Intelligent edge device and operation method and system thereof

Abstract

The invention provides an intelligent edge device and an operation method and system thereof, the intelligent edge device is deployed with a main model with a reasoning function, and the operation method comprises the following steps: calling the main model to perform reasoning based on an obtained feature data set to obtain a reasoning result data set; automatically transmitting and storing reasoning data when the main model is called for reasoning, wherein the reasoning data at least comprises a feature data set and a reasoning result data set; an incremental data set is created based on the collected real result data set and the feature data set for updating the master model. According to the method, the reasoning data and the real result data set are collected, and the incremental data set is constructed, so that automatic updating and optimization of the model are supported.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to an intelligent edge device and an operating method and system thereof. Background Art

[0002] With the development of artificial intelligence technology, AI models are being used more and more widely. Currently, the mainstream artificial intelligence scenarios are training AI models in the cloud and applying AI model reasoning on the edge.

[0003] After deploying the AI ​​model on the edge, the business scenarios on the edge need to continuously update the AI ​​model for reasoning. At the same time, new data will be generated in the business scenarios on the edge, but the AI ​​model deployed on the edge cannot learn automatically. If you want the AI ​​model to become more and more suitable for business scenarios, you need to manually collect these new data, perform data cleaning, data labeling, and then retrain the AI ​​model through the cloud and send it to the edge. The above process is time-consuming and requires the intervention of professional algorithm tuning personnel. Summary of the invention

[0004] In order to solve the above technical problems, the present disclosure provides an intelligent edge device that can automatically update a main model and an operating method and system thereof.

[0005] In a first aspect of the present disclosure, a method for operating an intelligent edge device is provided, wherein the intelligent edge device is deployed with a main model having an inference function, and the method comprises:

[0006] Calling the main model to perform reasoning based on the acquired feature data set to obtain a reasoning result data set;

[0007] Automatically transmit and store the inference data when calling the main model for inference, the inference data at least including the feature data set and the inference result data set;

[0008] An incremental dataset is created based on the collected real result dataset and the feature dataset for updating the main model.

[0009] Optionally, it also includes:

[0010] An interface is defined to automatically transfer and store said real result data set.

[0011] Optionally, it also includes:

[0012] Triggering updating of the main model according to a set time or according to a performance evaluation result of the main model;

[0013] Outputting the incremental data set when updating the main model is triggered; and

[0014] Receive and store the optimized model updated based on the incremental data set.

[0015] Optionally, triggering updating of the main model according to the performance evaluation result of the main model includes:

[0016] Evaluate the performance of the main model based on the inference result dataset and the real result dataset;

[0017] When the performance evaluation result exceeds a performance decay threshold, updating of the main model is triggered.

[0018] Optionally, it also includes:

[0019] Automatically replacing the optimized model with the main model;

[0020] Alternatively, the optimization model is replaced by the main model according to multiple performance evaluation results of the main model and the optimization model.

[0021] Optionally, it also includes:

[0022] Performing performance evaluation on the main model and a plurality of stored auxiliary models, wherein the auxiliary model is at least one of an alternative model and other versions of the main model;

[0023] Determining whether to replace the main model is determined according to performance evaluation results of the plurality of auxiliary models and the main model.

[0024] Optionally, performing performance evaluation on the main model and the stored plurality of auxiliary models includes:

[0025] The same feature data set or different feature data sets are input into the main model and the stored multiple auxiliary models, and the performance evaluation results of the multiple auxiliary models and the main model are obtained based on the inference result data sets output by the main model and the multiple auxiliary models and the true result data sets.

[0026] Optionally, it also includes:

[0027] Pushing the performance evaluation result to the user;

[0028] receiving instructions generated based on the instructions sent back by the user; and

[0029] The instructions are executed to trigger updating of the main model or replacing of the main model.

[0030] A second aspect of the present disclosure provides an intelligent edge device, including:

[0031] The reasoning module deploys the main model with reasoning capabilities;

[0032] An application module calls the main model to perform reasoning based on the acquired feature data set to obtain a reasoning result data set, and obtains a real result data set; and

[0033] A management module automatically transmits and stores inference data when calling the main model for inference, wherein the inference data at least includes the feature data set and the inference result data set, and creates an incremental data set based on the real result data set and the feature data set for updating the main model.

[0034] Optionally, the management module further defines an interface to automatically transmit and store the real result dataset, and creates the incremental dataset based on the stored real result dataset and the feature dataset.

