Cloud-edge collaborative systems, data processing methods, electronic devices and storage media

Abstract

This invention provides a cloud-edge collaborative system, a data processing method, an electronic device, and a storage medium. The cloud-edge collaborative system may include an online model update module, a cloud device, and an edge device communicating with the cloud device. The cloud device can be used to obtain the target business scenario and the scene model corresponding to the target business scenario from the edge device. The online model update module can obtain the model performance information of the current scene model during the running of the current scene model. If the model performance information indicates that the current scene model has performance anomalies, it obtains model training data for the current scene model. Based on the model training data, it updates the current scene model to generate a target scene model corresponding to the target business scenario.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to a cloud-edge collaboration system, a cloud-edge collaboration data processing method, an electronic device, and a computer-readable storage medium. Background Technology

[0002] Edge computing refers to an open platform that integrates core capabilities of networking, computing, storage, and applications at the network edge, close to or near the data source. It provides edge intelligence services locally, meeting key industry needs in areas such as agile connectivity, real-time business processes, data optimization, application intelligence, and security and privacy protection. For example, in application scenarios such as energy management, power grid management, smart homes, and industrial production, edge computing can be implemented by configuring corresponding edge devices and cloud devices to process scenario data, thereby improving processing efficiency and effectively reducing the load on the cloud.

[0003] While different AI models can be deployed to suit various edge computing scenarios, the following challenges hinder their large-scale effectiveness in real-world applications:

[0004] 1. Effectiveness Issues: When AI models are deployed to different business scenarios, the characteristics of these different scenarios mean that a model successfully deployed in scenario A cannot be effectively applied directly to scenario B. This requires a large number of algorithm engineers to redesign and train the model, resulting in low production efficiency and hindering large-scale deployment. Furthermore, even if a model has been successfully deployed in the same scenario, the complexity and changing nature of the scenario itself can cause the model's data processing performance to gradually decline, even leading to model failure. These performance degradations or failures often necessitate manual updates and upgrades, further hindering the large-scale deployment of AI models.

[0005] 2. High Cost Issues: To mitigate the gradual failure of models after deployment, a large amount of rich data needs to be collected during model production. This requires significant manpower and time, and in most cases, the collected data is insufficient to meet actual needs, making data collection challenging. Furthermore, edge computing involves data transmission and data storage pressures between the edge and the cloud, easily leading to high data maintenance costs. Summary of the Invention

[0006] This invention provides a cloud-edge collaborative system, a data processing method, an electronic device, and a computer-readable storage medium to solve or partially solve the problems of poor model effectiveness and high data processing costs in edge computing.

[0007] This invention discloses a cloud-edge collaborative system, which includes an online model update module, a cloud device, and an edge device communicatively connected to the cloud device; wherein,

[0008] The cloud device is used to acquire the target business scenario corresponding to the edge device and the scenario model corresponding to the target business scenario.

[0009] The online model update module is used to obtain model performance information of the current scene model during the running of the current scene model; if the model performance information indicates that the current scene model has performance anomalies, then obtain model training data for the current scene model; update the current scene model according to the model training data to generate a target scene model corresponding to the target business scene.

[0010] Optionally, the cloud-edge collaborative system is used for:

[0011] Determine the amount of data transmission between the cloud device and the edge device, and / or the amount of data storage of the cloud device;

[0012] If the data transmission volume is greater than or equal to a first preset quantity threshold, and / or the data storage volume is greater than or equal to a second preset quantity threshold, then the online model update module will be deployed on the edge device.

[0013] If the amount of data transmitted is less than the first preset threshold and the amount of data stored is less than the second preset threshold, then the online model update module will be deployed on the cloud device.

[0014] Optionally, the cloud-edge collaborative system is used for:

[0015] In response to a module processing instruction, obtain the device identifier corresponding to the module processing instruction;

[0016] If the device identifier is the identifier of a cloud device, then the online model update module is deployed on the cloud device or the online model update module preset on the cloud device is activated;

[0017] If the device identifier is an edge device identifier, then the online model update module is deployed on the edge device or the online model update module preset on the edge device is activated.

[0018] Optionally, the model performance information includes scenario prediction information corresponding to the target business scenario and the prediction data distribution characteristics of the scenario prediction information corresponding to the target business scenario; wherein, the online model update module is specifically used for:

[0019] Obtain the scene reference information corresponding to the current scene model and the reference data distribution characteristics corresponding to the current scene model;

[0020] The scene prediction information is compared with the scene reference information to obtain prediction difference information;

[0021] The predicted data distribution characteristics are compared with the reference data distribution characteristics to obtain data distribution difference information;

[0022] If the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then the model training data corresponding to the prediction difference information is obtained.

[0023] Optionally, the online model update module is specifically used for:

[0024] If the preset difference information satisfies the first prediction anomaly condition corresponding to the target business scenario, and the data distribution difference information satisfies the first data distribution anomaly condition corresponding to the target business scenario, then the first model training data for the current scenario model is obtained.

[0025] If the preset difference information satisfies the second prediction anomaly condition corresponding to the target business scenario or the data distribution difference information satisfies the second data distribution anomaly condition corresponding to the target business scenario, then the second model training data for the current scenario model is obtained.

[0026] Optionally, the first model training data includes real-time scene data and historical scene data corresponding to the target business scenario, and the online model update module is further used for:

[0027] Obtain annotation information for the historical scene data;

[0028] The annotation information is used to annotate all or part of the historical scene data to obtain the target historical scene data;

[0029] The current scene model is trained using the real-time scene data and the target historical scene data to generate a target scene model corresponding to the target business scene.

[0030] Optionally, the second model training data includes all or part of the real-time scene data and historical scene data corresponding to the target business scenario, and the online model update module is further used for:

[0031] Obtain annotation information for all or part of the real-time scene data;

[0032] The annotation information is used to annotate all or part of the real-time scene data to generate target business scene data;

[0033] The target business scenario data and the historical scenario data are used to train the current scenario model to generate a target scenario model corresponding to the target business scenario.

[0034] Optionally, the cloud device further includes a seed model library; wherein,

[0035] The cloud device is used to acquire model feature data corresponding to the target scene model, and store the model feature data and the target scene model in the seed model library.

[0036] Optionally, the cloud-edge collaborative system further includes a database, and the cloud device and / or the edge device includes a data acquisition and annotation module; wherein,

[0037] The data acquisition and annotation module is used to acquire scene data corresponding to the target business scenario, annotate the scene data to obtain annotated data corresponding to the scene data, extract features from the annotated data to obtain feature vectors corresponding to the scene data, and store the scene data, the feature vectors and the annotated data in the database.

[0038] Optionally, the cloud-edge collaborative system further includes a database, the edge device includes a first data acquisition and annotation module, and the cloud device includes a second data acquisition and annotation module; wherein,

[0039] The first data acquisition and annotation module is used to acquire scene data corresponding to the target business scenario, encode the scene data, and generate scene coded data;

[0040] The second data acquisition and annotation module is used to decode the scene encoded data to obtain the scene data; to annotate the scene data to obtain annotation data corresponding to the scene data; to extract features from the annotation data to obtain feature vectors corresponding to the scene data; and to store the scene data, the annotation data, and the feature vectors in the database.

[0041] or,

[0042] The first data acquisition and annotation module is used to acquire scene data corresponding to the target business scenario, and to annotate the scene data to obtain annotated data corresponding to the scene data; and to encode the scene data and the annotated data to generate scene coded data.

[0043] The second data acquisition and annotation module is used to decode the scene encoded data to obtain the scene data and the annotation data; to extract features from the annotation data to obtain feature vectors corresponding to the scene data; and to store the scene data, the annotation data and the feature vectors in the database.

[0044] Optionally, the cloud device further includes a model training module; wherein,

[0045] The model training module is used to obtain a seed model corresponding to the feature vector from the seed model library; and to train the seed model according to the scene data and the labeled data to generate a scene model corresponding to the target business scene.

[0046] Optionally, the cloud device includes a first model deployment module, and the edge device includes a second model deployment module; wherein,

[0047] The first model deployment module is used to obtain the target business scenario corresponding to the edge device and the scenario model corresponding to the target business scenario;

[0048] The second model deployment module is used to deploy the scene model and run the scene model after deployment.

[0049] This invention also discloses a cloud-edge collaborative data processing method applied to a cloud-edge collaborative system. The cloud-edge collaborative system includes an online model update module, a cloud device, and an edge device communicatively connected to the cloud device. The method includes:

[0050] The target business scenario corresponding to the edge device and the scene model corresponding to the target business scenario are obtained through the cloud device.

[0051] During the running of the current scene model, the online model update module obtains the model performance information of the current scene model through the edge device.

[0052] If the model performance information indicates that the current scene model has performance anomalies, then the model training data for the current scene model is obtained through the online model update module;

[0053] The online model update module updates the current scene model based on the model training data to generate a target scene model corresponding to the target business scene.

[0054] Optionally, it also includes:

[0055] Determine the amount of data transmission between the cloud device and the edge device, and / or the amount of data storage of the cloud device;

[0056] If the data transmission volume is greater than or equal to a first preset quantity threshold, and / or the data storage volume is greater than or equal to a second preset quantity threshold, then the online model update module will be deployed on the edge device.

[0057] If the amount of data transmitted is less than the first preset threshold and the amount of data stored is less than the second preset threshold, then the online model update module will be deployed on the cloud device.

[0058] Optionally, it also includes:

[0059] In response to a module processing instruction, obtain the device identifier corresponding to the module processing instruction;

[0060] If the device identifier is the identifier of a cloud device, then the online model update module is deployed on the cloud device or the online model update module preset on the cloud device is activated;

[0061] If the device identifier is an edge device identifier, then the online model update module is deployed on the edge device or the online model update module preset on the edge device is activated.

[0062] Optionally, the model performance information includes scenario prediction information corresponding to the target business scenario and the prediction data distribution characteristics of the scenario prediction information corresponding to the target business scenario; wherein, if the model performance information indicates that the current scenario model has performance anomalies, then the model training data for the current scenario model is obtained through the online model update module, including:

[0063] The online model update module obtains the scene reference information corresponding to the current scene model and the reference data distribution characteristics corresponding to the current scene model.

[0064] The online model update module compares the scene prediction information with the scene reference information to obtain prediction difference information.

[0065] The online model update module compares the predicted data distribution characteristics with the reference data distribution characteristics to obtain data distribution difference information.

[0066] If the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then the model training data corresponding to the prediction difference information is obtained through the online model update module.

[0067] Optionally, if the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then the model training data corresponding to the prediction difference information is obtained through the online model update module, including:

[0068] If the preset difference information satisfies the first prediction anomaly condition corresponding to the target business scenario, and the data distribution difference information satisfies the first data distribution anomaly condition corresponding to the target business scenario, then the first model training data for the current scenario model is obtained through the online model update module.

[0069] If the preset difference information satisfies the second prediction anomaly condition corresponding to the target business scenario, or if the data distribution difference information satisfies the second data distribution anomaly condition corresponding to the target business scenario, then the second model training data for the current scenario model is obtained through the online model update module.

