a terminal device; a confidence first obtaining module configured to input the offloading feature statistical data of the target data packet into a first data offloading decision network to obtain a first offloading confidence of the target data packet matching each transmission path, the first data offloading decision network being obtained by learning network parameters of a pre-constructed convolutional neural network according to a meta-cognition learning algorithm; a confidence second obtaining module configured to input the offloading feature statistical data of the target data packet into a second data offloading decision network to obtain a second offloading confidence of the target data packet matching each transmission path, the second data offloading decision network being obtained by learning network parameters of a pre-constructed multi-layer perception network according to a bio-inspired heuristic algorithm; and a confidence third obtaining module configured to input the first offloading confidence and the second offloading confidence of the target data packet matching each transmission path into a third data offloading decision network to determine a target offloading strategy of the target data packet, the third data offloading decision network being obtained by learning network parameters of a pre-constructed evaluation network according to a meta-cognition learning algorithm. The shunt strategy determination module is configured to determine a target shunt strategy of the target data packet based on the first shunt confidence and the second shunt confidence of each transmission path matched by the target data packet.

9. An electronic device, comprising: The method comprises the following steps: A memory is configured to store a computer program. A processor is configured to execute the computer program to implement the steps of the data shunt method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer readable storage medium, and the computer program is executed by the processor to implement the steps of the data shunt method according to any one of claims 1 to 7.

Images