[0035] Optionally, the management module also triggers updating of the main model according to a set time or according to a performance evaluation result of the main model, outputs the incremental data set when updating of the main model is triggered, and receives and stores the updated optimization model based on the incremental data set.

[0036] Optionally, the management module also performs performance evaluation on the main model and multiple stored auxiliary models, and determines whether to replace the main model based on the performance evaluation results of the multiple auxiliary models and the main model, wherein the auxiliary model is at least one of the alternative models and other versions of the main model.

[0037] Optionally, the management module also pushes the performance evaluation result to the user, receives instructions generated based on the instructions sent back by the user, and executes the instructions to trigger updating the main model or replacing the main model.

[0038] A third aspect of the present disclosure provides an operating system for an intelligent edge device, including:

[0039] The smart edge device as described above or the smart edge device executing the method as described above; and

[0040] The cloud service center provides the master model to the smart edge device, and updates the master model based on the received incremental data set to provide the updated optimization model to the smart edge device.

[0041] The intelligent edge device and its operation method and system provided by the present disclosure can quickly realize the automatic collection of incremental reasoning data and real result data sets by defining interfaces at the intelligent edge, and automatically build incremental data sets to realize automatic model updates. Ultimately, the model self-learning capability with zero human intervention is realized.

[0042] It should be noted that the above general description and the following detailed description are merely exemplary and explanatory and do not limit the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A schematic diagram showing the structure of an operating system of an intelligent edge device provided according to an embodiment of the present application is shown;

[0044] Figure 2 A schematic diagram showing the structure of a management module in an intelligent edge device provided according to an embodiment of the present application is shown;

[0045] Figure 3 A schematic diagram of a process of updating a main model of an operating system of an intelligent edge device provided according to an embodiment of the present application is shown;

[0046] Figure 4 A schematic diagram showing a process of updating a main model of an operating system of another intelligent edge device provided in an embodiment of the present application is shown;

[0047] Figure 5 A schematic diagram of a process of replacing a main model of an operating system of an intelligent edge device provided in an embodiment of the present application is shown;

[0048] Figure 6 A schematic diagram showing a process of sharing feature information of an operating system of an intelligent edge device provided according to an embodiment of the present application;

[0049] Figure 7 A schematic diagram of a process of interaction between a smart edge device and a user provided according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0050] In order to facilitate the understanding of the present application, the present application will be described more comprehensively below with reference to the relevant drawings. The preferred embodiments of the present application are given in the drawings. However, the present application can be implemented in different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive.

[0051] Figure 1 A structural schematic diagram of an operating system of an intelligent edge device provided according to an embodiment of the present application is shown. Figure 2 A schematic diagram of the structure of a management module in an intelligent edge device provided according to an embodiment of the present application is shown. Figure 3 A schematic diagram of a process for updating a main model of an operating system of an intelligent edge device provided according to an embodiment of the present application is shown. Figure 4 A schematic diagram of a process for updating a main model of an operating system of another intelligent edge device provided according to an embodiment of the present application is shown. Figure 5A schematic diagram of a process of replacing a main model of an operating system of an intelligent edge device provided according to an embodiment of the present application is shown. Figure 6 A schematic diagram of a process for sharing feature information of an operating system of an intelligent edge device provided according to an embodiment of the present application is shown. Figure 7 A schematic diagram of a process of interaction between a smart edge device and a user provided according to an embodiment of the present application is shown.

[0052] See also Figure 1 The operating system of the intelligent edge device includes an intelligent edge device 110 and a cloud service center 120. The intelligent edge device 110 includes an inference module 112, an application module 111, a management module 113, a container engine 114, and a processing chip 115.

[0053] The intelligent edge device 110 is used for inference and prediction in business scenarios to achieve the purpose of guiding industrial production. It can also automatically trigger the update of the main model, automatically replace the main model, analyze the important features of the main model and share them, interact with users to push model evaluation results and guide the update or replacement of the main model.

[0054] The cloud service center 120 provides the master model to the smart edge device 110, and updates the master model to provide the updated optimization model to the smart edge device 110. The cloud service center 120 can be a private AI training center or an AI center of a mainstream public cloud vendor.

[0055] The reasoning module 112 deploys a main model with reasoning function. The main model takes the AI ​​reasoning model as an example. The AI ​​model is a type of mathematical algorithm model that uses machine learning ideas to solve practical problems.

[0056] The application module 111 calls the main model deployed by the reasoning module 112 to reason the feature data set X to obtain the reasoning result data set Y, and obtain the real result data set Z.