[0070] Optionally, the first model training data includes real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model using the online model update module based on the model training data to generate a target scene model corresponding to the target business scenario includes:

[0071] The model online update module obtains annotation information for the historical scene data.

[0072] The online model update module uses the annotation information to annotate all or part of the historical scene data to obtain the target historical scene data.

[0073] The online model update module uses the real-time scene data and the target historical scene data to train the current scene model, generating a target scene model corresponding to the target business scene.

[0074] Optionally, the second model training data includes all or part of the real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model using the online model update module based on the model training data to generate a target scene model corresponding to the target business scenario includes:

[0075] The model online update module obtains annotation information for all or part of the real-time scene data;

[0076] The online model update module uses the annotation information to annotate all or part of the real-time scene data to generate target business scene data.

[0077] The online model update module uses the target business scenario data and the historical scenario data to train the current scenario model, generating a target scenario model corresponding to the target business scenario.

[0078] Optionally, the cloud device further includes a seed model library, and the method further includes:

[0079] The cloud device acquires model feature data corresponding to the target scene model, and stores the model feature data and the target scene model in the seed model library.

[0080] Optionally, the cloud-edge collaborative system further includes a database, and the cloud device and / or the edge device includes a data acquisition and annotation module. The method further includes:

[0081] The data acquisition and annotation module acquires scene data corresponding to the target business scenario, annotates the scene data to obtain annotated data corresponding to the scene data, extracts features from the annotated data to obtain feature vectors corresponding to the scene data, and stores the scene data, the feature vectors, and the annotated data in the database.

[0082] Optionally, the cloud-edge collaborative system further includes a database, the edge device includes a first data acquisition and annotation module, the cloud device includes a second data acquisition and annotation module, and the method further includes:

[0083] The first data acquisition and annotation module acquires scene data corresponding to the target business scenario, encodes the scene data, and generates scene-coded data.

[0084] The scene data is obtained by decoding the scene encoded data through the second data acquisition and annotation module; the scene data is annotated to obtain annotated data corresponding to the scene data; and the annotated data is used to extract features to obtain a feature vector corresponding to the scene data. The scene data, the annotated data, and the feature vector are then stored in the database.

[0085] or,

[0086] The first data acquisition and annotation module acquires scene data corresponding to the target business scenario, and annotates the scene data to obtain annotated data corresponding to the scene data; the scene data and the annotated data are encoded to generate scene coded data.

[0087] The scene data and the annotation data are obtained by decoding the scene encoding data through the second data acquisition and annotation module; feature extraction is performed on the annotation data to obtain the feature vector corresponding to the scene data; and the scene data, the annotation data and the feature vector are stored in the database.

[0088] Optionally, the cloud device further includes a model training module, and the method further includes:

[0089] The model training module obtains a seed model corresponding to the feature vector from the seed model library; and trains the seed model based on the scene data and the labeled data to generate a scene model corresponding to the target business scene.

[0090] Optionally, the cloud device includes a first model deployment module, the edge device includes a second model deployment module, and the method further includes:

[0091] The first model deployment module obtains the target business scenario corresponding to the edge device and the scenario model corresponding to the target business scenario.

[0092] The scene model is deployed through the second model deployment module, and then the scene model is run after deployment.

[0093] This invention also discloses a cloud-edge collaborative data processing method applied to an edge device, wherein the edge device is configured with a scenario model corresponding to a target business scenario, and the method includes:

[0094] Obtain model performance information for the current scene model;

[0095] If the model performance information indicates that the current scene model has performance anomalies, then obtain the model training data for the current scene model;

[0096] The current scene model is updated based on the model training data to generate a target scene model corresponding to the target business scene.

[0097] Optionally, the model performance information includes scenario prediction information corresponding to the target business scenario. If the model performance information indicates that the current scenario model has performance anomalies, then model training data for the current scenario model is obtained, including:

[0098] Obtain the scene reference information corresponding to the current scene model and the reference data distribution characteristics corresponding to the current scene model;

[0099] The scene prediction information is compared with the scene reference information to obtain prediction difference information;

[0100] The predicted data distribution characteristics are compared with the reference data distribution characteristics to obtain data distribution difference information;

[0101] If the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then the model training data corresponding to the prediction difference information is obtained.

[0102] Optionally, if the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then obtaining the model training data corresponding to the prediction difference information includes:

[0103] If the preset difference information satisfies the first prediction anomaly condition corresponding to the target business scenario, and the data distribution difference information satisfies the first data distribution anomaly condition corresponding to the target business scenario, then the first model training data for the current scenario model is obtained.

[0104] If the preset difference information satisfies the second prediction anomaly condition corresponding to the target business scenario or the data distribution difference information satisfies the second data distribution anomaly condition corresponding to the target business scenario, then the second model training data for the current scenario model is obtained.

[0105] Optionally, the first model training data includes real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model based on the model training data to generate a target scene model corresponding to the target business scenario includes:

[0106] Obtain annotation information for the historical scene data;

[0107] The annotation information is used to annotate all or part of the historical scene data to obtain the target historical scene data;

[0108] The current scene model is trained using the real-time scene data and the target historical scene data to generate a target scene model corresponding to the target business scene.

[0109] Optionally, the second model training data includes all or part of the real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model based on the model training data to generate a target scene model corresponding to the target business scenario includes:

[0110] Obtain annotation information for all or part of the real-time scene data;

[0111] The annotation information is used to annotate all or part of the real-time scene data to generate target business scene data;

[0112] The target business scenario data and the historical scenario data are used to train the current scenario model to generate a target scenario model corresponding to the target business scenario.

[0113] This invention also discloses an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0114] The memory is used to store computer programs;

[0115] When the processor executes a program stored in the memory, it implements the method described in the embodiments of the present invention.

[0116] This invention also discloses a computer-readable storage medium storing instructions that, when executed by one or more processors, cause the processors to perform the methods described in this invention.

[0117] The embodiments of the present invention have the following advantages:

[0118] In this embodiment of the invention, the cloud-edge collaborative system may include an online model update module, a cloud device, and an edge device communicating with the cloud device. The cloud device can be used to obtain the target business scenario and the scene model corresponding to the target business scenario for the edge device. The online model update module can obtain the model performance information of the current scene model during the running of the current scene model. If the model performance information indicates that the current scene model has performance anomalies, it obtains the model training data for the current scene model. Based on the model training data, the current scene model is updated to generate a target scene model corresponding to the target business scenario. Thus, in the edge computing scenario, by performing performance detection and updating of the model, not only is the model maintenance cost reduced, but the effectiveness of the model can also be guaranteed and manual model upgrades can be avoided, which greatly promotes the large-scale production and deployment of the model. In addition, by obtaining the corresponding training data during the model update process, the data requirements for model updates can be effectively reduced. By collecting a small amount of sample data from the actual scenario, the difficulty of data collection is greatly reduced. Attached Figure Description

[0119] Figure 1 This is a structural block diagram of a cloud-edge collaborative system provided in an embodiment of the present invention;

[0120] Figure 2 This is a structural block diagram of the cloud-edge collaborative system provided in this embodiment of the invention;

[0121] Figure 3 This is a schematic diagram of the edge computing process provided in an embodiment of the present invention;

[0122] Figure 4 This is a schematic diagram of the data acquisition and annotation process provided in an embodiment of the present invention;

[0123] Figure 5 This is a schematic diagram of the model training process provided in an embodiment of the present invention;

[0124] Figure 6 This is a schematic diagram of the model update process provided in an embodiment of the present invention;

[0125] Figure 7 This is a flowchart illustrating a cloud-edge collaborative data processing method provided in an embodiment of the present invention;

[0126] Figure 8 This is a flowchart illustrating a cloud-edge collaborative data processing method provided in an embodiment of the present invention;

[0127] Figure 9 This is a block diagram of an electronic device provided in an embodiment of the present invention;

[0128] Figure 10 This is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. Detailed Implementation

[0129] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0130] The Industrial Internet is a product of the deep integration of next-generation information and communication technologies with modern industrial technologies. It serves as a crucial carrier for the digitalization, networking, and intelligentization of manufacturing, and represents a commanding height in the new round of industrial competition. By constructing a fundamental network connecting machines, materials, people, and information systems, the Industrial Internet enables comprehensive perception, dynamic transmission, and real-time analysis of industrial data, leading to scientific decision-making and intelligent control, and improving the efficiency of manufacturing resource allocation. The greatest value of the Industrial Internet lies in its comprehensive connection of data and physical entities, and based on this connection, it fosters computation and control, making real-time decision-making possible.

[0131] As an example, the power industry, one of the largest sub-markets in the Industrial Internet of Things (IIoT) sector, suffers from massive daily data collection volumes coupled with significant deficiencies in the overall construction of the Internet of Things (IoT). This results in data inconsistencies and weak real-time sharing during power grid operation, failing to fully reflect the power grid's safe operation, efficiency, and work quality. Furthermore, infrastructure construction suffers from insufficient flexibility in data storage, processing, and application, and inadequate ability to respond quickly to changing demands. To address this, the ubiquitous power IoT, based on existing power grid infrastructure, can play a greater role on the distribution side, particularly in control and data collection, improving power grid efficiency and power supply reliability. The key technologies involved in the ubiquitous power IoT are "cloud-pipe-edge-device." "Cloud" refers to big data, artificial intelligence, and data mining platforms such as the State Grid Cloud, integrating dozens of business applications including project management and marketing systems. "Pipe" refers to the communication network, providing reliable communication channels for power grid information transmission. "Edge" refers to deploying edge computing devices to build a distributed data center for the power grid, enhancing overall edge computing capabilities. "Device" refers to creating intelligent terminals and sensor products to achieve real-time data collection and embedding security management products such as identity authentication and behavioral analysis.

[0132] In the entire system, edge computing occupies a crucial, bridging position. First, edge computing can push the computing power of the cloud platform closer to the terminal, integrating innovative technologies such as network transmission and storage to improve the data processing efficiency of each node and reduce latency. Second, edge computing can provide more precise and localized data processing, only needing to feed the results back to the backend, thus protecting data security and privacy from the source. Third, edge computing devices significantly reduce data transmission volume, saving bandwidth and enabling real-time responses to front-end demands, shortening response time. Fourth, edge computing maintains uninterrupted service even in weak network or out-of-network environments, improving operational stability. Fifth, edge computing devices and edge devices can form a unified whole; AI (Artificial Intelligence) components can be deployed on edge computing devices to meet customized and intelligent management needs. Therefore, edge computing can provide customers with one-stop management. In the overall construction and layout of the energy internet, edge computing is a vital component, playing an irreplaceable role in breaking down barriers between IoT terminals and promoting data integration.

[0133] For edge computing, different artificial intelligence models are deployed to suit different edge computing scenarios. However, the following difficulties exist in the implementation of these models, making it difficult for them to play a significant role on a large scale in real-world scenarios:

[0134] 1. Effectiveness Issues: When AI models are deployed to different business scenarios, the characteristics of these different scenarios mean that a model successfully deployed in scenario A cannot be effectively applied directly to scenario B. This requires a large number of algorithm engineers to redesign and train the model, resulting in low production efficiency and hindering large-scale deployment. Furthermore, even if a model has been successfully deployed in the same scenario, the complexity and changing nature of the scenario itself can cause the model's data processing performance to gradually decline, even leading to model failure. These performance degradations or failures often necessitate manual updates and upgrades, further hindering the large-scale deployment of AI models.

[0135] 2. High Cost Issues: To mitigate the gradual failure of models after deployment, a large amount of rich data needs to be collected during model production. This requires significant manpower and time, and in most cases, the collected data is insufficient to meet actual needs, making data collection challenging. Furthermore, edge computing involves data transmission and data storage pressures between the edge and the cloud, easily leading to high data maintenance costs.

[0136] One of the core inventive points of this invention is the deployment of a corresponding online model update module in the cloud-edge collaborative system. This module detects and updates the model's performance during its operation. In edge computing scenarios, performance detection and updates not only reduce model maintenance costs but also ensure model effectiveness and avoid manual model upgrades, greatly promoting large-scale model production and deployment. Furthermore, by acquiring corresponding training data during the model update process, the data requirements for model updates can be effectively reduced. By collecting a small amount of sample data from real-world scenarios, the difficulty of data collection is significantly reduced.

[0137] Reference Figure 1 This diagram illustrates a structural block diagram of a cloud-edge collaborative system provided in an embodiment of the present invention. The cloud-edge collaborative system includes an online model update module, at least one cloud device, and at least one edge device communicatively connected to each of the cloud devices.

[0138] The cloud device is used to acquire the target business scenario corresponding to the edge device and the scenario model corresponding to the target business scenario.

[0139] The online model update module is used to obtain model performance information of the current scene model during the running of the current scene model; if the model performance information indicates that the current scene model has performance anomalies, then obtain model training data for the current scene model; update the current scene model according to the model training data to generate a target scene model corresponding to the target business scene.

[0140] Specifically, the online model update module can be flexibly configured according to actual needs. In one scenario, equipment manufacturers can calculate data transmission costs between cloud and edge devices, as well as storage costs on cloud devices, during the equipment production phase, and deploy the online model update module on either cloud or edge devices based on the calculation results. In another scenario, a detection module can be added to the cloud-edge collaboration system to detect data transmission costs between cloud and edge devices, as well as storage costs on cloud devices, and deploy the online model update module on either cloud or edge devices based on the detection results. In yet another scenario, users can communicate with the cloud-edge collaboration system via applications or hardware, sending module processing commands to the system to deploy the online model update module on cloud or edge devices, or to activate a pre-installed online model update module on either cloud or edge devices. By calculating actual data transmission and storage costs, the deployment or running location of the online model update module can be flexibly adjusted, effectively meeting different user needs and greatly improving the versatility of data processing in the cloud-edge collaboration system.

[0141] Specifically, the adaptive deployment or activation of the online model update module in the cloud-edge system can be achieved as follows: The cloud-edge collaborative system can first determine the data transmission volume between the cloud device and the edge device, and / or the data storage volume of the cloud device. If the data transmission volume is greater than or equal to a first preset threshold, and / or the data storage volume is greater than or equal to a second preset threshold, then the online model update module is deployed on the edge device; if the data transmission volume is less than the first preset threshold and the data storage volume is less than the second preset threshold, then the online model update module is deployed on the cloud device.

[0142] In a specific implementation, a model deployment and startup module can be configured in the cloud device of the cloud-edge collaborative system. When the model deployment and startup module detects that the data transmission volume and data storage volume meet the conditions, if the cloud device already has a model online update module, it can start that module. If the cloud device does not have a corresponding model online update module, it can obtain a model online update module from the edge device or other cloud devices, and then deploy that model online update module. For example, for the former, the model online update module can be pre-deployed in the cloud device in hardware or software. When both the data transmission volume and data storage volume are less than the corresponding threshold, the cloud device can directly start the model online update module to detect and update the model's running performance to ensure the model's effectiveness. For the latter, the model online update module can be in software form. When both the data transmission volume and data storage volume are less than the corresponding threshold, since the cloud device does not have its own model online update module, it can obtain it through appropriate means and deploy it to detect and update the model's running performance to ensure the model's effectiveness. This invention does not impose any limitations on this.

[0143] For personalized deployment of the online model update module by users, the cloud-edge collaborative system can respond to the module processing command and obtain the device identifier corresponding to the command. If the device identifier is that of a cloud device, the online model update module will be deployed on the cloud device or a pre-installed online model update module on the cloud device will be activated. If the device identifier is that of an edge device, the online model update module will be deployed on the edge device or a pre-installed online model update module on the edge device will be activated. It should be noted that the deployment of the online model update module on the edge device is described above and will not be repeated here.

[0144] In one example, the deployment of the online model update module on an edge device is illustrated by reference. Figure 2This diagram illustrates the structural block diagram of the cloud-edge collaborative system provided in this embodiment of the invention. The cloud device may include a data acquisition and annotation module, a model training module, a model deployment module, and a seed model library, etc. The edge device may include a data acquisition and annotation module, a model online update module, and a model deployment module, etc. By deploying the model online update module on the edge device, the scene model can be deployed on the edge device, which can then run, detect, and update the scene model. This reduces data transmission between the edge device and the cloud device, lowering data transmission costs and ensuring data privacy and security. Furthermore, by performing performance testing and updating the model on the edge device, model maintenance costs are reduced. Collecting data corresponding to the business scenario required for model updates on the edge device lowers the data requirements for model updates, allowing model updates to be completed based on a small amount of sample data, greatly promoting the large-scale production and deployment of models.

[0145] Reference Figure 3 This diagram illustrates the edge computing process provided in this embodiment of the invention. For a cloud-edge collaborative system, upon receiving a data stream from a client, the data acquisition and annotation modules located on edge devices and / or cloud devices can be used to evaluate and clean the transmitted scene data. The cleaned data is then annotated using semi-automatic data annotation. After annotation, data features are extracted, and the original image, extracted data features, and annotation information are stored in a database. Next, the annotated data and data features can be input into the model training module in the cloud. The model training module compares the current data features with the data features in the seed model library, extracts the seed model most similar to the scene, and trains and fine-tunes it based on this model. The trained model is then stored in the seed model library and simultaneously input into the model deployment module for model compression, quantization, and deployment on the corresponding edge devices. During model operation, the online model update module located on the edge device can evaluate the performance of the deployed model. When a performance degradation or data distribution change is detected, online learning is initiated on the edge device. The model undergoes supervised or unsupervised incremental training according to the degree of performance degradation, completing the model's self-evolution and upgrade. By building a cloud-edge collaborative system, the scene model is deployed on the edge device, reducing data transmission between the edge device and the cloud device. This not only reduces data transmission costs but also ensures data privacy and security. At the same time, by performing performance testing and updates on the model on the edge device, model maintenance costs are reduced.

[0146] In this embodiment of the invention, the data acquisition and labeling module can be deployed on cloud devices, edge devices, or both cloud devices and edge devices.

[0147] In one optional embodiment, the cloud-edge collaborative system further includes a database. The cloud device and / or edge device include a data acquisition and annotation module. The data acquisition and annotation module can be used to acquire scene data corresponding to the target business scenario, annotate the scene data to obtain annotated data corresponding to the scene data, and extract features from the annotated data to obtain feature vectors corresponding to the scene data. The scene data, feature vectors, and annotated data are then stored in the database. By deploying the system on one side of the cloud device or edge device, the structure of the cloud-edge collaborative system can be effectively simplified and the system's operating efficiency can be improved. Furthermore, by deploying the system on both sides, the computing power of both the cloud device and the edge device can be fully utilized, improving the accuracy of data annotation and feature extraction.

[0148] In another optional embodiment, assuming the edge device includes a first data acquisition and annotation module and the cloud device includes a second data acquisition and annotation module, then in one case, the first data acquisition and annotation module is used to acquire scene data corresponding to the target business scenario, encode the scene data, and generate scene encoded data; the second data acquisition and annotation module is used to decode the scene encoded data to obtain scene data; annotate the scene data to obtain annotation data corresponding to the scene data, and extract features from the annotation data to obtain feature vectors corresponding to the scene data, and store the scene data, annotation data, and feature vectors in the database.

[0149] In another scenario, the first data acquisition and annotation module can be used to acquire scene data corresponding to the target business scenario, and annotate the scene data to obtain annotated data corresponding to the scene data; encode the scene data and annotated data to generate scene encoded data; the second data acquisition and annotation module can be used to decode the scene encoded data to obtain scene data and annotated data; extract features from the annotated data to obtain feature vectors corresponding to the scene data, and store the scene data, annotated data, and feature vectors in the database.

[0150] It's important to note that for the former, edge devices can act as data acquisition devices. After collecting relevant scene data, they compress the data and send it to cloud devices. The cloud devices then annotate and extract features from the scene data to obtain the target data. Utilizing edge devices for data acquisition effectively reduces the workload of cloud devices. Simultaneously, leveraging the powerful data processing capabilities of cloud devices for data annotation and feature extraction ensures data processing accuracy, providing a better foundation for subsequent model training. For the latter, edge devices can function as data acquisition devices and perform preliminary data quality assessment and cleaning, including data annotation and feature extraction. The cleaned data is then sent to cloud devices for secondary data quality assessment and cleaning. These multiple assessments and cleanings effectively improve data accuracy and validity, providing a better foundation for subsequent model training. Therefore, in edge computing, users can flexibly choose the deployment method of the data acquisition and annotation modules according to their actual needs, effectively meeting the requirements of different edge scenarios and improving the versatility of cloud-edge collaborative system data processing.

[0151] In one example, when the data transmission and storage pressure between cloud and edge devices is relatively low, the data acquisition and annotation module can be deployed on the cloud device. This module performs data quality evaluation and cleaning (quality evaluation includes, but is not limited to, image blur detection, anomaly detection, and missing value statistics; cleaning includes, but is not limited to, redundant data removal, missing value imputation, and outlier removal or correction), semi-automatic annotation, and data feature extraction on the collected scene data. The raw data, data features, and annotation information are then stored in a database. Data feature extraction can be divided into explicit feature extraction and implicit feature extraction. Explicit features include identifying the detected target category as a cat, while implicit features include using a feature extractor to extract image features and form feature vectors. This single-sided deployment fully leverages the performance of each device.

[0152] When the data transmission and storage pressure between cloud devices and edge devices is high, data acquisition and annotation modules can be deployed simultaneously on both cloud devices and the edge. The data acquisition and annotation module on the edge device can be used to perform data quality evaluation and cleaning, data compression and encoding, and transmission to the cloud device in sequence on the collected scene data. The data acquisition and annotation module on the cloud device is used to decode the data. The decoded data is then labeled using semi-automatic data annotation. Data feature extraction is performed on the labeled data. Finally, the original data, data features, and annotation information are stored in the database. This multi-sided deployment effectively reduces the data processing pressure on each side. By leveraging the scene advantages of the edge devices and cloud devices (edge ​​devices can collect scene data corresponding to the business scenario in real time, and cloud devices have strong data computing capabilities, etc.) for data processing, the data processing efficiency of cloud-edge collaborative computing is effectively improved.