[0057] The management module 113 automatically transmits and stores the inference data when calling the main model for inference, and the inference data at least includes the feature data set and the inference result data set. And based on the real result data set and the feature data set, an incremental data set is created for updating the main model. The management module 113 is also used to automatically trigger the update of the main model, automatically replace the main model, analyze the important features of the main model and share them, interact with the user to push the model evaluation results and guide the update or replacement of the main model.

[0058] The container engine 114 utilizes the kernel of the operating system to implement isolation and packaging of applications, so that the applications can run in different environments.

[0059] The processing chip 115 may be, for example, an ARM (Advanced RISC Machines) or a RISC-V (Reduced Instruction Set Computer-five) computing chip.

[0060] See also Figure 2 The management module 113 is, for example, a program deployed in a containerized manner, which is a preset program of the edge terminal device image, and is automatically started when the edge terminal device is running. The above program is adapted to ARM chips and RISC-V, and can efficiently schedule the performance of the above chips, while being compatible with most mainstream model frameworks.

[0061] The management module 113 includes an application area 1131 , a configuration area 1132 , a storage area 1133 and an interface area 1134 .

[0062] The application area 1131 includes at least a communication unit, a model version information unit, a model performance display unit, a feature management unit, and a model operation unit. The application area is displayed in an interface manner. For example, the model version information unit intuitively displays the model version information. The model performance display unit displays the model performance information. The feature management unit supports feature importance analysis and sharing of feature information containing important features. The model operation unit supports online model performance evaluation and model replacement. The communication unit supports interaction with users through SMS / official account applets.

[0063] The configuration area 1132 includes at least a model update strategy configuration unit, a model performance evaluation loss function configuration unit, and an optimization model online strategy configuration unit.

[0064] The model update policy configuration unit can automatically trigger the update of the main model according to the set time or according to the performance evaluation result of the main model. Exemplarily, the model update policy configuration unit can set the main model to be automatically updated once a month, quarter, half a year, or a custom time period. The model update policy configuration unit can also be set to automatically update when the performance of the main model decays to 95%, 90%, or 85%. Further, when the automatic update of the main model is triggered, for example, a text message notification to the user 130 will be synchronously triggered so that the user (for example, the administrator) can clearly know the progress of the automatic update link.

[0065] The model performance evaluation loss function configuration unit can, for example, set the performance evaluation function (loss function) of the current model to be evaluated. For example, the performance evaluation function for distance measurement in a business scenario at least includes the Euclidean distance loss function. The performance evaluation function for probability distribution measurement in a business scenario at least includes the cross-entropy loss function.

[0066] The optimization model online strategy configuration unit sets the online strategy after the main model is updated, for example, including the optimization model being online immediately after being decentralized or the optimization model being online as an auxiliary model. Exemplarily, after the optimization model is online as an auxiliary model, the optimization model is replaced with the main model according to the multiple performance evaluation results of the main model and the optimization model. Further, for example, when the performance of the optimization model is better than the performance of the main model for three consecutive times, the optimization model is upgraded to the main model. Exemplarily, after the optimization model is online as an auxiliary model, the optimization model can also be replaced with the main model according to user instructions.

[0067] Storage area 1133 includes multiple storage units. In the actual reasoning process, the reasoning data will be automatically stored in storage area 1133. At the same time, it supports interface calls to store the real result data set Z. It also realizes the creation of incremental data sets (for example, including training sets, test sets, and verification sets) when the main model update is triggered, and stores all versions of reasoning models. The multiple storage units include at least a first storage unit, a second storage unit, a third storage unit, and a fourth storage unit. The first storage unit stores the feature data set X, the second storage unit stores the real result data set Z, the third storage unit stores the reasoning result data set Y, and the fourth storage unit stores the reasoning model.

[0068] The interface area 1134 provides a standard Restful interface for transmitting inference data and real result data set Z, as well as the management of inference model versions and the ability to view model performance evaluation results. The RestFul interface provides SDK packages (Software Development Kit) based on python, java, and C languages, which can almost cover various business applications of edge computing. The Restful interface refers to an interface that complies with the REST (Representational State Transfer) principle. The interface area 1134 includes at least a real result data set import interface, a model version management interface, and a model performance viewing interface.