[0153] Optionally, the scene data can be data corresponding to a specific business scenario. For example, in a light intensity detection scenario, the scene data could be hourly brightness data during the day; in a substation switch opening and closing recognition scenario, the scene data could be a switch video stream, etc. Feature vectors can be used to characterize the data features of the scene data, such as explicit and implicit features. Annotated data can be data obtained by annotating the scene data. Different data storage types of scene data can correspond to different labeled data, such as image-type data corresponding to image labeled data, video-type data corresponding to video labeled data, etc. This invention does not impose any limitations on this.

[0154] In one example, refer to Figure 4 This diagram illustrates a data acquisition and annotation process provided in an embodiment of the present invention. Assuming corresponding data acquisition and annotation modules are deployed on both the edge device and the cloud device, the first data acquisition and annotation module deployed on the edge device, upon receiving the customer data stream, can perform quality assessment and cleaning of the customer data stream. After completing the quality assessment and cleaning, the data is compressed and transmitted to the cloud device. Upon receiving the compressed data from the edge device, the cloud device can decode the compressed data using the second data acquisition and annotation module. Then, it performs semi-automatic annotation and data feature extraction on the decoded data and stores the processed data in a corresponding database, such as a seed model library. This completes the quality assessment and cleaning of the scene data, obtaining the target data for subsequent model training.

[0155] In this embodiment of the invention, the cloud device also includes a model training module. After the data quality evaluation and cleaning are completed, the cloud device can obtain the seed model corresponding to the feature vector from the seed model library through the model training module; and train the seed model according to the scene data and labeled data to generate a scene model corresponding to the target business scene.

[0156] In its implementation, the model training module can first perform scene feature matching between the feature vectors and the data features of past business scenarios stored in the seed model library. This can be achieved through category similarity matching (e.g., if the new scenario is image recognition, specifically Persian cat recognition, and the seed model library contains images of cats, then the similarity is high), feature matching (e.g., similarity based on the cosine distance of feature vectors), or a combination of both. By using scene feature matching, the most matching seed model is selected from the seed model library as the initial model. Then, based on this initial model, the initial model is trained using the corresponding scene data and labeled data to obtain a scene model that meets the business scenario. Furthermore, by storing previously successfully deployed scene models in the seed model library, model training can be performed based on existing seed models when training scene models for new business scenarios, effectively improving the efficiency of model training.

[0157] In addition, cloud devices and edge devices may also include model deployment modules for deploying scene models. Specifically, cloud devices may include a first model deployment module and edge devices may include a second model deployment module. The first model deployment module is used to obtain the target business scenario corresponding to the edge device and the scene model corresponding to the target business scenario. The second model deployment module is used to deploy the scene model and run the scene model after deployment.

[0158] The business scenarios can be those where edge devices perform data processing in edge computing. For example, in the field of energy intelligence, business scenarios could include substation switch opening and closing identification, photovoltaic power generation prediction, and residential park electricity consumption monitoring. Different business scenarios may correspond to different edge devices, and these devices can be equipped with scenario models adapted to the specific business scenario. These scenario models can then be used to perform data processing corresponding to the business scenario, thereby realizing edge computing.

[0159] In one example, refer to Figure 5This diagram illustrates the model training process provided in this embodiment of the invention. The cloud device can use the model training module to match the acquired labeled data and scene features with scene features stored in the seed model library, extracting a seed model corresponding to the business scenario requiring model training. Then, the seed model is trained and fine-tuned based on the labeled data and scene features to obtain the trained model. Simultaneously, the trained model can be stored in the seed model library, enriching the types of stored seed models. This allows for model training based on existing seed models when training scene models for new business scenarios, effectively improving the efficiency of model training.

[0160] Once the scene model is deployed on the edge device, the online model update module deployed on the edge device can detect the running performance of the scene model during its operation. This allows for training updates when the model performance degrades, ensuring the effectiveness of the scene model.

[0161] In its implementation, the online model update module deployed on edge devices can include functions such as model performance evaluation, data feature extraction, and model update. Specifically, model performance evaluation assesses the performance of the currently running scene model; data feature extraction extracts features from scene data (e.g., data features can be explicit or implicit features; explicit features include detecting the target category as a cat, while implicit features are extracted from image features using a feature extractor to form feature vectors); and model update selects the appropriate model update training method based on the degree of degradation in model performance.

[0162] In this embodiment of the invention, the online model update module can be used to detect model performance and data distribution during the online operation of a trained model or the operation of a model in the current scenario, and record the corresponding data features through the model performance evaluation function in the module. When a potential decrease in model performance or a change in data distribution characteristics is detected, the degree of model performance decrease can be judged. If the model performance decrease is small, unsupervised model update training is started. If the model performance decrease is large, a semi-automatic annotation method can be used to annotate the real-time data stream images collected on-site, and then supervised model update training is started. This reduces model maintenance costs by performing model performance detection and updates on edge devices. By collecting data corresponding to the business scenario required for model updates on edge devices, the data requirements for model updates are reduced, and model updates can be completed based on a small amount of sample data, greatly promoting the large-scale production and deployment of models.

[0163] Optionally, taking the deployment of the scene model on an edge device as an example, the online model update module can be deployed on the edge device, and the edge device can handle the model operation, detection and update. It is understood that the scene model and the online model update module can also be deployed on a cloud device, and its data processing process is the same as that of the edge device. This invention does not limit this.

[0164] Specifically, the current scene model running on the edge device can be the first scene model deployed on the edge device, which has not been updated at this time; it can be a model trained by the cloud device by collecting scene data corresponding to the business scenario; or it can be a model deployed on the edge device that has been updated at least once during operation. This invention does not limit this. Optionally, the model performance information can be information characterizing whether the model's running performance meets the data processing requirements in the business scenario, such as whether the model's prediction performance meets the prediction requirements, whether the model's detection performance meets the detection requirements, and so on.

[0165] In this embodiment of the invention, the edge device can detect the performance of the currently running scene model through model performance information to determine whether to update the current scene model. The model performance information may include scene prediction information corresponding to the target business scene and the distribution characteristics of the prediction data corresponding to the scene prediction information. When the performance of the scene model can be directly detected, it can be evaluated using the scene prediction information; when the performance of the scene model cannot be directly detected, it can be evaluated using the distribution characteristics of the prediction data corresponding to the scene prediction information. The former evaluates the performance of the scene model at the model level, while the latter evaluates it at the data level. Therefore, for different scene models, an appropriate evaluation method can be selected to evaluate the model's performance and determine whether to update the scene model currently running on the edge device.

[0166] In the specific implementation, if the preset difference information meets the first prediction anomaly condition corresponding to the target business scenario, and the data distribution difference information meets the first data distribution anomaly condition corresponding to the target business scenario, then the first model training data for the current scenario model is obtained; if the preset difference information meets the second prediction anomaly condition corresponding to the target business scenario or the data distribution difference information meets the second data distribution anomaly condition corresponding to the target business scenario, then the second model training data for the current scenario model is obtained.

[0167] For scenario models whose performance can be evaluated at the model level, edge devices can first obtain scenario reference information corresponding to the current scenario model through the online model update module. Then, the scenario prediction information is compared with the scenario reference information to obtain prediction difference information. If the prediction difference information indicates that the current scenario model has performance anomalies, then the model training data corresponding to the prediction difference information is obtained. The prediction difference information can be used to characterize the degree of anomaly in the scenario model's running performance. Different model update methods can be selected for different degrees of anomaly. Specifically, if the prediction difference information meets the first prediction anomaly condition corresponding to the target business scenario, then the first model training data corresponding to the first prediction anomaly condition is obtained; if the prediction difference information meets the second prediction anomaly condition corresponding to the target business scenario, then the second model training data corresponding to the second prediction anomaly condition is obtained.

[0168] The scene prediction information includes the prediction results of the current scene model based on scene data in real time, and the scene reference information, which serves as a reference for evaluating the current scene model. By comparing the scene prediction information with the scene reference information, the prediction results of the current scene model can be compared with the reference results, and the model performance can be evaluated using the prediction difference information obtained after the comparison. Furthermore, the model update conditions determine the appropriate update method for the scene model. The first prediction anomaly condition allows for selecting the first update method when the scene model performance degradation is minor, while the second prediction anomaly condition allows for selecting the second update method when the scene model performance degradation is significant. Different update methods correspond to different model training data. Optionally, regarding model update conditions, since different business scenarios correspond to different scenario models, the model update conditions can be adjusted according to the actual business scenario. For example, in the power scenario, when the business scenario uses detection accuracy as the standard, the model update condition can be an accuracy threshold (such as a first accuracy threshold, a second accuracy threshold, etc.); when the detection quantity is used as the standard, the model update condition can be a quantity threshold (such as a first quantity threshold, a second quantity threshold, etc.). This invention does not impose any restrictions on this.

[0169] In one example, scene prediction information can be the scene prediction value of the current scene model, and scene reference information can be the scene reference value of the current scene model. The scene prediction value can then be compared with the scene reference value to obtain the difference between the two. For example, for the current scene model, a first prediction anomaly condition can be set as a first threshold, and a second update condition can be set as a second threshold, where the second threshold is greater than the first threshold, and both the first and second thresholds are positive values. When the difference between the scene prediction value and the scene reference value is greater than the first threshold and less than the second threshold, the online model update module can obtain the first model training data corresponding to the first update condition; when the difference between the scene prediction value and the scene reference value is greater than the second threshold, the online model update module can obtain the second model training data corresponding to the second update condition, so as to train the model based on the corresponding model training data.

[0170] For scenario models whose performance can be evaluated at the data level, edge devices can first obtain the reference data distribution characteristics corresponding to the current scenario model through the online model update module. Then, they compare the scenario data distribution characteristics with the reference data distribution characteristics to obtain data distribution difference information. If the data distribution difference information indicates that the current scenario model has performance anomalies, then the model training data corresponding to the data distribution difference information is obtained. Specifically, the data distribution difference information can be used to characterize abnormal programs in the scenario model's running performance. If the data distribution difference information meets the first data distribution anomaly condition corresponding to the target business scenario, then the first model training data corresponding to the first data distribution anomaly condition is obtained; if the data distribution difference information meets the second data distribution anomaly condition corresponding to the target business scenario, then the second model training data corresponding to the second data distribution anomaly condition is obtained.

[0171] The predicted data distribution characteristics can be the data distribution characteristics corresponding to the model prediction results output during the current scene model's operation; the reference data distribution characteristics can be the data distribution characteristics corresponding to the model prediction results output under the current scene model's standard operating state. When the model operation is normal, the data distribution characteristics corresponding to its prediction results can be the reference data distribution characteristics. By comparing the predicted data distribution characteristics with the reference data distribution characteristics, the data characteristics corresponding to the current scene model's prediction results and the data characteristics corresponding to the reference results can be compared. The model performance of the current scene model can then be evaluated based on the differences in the compared data characteristics. Optionally, the model update conditions can be referred to the aforementioned description and will not be repeated here.

[0172] In one example, the predicted data distribution characteristics can be the data distribution corresponding to the prediction results of the current scene model during operation, while the reference data distribution characteristics can be the data distribution corresponding to the ideal prediction results of the current scene model during operation. The two can then be compared to determine if there are any anomalies in the predicted data distribution characteristics. For example, in a brightness detection scenario, the scene model can detect brightness values ​​and output a corresponding brightness curve. When the scene model's performance is normal, the brightness curve can be a smooth curve; when the scene model's performance degrades or becomes abnormal, the brightness curve can be a curve with multiple peaks. By comparing the two curves, the curve offset is determined, and then compared with a preset first offset threshold and a second offset threshold, where the second offset threshold is greater than the first offset threshold. When the difference between the scene predicted value and the scene reference value is greater than the first offset threshold but less than the second offset threshold, the online model update module can obtain the first model training data corresponding to the first update condition; when the difference between the scene predicted value and the scene reference value is greater than the second offset threshold, the online model update module can obtain the second model training data corresponding to the second update condition, so as to train the model based on the corresponding model training data.

[0173] It should be noted that the embodiments of the present invention use the first and second prediction anomaly conditions as examples for illustrative purposes. It can be understood that, under the guidance of the embodiments of the present invention, the update method of the gradient descent strategy can also be set for the model performance. In this way, by judging the degree of degradation of the model performance, an appropriate model update method can be selected to update the model according to the degree of degradation, thereby improving the accuracy and effectiveness of model prediction.

[0174] In addition, for model anomaly detection, other methods such as deep learning autoencoder-based anomaly detection, autoencoder anomaly detection scheme combined with background modeling, data classification-based methods, and data pattern drift detection methods can also be used, and this invention does not limit these methods.

[0175] In this embodiment of the invention, due to differences in model update conditions, the performance degradation of the scene model varies. Edge devices can use different update methods through the online model update module to update the model. Different update methods result in different model training data.

[0176] For scenario models whose performance can be evaluated at the model level, edge devices can update the model using the first model training data through the online model update module. The first model training data includes real-time scenario data and historical scenario data corresponding to the target business scenario. The online model update module can first obtain the annotation information for the historical scenario data, then use the annotation information to annotate all or part of the historical scenario data to obtain the target historical scenario data. Finally, it uses the real-time scenario data and the target historical scenario data to train the current scenario model, generating a target scenario model corresponding to the target business scenario.

[0177] For scenario models whose performance can be evaluated at the data level, edge devices can update the model using the second model training data through the online model update module. The second model training data includes all or part of the real-time scenario data and historical scenario data corresponding to the target business scenario. The online model update module can first obtain the annotation information for all or part of the real-time scenario data, then use the annotation information to annotate all or part of the real-time scenario data to generate the target business scenario data. Finally, it uses the target business scenario data and historical scenario data to train the current scenario model, generating a target scenario model corresponding to the target business scenario.

[0178] Specifically, when model performance degradation is minor, the current scene model can be updated using unsupervised model updates to reduce update time while ensuring effective model updates. Real-time scene data can be scene data generated by the business scenario. The online model update module can input real-time scene data into the current scene model for data processing to obtain corresponding prediction results. Historical scene data can include prior scene data and annotation information for that data, such as scene data ① - annotation information A, scene data ② - annotation information B, and scene data ③ - annotation information C. Scene data can serve as input data for the scene model, while annotation information can serve as output data. By updating the current scene model using both historical and real-time scene data, training can be performed while retaining the memory of historical data distribution, conforming to the current data distribution. This not only ensures the accuracy and effectiveness of the model but also effectively reduces the collection of training samples and lowers data collection costs by using a small amount of real-time scene data. When model performance degrades significantly, and unsupervised model update strategies are insufficient to restore it to a usable level, supervised model updates can be used to update the current scenario model. Specifically, semi-automatic annotation can be used to label real-time scene data first. Then, the labeled real-time scene data and historical scene data are combined to perform supervised update training on the current scenario model to ensure its effectiveness. By using an online model update module to perform performance testing and updates, model maintenance costs are reduced. By collecting data corresponding to the business scenario required for model updates from edge devices, the data requirements for model updates are reduced, and model updates can be completed based on a small amount of sample data, greatly promoting the large-scale production and deployment of models.

[0179] In one example, when the performance of the prediction difference information characterization model slightly decreases or the prediction data distribution characteristics characterize a small shift in the current scene model, the online model update module can directly use the unlabeled real-time data stream and labeled historical data to update the model. Specifically, the labeled historical data is used for normal training, while the unlabeled real-time data can be used for model training in various ways, including but not limited to: 1. A pseudo-label-based model training and update scheme, which uses the current scene model to predict pseudo-labels for the unlabeled real-time data, and then combines this with the labeled historical data to update the model; 2. A consistency-based model training scheme, for example, performing simple and complex data augmentation transformations on the same unlabeled image to obtain two different images. Since the target semantic information in the images does not change during the data augmentation transformation, the semantic information of the two images does not change. Therefore, the model is trained by increasing the constraint on the consistency of the semantic information of the two images.

[0180] When the performance of the prediction difference information characterization model drops significantly or the prediction data distribution characteristics characterize the current scene model with a large deviation, the real-time scene data can be labeled first through the online model update module. For example, real-time scene data can be labeled using image labeling, voice labeling, text labeling, and video labeling. Then, the labeled real-time scene data is combined with the labeled historical scene data to update and train the current scene model. Thus, the model can be updated and trained in different ways according to the degree of performance degradation of the scene model. While ensuring the effectiveness of model operation performance through model update training, the flexibility of model training is improved by using different update training methods. Furthermore, by collecting the data corresponding to the business scenario required for model update on edge devices, the data requirements for model update are reduced. Model update can be completed based on a small amount of sample data, which greatly promotes the large-scale production and deployment of the model.

[0181] In one example, refer to Figure 6 This diagram illustrates the model update process provided in this embodiment of the invention. For scene models successfully deployed on edge devices, the online model update module can evaluate the performance of the current scene model. When a performance degradation is detected, real-time data streams and historical labeled cached data can be acquired. Then, based on the degree of performance degradation, unsupervised or supervised model updates are selected. After the update is completed, the updated target scene model is sent to a seed model library for storage to enrich the variety of seed models in the library. Furthermore, while acquiring the real-time data stream, data features can be extracted and stored in the seed model library. During supervised model training and updates, training data can be labeled using semi-automatic annotation methods, and the labeled data is stored in the database. This data storage facilitates direct retrieval during model training, improving training efficiency.

[0182] It should be noted that the embodiments of the present invention include, but are not limited to, the examples described above. It is understood that those skilled in the art can make further settings according to actual needs under the guidance of the ideas in the embodiments of the present invention. For example, they can use an adaptive incremental learning approach, or add other structures and techniques that are beneficial to model training based on the examples described above. The present invention does not limit these options.

[0183] Furthermore, after completing the model update training, the old scene model can be replaced. During the operation of the new scene model, model performance continues to be monitored. This allows for continued update training even if performance degrades, ensuring the scene model's effectiveness. Simultaneously, after completing the scene model update training, the updated scene model can be sent to a cloud device for storage in its seed model library. Simultaneously, the training data corresponding to the updated scene model is also stored. This allows for rapid training of scene models adapted to different business scenarios using the seed models and training data stored in the seed model library when needed for model deployment or training.

[0184] For the seed model library, after the scene model is updated and trained on the edge device through the online model update module, the updated target scene model and the corresponding data during the update training process can be sent to the cloud device, where the seed model library stores the model feature data and the target scene model.

[0185] It should be noted that the embodiments of the present invention include, but are not limited to, the examples described above. It is understood that those skilled in the art can make further settings according to actual needs under the guidance of the ideas in the embodiments of the present invention, and the present invention does not limit such settings.

[0186] In this embodiment of the invention, the cloud-edge collaborative system may include an online model update module, a cloud device, and an edge device communicating with the cloud device. The cloud device can be used to obtain the target business scenario and the scene model corresponding to the target business scenario for the edge device. The online model update module can obtain the model performance information of the current scene model during the running of the current scene model. If the model performance information indicates that the current scene model has performance anomalies, it obtains the model training data for the current scene model. Based on the model training data, the current scene model is updated to generate a target scene model corresponding to the target business scenario. Thus, in the edge computing scenario, by performing performance detection and updating of the model, not only is the model maintenance cost reduced, but the effectiveness of the model can also be guaranteed and manual model upgrades can be avoided, which greatly promotes the large-scale production and deployment of the model. In addition, by obtaining the corresponding training data during the model update process, the data requirements for model updates can be effectively reduced. By collecting a small amount of sample data from the actual scenario, the difficulty of data collection is greatly reduced.

[0187] Reference Figure 7 This diagram illustrates a flowchart of a cloud-edge collaborative data processing method provided in an embodiment of the present invention. The method is applied to a cloud-edge collaborative system, which includes an online model update module, a cloud device, and an edge device communicatively connected to the cloud device. Specifically, the method may include the following steps:

[0188] Step 701: Obtain the target business scenario corresponding to the edge device and the scene model corresponding to the target business scenario through the cloud device;

[0189] Step 702: During the running of the current scene model, the online model update module obtains the model performance information of the current scene model through the edge device.

[0190] Step 703: If the model performance information indicates that the current scene model has performance anomalies, then obtain the model training data for the current scene model through the online model update module;

[0191] Step 704: The online model update module updates the current scene model based on the model training data to generate a target scene model corresponding to the target business scene.

[0192] In one alternative embodiment, it further includes:

[0193] Determine the amount of data transmission between the cloud device and the edge device, and / or the amount of data storage of the cloud device;

[0194] If the data transmission volume is greater than or equal to a first preset quantity threshold, and / or the data storage volume is greater than or equal to a second preset quantity threshold, then the online model update module will be deployed on the edge device.

[0195] If the amount of data transmitted is less than the first preset threshold and the amount of data stored is less than the second preset threshold, then the online model update module will be deployed on the cloud device.

[0196] In one alternative embodiment, it further includes:

[0197] In response to a module processing instruction, obtain the device identifier corresponding to the module processing instruction;

[0198] If the device identifier is the identifier of a cloud device, then the online model update module is deployed on the cloud device or the online model update module preset on the cloud device is activated;

[0199] If the device identifier is an edge device identifier, then the online model update module is deployed on the edge device or the online model update module preset on the edge device is activated.

[0200] In one optional embodiment, the model performance information includes scenario prediction information corresponding to the target business scenario and the prediction data distribution characteristics of the scenario prediction information corresponding to the target business scenario; wherein, if the model performance information indicates that the current scenario model has performance anomalies, then the model training data for the current scenario model is obtained through the online model update module, including:

[0201] The online model update module obtains the scene reference information corresponding to the current scene model and the reference data distribution characteristics corresponding to the current scene model.

[0202] The online model update module compares the scene prediction information with the scene reference information to obtain prediction difference information.

[0203] The online model update module compares the predicted data distribution characteristics with the reference data distribution characteristics to obtain data distribution difference information.

[0204] If the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then the model training data corresponding to the prediction difference information is obtained through the online model update module.

[0205] In one optional embodiment, if the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then obtaining the model training data corresponding to the prediction difference information through the online model update module includes:

[0206] If the preset difference information satisfies the first prediction anomaly condition corresponding to the target business scenario, and the data distribution difference information satisfies the first data distribution anomaly condition corresponding to the target business scenario, then the first model training data for the current scenario model is obtained through the online model update module.

[0207] If the preset difference information satisfies the second prediction anomaly condition corresponding to the target business scenario, or if the data distribution difference information satisfies the second data distribution anomaly condition corresponding to the target business scenario, then the second model training data for the current scenario model is obtained through the online model update module.

[0208] In one optional embodiment, the first model training data includes real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model using the online model update module based on the model training data to generate a target scene model corresponding to the target business scenario includes:

[0209] The model online update module obtains annotation information for the historical scene data.

[0210] The online model update module uses the annotation information to annotate all or part of the historical scene data to obtain the target historical scene data.

[0211] The online model update module uses the real-time scene data and the target historical scene data to train the current scene model, generating a target scene model corresponding to the target business scene.

[0212] In one optional embodiment, the second model training data includes all or part of the real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model using the online model update module based on the model training data to generate a target scene model corresponding to the target business scenario includes:

[0213] The model online update module obtains annotation information for all or part of the real-time scene data;

[0214] The online model update module uses the annotation information to annotate all or part of the real-time scene data to generate target business scene data.

[0215] The online model update module uses the target business scenario data and the historical scenario data to train the current scenario model, generating a target scenario model corresponding to the target business scenario.

[0216] In one alternative embodiment, the cloud device further includes a seed model library, and the method further includes:

[0217] The cloud device acquires model feature data corresponding to the target scene model, and stores the model feature data and the target scene model in the seed model library.

[0218] In one optional embodiment, the cloud-edge collaborative system further includes a database, the cloud device and / or the edge device include a data acquisition and annotation module, and the method further includes:

[0219] The data acquisition and annotation module acquires scene data corresponding to the target business scenario, annotates the scene data to obtain annotated data corresponding to the scene data, extracts features from the annotated data to obtain feature vectors corresponding to the scene data, and stores the scene data, the feature vectors, and the annotated data in the database.

[0220] In one optional embodiment, the cloud-edge collaborative system further includes a database, the edge device includes a first data acquisition and annotation module, the cloud device includes a second data acquisition and annotation module, and the method further includes:

[0221] The first data acquisition and annotation module acquires scene data corresponding to the target business scenario, encodes the scene data, and generates scene-coded data.

[0222] The scene data is obtained by decoding the scene encoded data through the second data acquisition and annotation module; the scene data is annotated to obtain annotated data corresponding to the scene data; and the annotated data is used to extract features to obtain a feature vector corresponding to the scene data. The scene data, the annotated data, and the feature vector are then stored in the database.

[0223] or,

[0224] The first data acquisition and annotation module acquires scene data corresponding to the target business scenario, and annotates the scene data to obtain annotated data corresponding to the scene data; the scene data and the annotated data are encoded to generate scene coded data.

[0225] The scene data and the annotation data are obtained by decoding the scene encoding data through the second data acquisition and annotation module; feature extraction is performed on the annotation data to obtain the feature vector corresponding to the scene data; and the scene data, the annotation data and the feature vector are stored in the database.

[0226] In one alternative embodiment, the cloud device further includes a model training module, and the method further includes:

[0227] The model training module obtains a seed model corresponding to the feature vector from the seed model library; and trains the seed model based on the scene data and the labeled data to generate a scene model corresponding to the target business scene.

[0228] In one optional embodiment, the cloud device includes a first model deployment module, the edge device includes a second model deployment module, and the method further includes:

[0229] The first model deployment module obtains the target business scenario corresponding to the edge device and the scenario model corresponding to the target business scenario.

[0230] The scene model is deployed through the second model deployment module, and then the scene model is run after deployment.

[0231] In this embodiment of the invention, the cloud-edge collaborative system may include an online model update module, a cloud device, and an edge device communicating with the cloud device. The cloud device can be used to obtain the target business scenario and the scene model corresponding to the target business scenario for the edge device. The online model update module can obtain the model performance information of the current scene model during the running of the current scene model. If the model performance information indicates that the current scene model has performance anomalies, it obtains the model training data for the current scene model. Based on the model training data, the current scene model is updated to generate a target scene model corresponding to the target business scenario. Thus, in the edge computing scenario, by performing performance detection and updating of the model, not only is the model maintenance cost reduced, but the effectiveness of the model can also be guaranteed and manual model upgrades can be avoided, which greatly promotes the large-scale production and deployment of the model. In addition, by obtaining the corresponding training data during the model update process, the data requirements for model updates can be effectively reduced. By collecting a small amount of sample data from the actual scenario, the difficulty of data collection is greatly reduced.

[0232] Reference Figure 8 This diagram illustrates a cloud-edge collaborative data processing method provided in an embodiment of the present invention. Applied to an edge device, the edge device is configured with a scenario model corresponding to the target business scenario, and may specifically include the following steps:

[0233] Step 801: Obtain the model performance information of the current scene model;

[0234] Step 802: If the model performance information indicates that the current scene model has a performance anomaly, then obtain the model training data for the current scene model.

[0235] Step 803: Update the current scene model based on the model training data to generate a target scene model corresponding to the target business scene.

[0236] In one optional embodiment, the model performance information includes scenario prediction information corresponding to the target business scenario. If the model performance information indicates that the current scenario model has performance anomalies, then model training data for the current scenario model is obtained, including:

[0237] Obtain the scene reference information corresponding to the current scene model and the reference data distribution characteristics corresponding to the current scene model;

[0238] The scene prediction information is compared with the scene reference information to obtain prediction difference information;

[0239] The predicted data distribution characteristics are compared with the reference data distribution characteristics to obtain data distribution difference information;

[0240] If the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then the model training data corresponding to the prediction difference information is obtained.

[0241] In one optional embodiment, if the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then obtaining the model training data corresponding to the prediction difference information includes:

[0242] If the preset difference information satisfies the first prediction anomaly condition corresponding to the target business scenario, and the data distribution difference information satisfies the first data distribution anomaly condition corresponding to the target business scenario, then the first model training data for the current scenario model is obtained.

[0243] If the preset difference information satisfies the second prediction anomaly condition corresponding to the target business scenario or the data distribution difference information satisfies the second data distribution anomaly condition corresponding to the target business scenario, then the second model training data for the current scenario model is obtained.

[0244] In one optional embodiment, the first model training data includes real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model based on the model training data to generate a target scene model corresponding to the target business scenario includes:

[0245] Obtain annotation information for the historical scene data;

[0246] The annotation information is used to annotate all or part of the historical scene data to obtain the target historical scene data;

[0247] The current scene model is trained using the real-time scene data and the target historical scene data to generate a target scene model corresponding to the target business scene.

[0248] In one optional embodiment, the second model training data includes all or part of the real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model based on the model training data to generate a target scene model corresponding to the target business scenario includes:

[0249] Obtain annotation information for all or part of the real-time scene data;

[0250] The annotation information is used to annotate all or part of the real-time scene data to generate target business scene data;

[0251] The target business scenario data and the historical scenario data are used to train the current scenario model to generate a target scenario model corresponding to the target business scenario.

[0252] In this embodiment of the invention, the cloud-edge collaborative system may include an online model update module, a cloud device, and an edge device communicating with the cloud device. The cloud device can be used to obtain the target business scenario and the scene model corresponding to the target business scenario for the edge device. The online model update module can obtain the model performance information of the current scene model during the running of the current scene model. If the model performance information indicates that the current scene model has performance anomalies, it obtains the model training data for the current scene model. Based on the model training data, the current scene model is updated to generate a target scene model corresponding to the target business scenario. Thus, in the edge computing scenario, by performing performance detection and updating of the model, not only is the model maintenance cost reduced, but the effectiveness of the model can also be guaranteed and manual model upgrades can be avoided, which greatly promotes the large-scale production and deployment of the model. In addition, by obtaining the corresponding training data during the model update process, the data requirements for model updates can be effectively reduced. By collecting a small amount of sample data from the actual scenario, the difficulty of data collection is greatly reduced.

[0253] To enable those skilled in the art to better understand the technical solutions of the embodiments of the present invention, a specific example is provided below for explanation.

[0254] In the field of energy intelligence, relevant business scenarios include the identification of substation disconnector opening and closing, photovoltaic power generation prediction, and residential park electricity consumption monitoring. Edge devices can be edge boxes that do not include parallel computing hardware such as GPUs (Graphics Processing Units) and NPUs (Neural-network Processing Units), or they can be edge boxes that include parallel computing hardware such as GPUs and NPUs. Specifically, taking substation disconnector opening and closing identification as an example:

[0255] Due to the large number of monitoring cameras in substations, resulting in significant transmission and storage pressure, data acquisition and annotation modules are deployed at both the edge and cloud levels. Upon receiving the disconnector video stream from the substation, the edge data acquisition and annotation module first performs data quality evaluation and cleaning, deleting low-quality disconnector images caused by transmission errors or highly repetitive content. The cleaned images are then compressed and encoded before being uploaded to the cloud. The cloud decodes the images to reconstruct the original disconnector images, which are then annotated using semi-automatic data annotation. Feature extraction is performed on the annotated images, including but not limited to the target category name and target characteristics (e.g., feature vectors are extracted from images of the disconnector in open and closed states using a neural network). The original images, feature information, and annotation information are then stored in a database.

[0256] The labeled disconnect switch data and its features are input into the model training module. The module first compares the disconnect switch data features with those in the seed model library. Comparison methods include, but are not limited to, category name similarity matching (e.g., if a previously trained and deployed fuse-type disconnect switch model exists in the seed library, its category name is highly similar to the current category) and feature vector similarity matching (e.g., comparing two feature vectors using cosine similarity, Euclidean distance similarity, Manhattan distance similarity, etc.; higher similarity indicates a better match). Next, the seed model with the most similar scenario is selected from the seed model library, and fine-tuned based on this model. The trained model can then be stored in the seed model library and simultaneously input into the model deployment module for model compression, quantization, and deployment at the edge.

[0257] Because data transmission and storage costs are high for identifying the status of disconnect switches in substations, the online model update module is deployed at the edge. Once the trained model is online, the model performance evaluation function in the module will detect model performance or data distribution.

[0258] Model performance testing includes methods based on labeled data, methods based on unlabeled data, and combinations of both. One example is that annotators periodically (every 10 days / month, etc.) quickly annotate a switch video over a past period (5 minutes / 10 minutes / half an hour, etc.) and use it as a test set to test whether model performance degrades. Optionally, other methods that can test model performance without annotation can also be used, and this invention does not limit this approach.

[0259] Data distribution detection primarily examines whether the characteristics of the current disconnector video have changed significantly compared to previous videos. When a decrease in model performance or a change in data distribution is detected, the degree of this performance degradation is assessed.

[0260] If the model performance degradation is minor (e.g., accuracy less than 95% but greater than 90%), unsupervised model update training is initiated. Unsupervised model update comprises two parts: the first part uses labeled historical data for normal training; the second part uses unlabeled real-time data, which can be trained in various ways, including but not limited to: 1. A pseudo-label-based model training and update scheme, where the current model predicts pseudo-labels for unlabeled real-time data, and then updates the model based on labeled historical data; 2. A consistency-based regularization-based model training scheme, where the same unlabeled image undergoes both simple and complex data augmentation transformations to obtain two different images. Since the target semantic information in the images does not change during the data augmentation transformation, the semantic information of these two images remains unchanged. Therefore, the model is trained by increasing the constraint on the consistency of the semantic information between the two images. It should be noted that adaptive incremental learning can also be used, or other structures and techniques beneficial to model training can be added to the above examples; this invention does not limit these approaches.

[0261] If the model performance degrades significantly (e.g., accuracy falls below 90%), a semi-automatic annotation method is used to annotate the real-time data stream images collected on-site, and then supervised model training is initiated. It should be noted that any further improvements to this approach (such as adding other structures and techniques beneficial to model training) are covered by the patent.

[0262] Ultimately, the updated and trained model replaces the existing old model. Simultaneously, the new model and corresponding data features are uploaded to the seed model library. In edge computing scenarios, this reduces data transmission between edge and cloud devices by deploying the scenario model on edge devices, lowering data transmission costs and ensuring data privacy and security. Furthermore, performance testing and updates on edge devices reduce model maintenance costs. Collecting data relevant to the business scenario for model updates on edge devices lowers the data requirements for model updates, enabling updates based on a small amount of sample data, significantly promoting large-scale model production and deployment.

[0263] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention. As for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple; relevant details can be found in the descriptions of the method embodiments.

[0264] In addition, embodiments of the present invention also provide an electronic device, such as... Figure 9 As shown, it includes a processor 901, a communication interface 902, a memory 903, and a communication bus 904, wherein the processor 901, the communication interface 902, and the memory 903 communicate with each other through the communication bus 904.

[0265] Memory 903 is used to store computer programs;

[0266] When processor 901 executes a program stored in memory 903, it performs the following steps:

[0267] Obtain the target business scenario corresponding to the edge device and the scenario model corresponding to the target business scenario;

[0268] During the running of the current scene model, the model performance information of the current scene model is obtained through the edge device;

[0269] If the model performance information indicates that the current scene model has performance anomalies, then obtain the model training data for the current scene model;

[0270] The current scene model is updated based on the model training data to generate a target scene model corresponding to the target business scene.

[0271] In one alternative embodiment, it further includes:

[0272] Determine the amount of data transmission between the cloud device and the edge device, and / or the amount of data storage of the cloud device;

[0273] If the data transmission volume is greater than or equal to a first preset quantity threshold, and / or the data storage volume is greater than or equal to a second preset quantity threshold, then the online model update module will be deployed on the edge device.

[0274] If the amount of data transmitted is less than the first preset threshold and the amount of data stored is less than the second preset threshold, then the online model update module will be deployed on the cloud device.

[0275] In one alternative embodiment, it further includes:

[0276] In response to a module processing instruction, obtain the device identifier corresponding to the module processing instruction;

[0277] If the device identifier is the identifier of a cloud device, then the online model update module is deployed on the cloud device or the online model update module preset on the cloud device is activated;

[0278] If the device identifier is an edge device identifier, then the online model update module is deployed on the edge device or the online model update module preset on the edge device is activated.

[0279] In one optional embodiment, the model performance information includes scenario prediction information corresponding to the target business scenario and the prediction data distribution characteristics of the scenario prediction information corresponding to the target business scenario; wherein, if the model performance information indicates that the current scenario model has performance anomalies, then obtaining model training data for the current scenario model includes:

[0280] Obtain the scene reference information corresponding to the current scene model and the reference data distribution characteristics corresponding to the current scene model;

[0281] The scene prediction information is compared with the scene reference information to obtain prediction difference information;

[0282] The predicted data distribution characteristics are compared with the reference data distribution characteristics to obtain data distribution difference information;

[0283] If the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then the model training data corresponding to the prediction difference information is obtained.

[0284] In one optional embodiment, if the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then obtaining the model training data corresponding to the prediction difference information includes:

[0285] If the preset difference information satisfies the first prediction anomaly condition corresponding to the target business scenario, and the data distribution difference information satisfies the first data distribution anomaly condition corresponding to the target business scenario, then the first model training data for the current scenario model is obtained.

[0286] If the preset difference information satisfies the second prediction anomaly condition corresponding to the target business scenario or the data distribution difference information satisfies the second data distribution anomaly condition corresponding to the target business scenario, then the second model training data for the current scenario model is obtained.

[0287] In one optional embodiment, the first model training data includes real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model based on the model training data to generate a target scene model corresponding to the target business scenario includes:

[0288] Obtain annotation information for the historical scene data;

[0289] The annotation information is used to annotate all or part of the historical scene data to obtain the target historical scene data;

[0290] The current scene model is trained using the real-time scene data and the target historical scene data to generate a target scene model corresponding to the target business scene.

[0291] In one optional embodiment, the second model training data includes all or part of the real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model based on the model training data to generate a target scene model corresponding to the target business scenario includes:

[0292] Obtain annotation information for all or part of the real-time scene data;

[0293] The annotation information is used to annotate all or part of the real-time scene data to generate target business scene data;

[0294] The target business scenario data and the historical scenario data are used to train the current scenario model to generate a target scenario model corresponding to the target business scenario.

[0295] In an optional embodiment, the method further includes:

[0296] Obtain model feature data corresponding to the target scene model, and store the model feature data and the target scene model in the seed model library.

[0297] In an optional embodiment, the method further includes:

[0298] Obtain scenario data corresponding to the target business scenario;

[0299] The scene data is labeled to obtain labeled data corresponding to the scene data;

[0300] Feature extraction is performed on the labeled data to obtain a feature vector corresponding to the scene data;

[0301] The scene data, the feature vector, and the labeled data are stored in the database.

[0302] In an optional embodiment, the method further includes:

[0303] Obtain scenario data corresponding to the target business scenario, encode the scenario data, and generate scenario-encoded data;

[0304] The scene encoded data is decoded to obtain the scene data;

[0305] The scene data is labeled to obtain labeled data corresponding to the scene data;

[0306] Feature extraction is performed on the labeled data to obtain a feature vector corresponding to the scene data;

[0307] The scene data, the labeled data, and the feature vectors are stored in the database;

[0308] In an optional embodiment, the method further includes:

[0309] Obtain scenario data corresponding to the target business scenario;

[0310] The scene data is labeled to obtain labeled data corresponding to the scene data;

[0311] The scene data and the labeled data are encoded to generate scene-coded data;

[0312] Decode the scene encoded data to obtain the scene data and the annotation data;

[0313] Feature extraction is performed on the labeled data to obtain a feature vector corresponding to the scene data;

[0314] The scene data, the labeled data, and the feature vectors are stored in the database.

[0315] In an optional embodiment, the method further includes:

[0316] Obtain the seed model corresponding to the feature vector from the seed model library;

[0317] The seed model is trained based on the scene data and the labeled data to generate a scene model corresponding to the target business scenario.

[0318] In an optional embodiment, the method further includes:

[0319] Obtain the target business scenario corresponding to the edge device and the scenario model corresponding to the target business scenario;

[0320] The scene model is deployed, and then run after deployment.

[0321] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0322] The communication interface is used for communication between the aforementioned terminal and other devices.

[0323] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage. Alternatively, the memory may be at least one storage device located away from the processor.

[0324] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0325] like Figure 10 As shown, in another embodiment of the present invention, a computer-readable storage medium 1001 is also provided, which stores instructions that, when executed on a computer, cause the computer to perform the cloud-edge collaborative data processing method described in the above embodiment.

[0326] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the cloud-edge collaborative data processing method described in the above embodiments.

[0327] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0328] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0329] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0330] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A cloud-edge collaborative system, characterized in that, The cloud-edge collaborative system includes an online model update module, cloud devices, and edge devices that communicate with the cloud devices; wherein... The cloud device is used to acquire the target business scenario corresponding to the edge device and the scenario model corresponding to the target business scenario. The online model update module is used to obtain model performance information of the current scene model during the running of the current scene model; if the model performance information indicates that the current scene model has performance anomalies, then obtain model training data for the current scene model; update the current scene model according to the model training data to generate a target scene model corresponding to the target business scene; The cloud-edge collaborative system is used for: Determine the amount of data transmission between the cloud device and the edge device, and / or the amount of data storage of the cloud device; If the data transmission volume is greater than or equal to a first preset quantity threshold, and / or the data storage volume is greater than or equal to a second preset quantity threshold, then the online model update module will be deployed on the edge device. If the data transmission volume is less than the first preset threshold and the data storage volume is less than the second preset threshold, then the online model update module will be deployed on the cloud device. The online model update module is also used to judge the model performance information. When the model performance drops slightly, unsupervised model update training is started; when the model performance drops significantly, a semi-automatic annotation method is used to annotate the real-time data stream images collected on-site, and supervised model update training is started.

2. The cloud-edge collaborative system according to claim 1, characterized in that, The cloud-edge collaborative system is used for: In response to a module processing instruction, obtain the device identifier corresponding to the module processing instruction; If the device identifier is the identifier of a cloud device, then the online model update module is deployed on the cloud device or the online model update module preset on the cloud device is activated; If the device identifier is an edge device identifier, then the online model update module is deployed on the edge device or the online model update module preset on the edge device is activated.

3. The cloud-edge collaborative system according to claim 1, characterized in that, The model performance information includes scenario prediction information corresponding to the target business scenario and the prediction data distribution characteristics of the scenario prediction information corresponding to the target business scenario; wherein, the online model update module is specifically used for: Obtain the scene reference information corresponding to the current scene model and the reference data distribution characteristics corresponding to the current scene model; The scene prediction information is compared with the scene reference information to obtain prediction difference information; The predicted data distribution characteristics are compared with the reference data distribution characteristics to obtain data distribution difference information; If the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then the model training data corresponding to the prediction difference information is obtained.

4. The cloud-edge collaborative system according to claim 3, characterized in that, The online model update module is specifically used for: If the prediction difference information satisfies the first prediction anomaly condition corresponding to the target business scenario, and the data distribution difference information satisfies the first data distribution anomaly condition corresponding to the target business scenario, then the first model training data for the current scenario model is obtained. If the prediction difference information satisfies the second prediction anomaly condition corresponding to the target business scenario or the data distribution difference information satisfies the second data distribution anomaly condition corresponding to the target business scenario, then the second model training data for the current scenario model is obtained. The first prediction anomaly condition is when the model performance decreases only slightly; The second prediction anomaly condition is when the model performance deteriorates significantly.

5. The cloud-edge collaborative system according to claim 4, characterized in that, The first model training data includes real-time scene data and historical scene data corresponding to the target business scenario. The online model update module is further used for: Obtain annotation information for the historical scene data; The annotation information is used to annotate all or part of the historical scene data to obtain the target historical scene data; The current scene model is trained using the real-time scene data and the target historical scene data to generate a target scene model corresponding to the target business scene.

6. The cloud-edge collaborative system according to claim 4, characterized in that, The second model training data includes all or part of the real-time scene data and historical scene data corresponding to the target business scenario. The online model update module is further used for: Obtain annotation information for all or part of the real-time scene data; The annotation information is used to annotate all or part of the real-time scene data to generate target business scene data; The target business scenario data and the historical scenario data are used to train the current scenario model to generate a target scenario model corresponding to the target business scenario.

7. The cloud-edge collaborative system according to claim 1, characterized in that, The cloud device also includes a seed model library; wherein... The cloud device is used to acquire model feature data corresponding to the target scene model, and store the model feature data and the target scene model in the seed model library.

8. The cloud-edge collaborative system according to claim 1, characterized in that, The cloud-edge collaborative system also includes a database, and the cloud device and / or the edge device include a data acquisition and annotation module; wherein... The data acquisition and annotation module is used to acquire scene data corresponding to the target business scenario, annotate the scene data to obtain annotated data corresponding to the scene data, extract features from the annotated data to obtain feature vectors corresponding to the scene data, and store the scene data, the feature vectors and the annotated data in the database.

9. The cloud-edge collaborative system according to claim 1, characterized in that, The cloud-edge collaborative system also includes a database; the edge device includes a first data acquisition and annotation module; and the cloud device includes a second data acquisition and annotation module. The first data acquisition and annotation module is used to acquire scene data corresponding to the target business scenario, encode the scene data, and generate scene coded data; The second data acquisition and annotation module is used to decode the scene encoded data to obtain the scene data; to annotate the scene data to obtain annotation data corresponding to the scene data; to extract features from the annotation data to obtain feature vectors corresponding to the scene data; and to store the scene data, the annotation data, and the feature vectors in the database. or, The first data acquisition and annotation module is used to acquire scene data corresponding to the target business scenario, and to annotate the scene data to obtain annotated data corresponding to the scene data; and to encode the scene data and the annotated data to generate scene coded data. The second data acquisition and annotation module is used to decode the scene encoded data to obtain the scene data and the annotation data; to extract features from the annotation data to obtain feature vectors corresponding to the scene data; and to store the scene data, the annotation data and the feature vectors in the database.

10. The cloud-edge collaborative system according to claim 8 or 9, characterized in that, The cloud device also includes a model training module; wherein... The model training module is used to obtain a seed model corresponding to the feature vector from the seed model library; and to train the seed model according to the scene data and the labeled data to generate a scene model corresponding to the target business scene.

11. The cloud-edge collaborative system according to any one of claims 1-9, characterized in that, The cloud device includes a first model deployment module, and the edge device includes a second model deployment module; wherein... The first model deployment module is used to obtain the target business scenario corresponding to the edge device and the scenario model corresponding to the target business scenario; The second model deployment module is used to deploy the scene model and run the scene model after deployment.

12. A cloud-edge collaborative data processing method, characterized in that, The method, applied to a cloud-edge collaborative system, includes an online model update module, a cloud device, and an edge device communicatively connected to the cloud device. The target business scenario corresponding to the edge device and the scene model corresponding to the target business scenario are obtained through the cloud device. During the running of the current scene model, the online model update module obtains the model performance information of the current scene model through the edge device. If the model performance information indicates that the current scene model has performance anomalies, then the model training data for the current scene model is obtained through the online model update module; The online model update module updates the current scene model based on the model training data to generate a target scene model corresponding to the target business scene. Determine the amount of data transmission between the cloud device and the edge device, and / or the amount of data storage of the cloud device; If the data transmission volume is greater than or equal to a first preset quantity threshold, and / or the data storage volume is greater than or equal to a second preset quantity threshold, then the online model update module will be deployed on the edge device. If the data transmission volume is less than the first preset threshold and the data storage volume is less than the second preset threshold, then the online model update module will be deployed on the cloud device. The online model update module is also used to judge the model performance information. When the model performance drops slightly, unsupervised model update training is started; when the model performance drops significantly, a semi-automatic annotation method is used to annotate the real-time data stream images collected on-site, and supervised model update training is started.

13. The method according to claim 12, characterized in that, Also includes: In response to a module processing instruction, obtain the device identifier corresponding to the module processing instruction; If the device identifier is the identifier of a cloud device, then the online model update module is deployed on the cloud device or the online model update module preset on the cloud device is activated; If the device identifier is an edge device identifier, then the online model update module is deployed on the edge device or the online model update module preset on the edge device is activated.

14. The method according to claim 12, characterized in that, The model performance information includes scenario prediction information corresponding to the target business scenario and the prediction data distribution characteristics of the scenario prediction information corresponding to the target business scenario; wherein, if the model performance information indicates that the current scenario model has performance anomalies, then the model training data for the current scenario model is obtained through the online model update module, including: The online model update module obtains the scene reference information corresponding to the current scene model and the reference data distribution characteristics corresponding to the current scene model. The online model update module compares the scene prediction information with the scene reference information to obtain prediction difference information. The online model update module compares the predicted data distribution characteristics with the reference data distribution characteristics to obtain data distribution difference information. If the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then the model training data corresponding to the prediction difference information is obtained through the online model update module.

15. The method according to claim 14, characterized in that, If the prediction difference information and / or the data distribution difference information indicate that the current scene model has performance anomalies, then the model training data corresponding to the prediction difference information is obtained through the online model update module, including: If the prediction difference information satisfies the first prediction anomaly condition corresponding to the target business scenario, and the data distribution difference information satisfies the first data distribution anomaly condition corresponding to the target business scenario, then the first model training data for the current scenario model is obtained through the online model update module. If the prediction difference information satisfies the second prediction anomaly condition corresponding to the target business scenario or the data distribution difference information satisfies the second data distribution anomaly condition corresponding to the target business scenario, then the second model training data for the current scenario model is obtained through the online model update module. The first prediction anomaly condition is when the model performance decreases only slightly; The second prediction anomaly condition is when the model performance deteriorates significantly.

16. The method according to claim 15, characterized in that, The first model training data includes real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model using the online model update module based on the model training data to generate a target scene model corresponding to the target business scenario includes: The model online update module obtains annotation information for the historical scene data. The online model update module uses the annotation information to annotate all or part of the historical scene data to obtain the target historical scene data. The online model update module uses the real-time scene data and the target historical scene data to train the current scene model, generating a target scene model corresponding to the target business scene.

17. The method according to claim 15, characterized in that, The second model training data includes all or part of the real-time scene data and historical scene data corresponding to the target business scenario. The step of updating the current scene model using the online model update module based on the model training data to generate a target scene model corresponding to the target business scenario includes: The model online update module obtains annotation information for all or part of the real-time scene data; The online model update module uses the annotation information to annotate all or part of the real-time scene data to generate target business scene data. The online model update module uses the target business scenario data and the historical scenario data to train the current scenario model, generating a target scenario model corresponding to the target business scenario.

18. The method according to claim 12, characterized in that, The cloud device also includes a seed model library, and the method further includes: The cloud device acquires model feature data corresponding to the target scene model, and stores the model feature data and the target scene model in the seed model library.

19. The method according to claim 12, characterized in that, The cloud-edge collaborative system further includes a database, and the cloud device and / or the edge device include a data acquisition and annotation module. The method further includes: The data acquisition and annotation module acquires scene data corresponding to the target business scenario, annotates the scene data to obtain annotated data corresponding to the scene data, extracts features from the annotated data to obtain feature vectors corresponding to the scene data, and stores the scene data, the feature vectors, and the annotated data in the database.

20. The method according to claim 12, characterized in that, The cloud-edge collaborative system further includes a database, the edge device includes a first data acquisition and annotation module, the cloud device includes a second data acquisition and annotation module, and the method further includes: The first data acquisition and annotation module acquires scene data corresponding to the target business scenario, encodes the scene data, and generates scene-coded data. The scene data is obtained by decoding the scene encoded data through the second data acquisition and annotation module; the scene data is annotated to obtain annotated data corresponding to the scene data; and the annotated data is used to extract features to obtain a feature vector corresponding to the scene data. The scene data, the annotated data, and the feature vector are then stored in the database. or, The first data acquisition and annotation module acquires scene data corresponding to the target business scenario, and annotates the scene data to obtain annotated data corresponding to the scene data; the scene data and the annotated data are encoded to generate scene coded data. The scene data and the annotation data are obtained by decoding the scene encoding data through the second data acquisition and annotation module; feature extraction is performed on the annotation data to obtain the feature vector corresponding to the scene data; and the scene data, the annotation data and the feature vector are stored in the database.

21. The method according to claim 19 or 20, characterized in that, The cloud device also includes a model training module, and the method further includes: The model training module obtains a seed model corresponding to the feature vector from the seed model library; and trains the seed model based on the scene data and the labeled data to generate a scene model corresponding to the target business scene.

22. The method according to any one of claims 12-20, characterized in that, The cloud device includes a first model deployment module, the edge device includes a second model deployment module, and the method further includes: The first model deployment module obtains the target business scenario corresponding to the edge device and the scenario model corresponding to the target business scenario. The scene model is deployed through the second model deployment module, and then the scene model is run after deployment.

23. A cloud-edge collaborative data processing method, characterized in that, Applied to edge devices, the edge devices are configured with a scenario model corresponding to a target business scenario. An online model update module, a cloud device, and the edge device communicatively connected to the cloud device constitute a cloud-edge collaborative system. The method includes: Obtain model performance information for the current scene model; If the model performance information indicates that the current scene model has performance anomalies, then obtain the model training data for the current scene model; The current scene model is updated based on the model training data to generate a target scene model corresponding to the target business scene; the cloud-edge collaboration system is used to: determine the data transmission volume between the cloud device and the edge device, and / or the data storage volume of the cloud device; if the data transmission volume is greater than or equal to a first preset threshold, and / or the data storage volume is greater than or equal to a second preset threshold, then the online model update module is deployed on the edge device; if the data transmission volume is less than the first preset threshold and the data storage volume is less than the second preset threshold, then the online model update module is deployed on the cloud device; The online model update module is also used to judge the model performance information. When the model performance drops slightly, unsupervised model update training is started; when the model performance drops significantly, a semi-automatic annotation method is used to annotate the real-time data stream images collected on-site, and supervised model update training is started.

24. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes a program stored in the memory, it implements the method as described in any one of claims 12-23.

25. A computer-readable storage medium having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the method as described in any one of claims 12-23.

Images