[0069] The present invention uses a layered design concept to set up a management module, which can achieve unified model call interface, automatic storage of incremental data, import of real result data set interface, horizontal comparison of model performance, automatic triggering of model update, and customized model online capability. It solves the problems of high model update cost, high AI technology requirements, and difficulty in grasping model operation status brought by traditional edge devices, and reduces the cost of model development and operation and maintenance.

[0070] See also Figure 3 ,The process of automatically updating the master model includes the following steps.

[0071] The cloud service center 120 first sends the trained AI reasoning model to the smart edge device 110.

[0072] The management module 113 receives and stores the AI ​​reasoning model issued by the cloud service center 120. Furthermore, the management module 113 encapsulates the AI ​​reasoning model into one or more callable interfaces to facilitate its integration into different business scenarios and interact with different data sources and application modules 111. Exemplarily, for example, by pre-defining the Restful interface of the AI ​​reasoning model, the AI ​​reasoning model can be run and reasoning data can be obtained by simply calling the interface operation in the business scenario.

[0073] The reasoning module 112 deploys a main model with reasoning function. The main model takes an AI reasoning model as an example.

[0074] The application module 111 calls the main model to perform reasoning based on the acquired feature data set X to obtain the reasoning result data set Y. The application module 111 implements the calling operation through code, for example, and inputs the feature data set X into the main model for reasoning in the business scenario to obtain the reasoning result data set Y, thereby achieving the purpose of guiding production in the business scenario. The application module 111 also obtains the real result data set Z. Taking the business scenario related to industrial production prediction as an example, the real result data set Z is generally obtained by automatic measurement of sensors.

[0075] When the application module 111 calls the main model for reasoning through the calling interface, the management module 113 will automatically parse, transmit and store the reasoning data generated by the reasoning module 112. The reasoning data includes, for example, the calling time, the calling source IP, the calling input (feature data set X), and the calling result (inference result data set Y). The management module 113 also automatically transmits and stores the real result data set Z collected by the application module 111. Furthermore, for example, the management module 113 automatically obtains the real result data set Z obtained by the application module 111 in real time through a predefined Restful interface. Furthermore, it can also be replaced by obtaining and storing the real result data set Z from the application module 111 at a regular interval through a predefined Restful interface.

[0076] It should be noted that as time goes by, the incremental data sets acquired by the management module 113 will continue to increase. The quality of the incremental data can be ensured by downsampling or automatically reducing the proportion of data sets far from the current time.

[0077] Taking the industrial business scenario of predicting the natural bandwidth of the steel plate of the xx production line as an example, the inference result value is the inference width based on the feature data set X. At the same time, after the steel plate passes through the rolling conveyor belt, the conveyor belt sensor will automatically obtain the true width of the steel plate. The intelligent edge device 110, for example, obtains the true width data set by collecting real-time data from the sensor, and then transmits and stores the true width data set to the management module 113 through an interface call. Fully automated operation is achieved based on the above method. It should be noted that this application takes the industrial business scenario as an example, but the definition interface provided in the embodiment can also be extended to other fields of non-industrial scenarios for use in similar situations.

[0078] This embodiment defines an interface in the intelligent edge device 110 to quickly realize the automatic collection of incremental reasoning data and real result data sets, and automatically build incremental data sets to realize automatic model update, thereby finally realizing the model self-learning capability with zero human intervention.

[0079] See also Figure 4 ,The process of automatically updating the main model also includes the following steps.

[0080] The model update strategy configuration unit in the management module 113 triggers the update of the main model according to the set time or according to the performance evaluation result of the main model. And in the case of triggering the update of the main model, an incremental data set is created based on the stored real result data set Z and feature data set X, and the incremental data set is provided to the training platform 121 of the cloud service center 120. The training platform 121 retrains and updates the main model based on the incremental data set to obtain an optimized model and sends it to the management module 113 for storage in the fourth storage unit. Exemplarily, the triggering of the update of the main model is realized by calling the training interface, for example.

[0081] Furthermore, the performance of the main model is evaluated based on the inference result data set Y and the real result data set Z through the model performance evaluation loss function configuration unit of the management module 113. The model update strategy configuration unit triggers the update of the main model when the performance evaluation result exceeds the performance decay threshold. The performance decay threshold is, for example, 5% or 10%.

[0082] Furthermore, the optimization model is automatically put online as a main model or an auxiliary model through the optimization model online strategy configuration unit of the management module 113.

[0083] See also Figure 5The process of automatically replacing the main model includes: evaluating the performance of the main model and multiple stored auxiliary models, and determining whether to replace the main model according to the performance evaluation results of the multiple auxiliary models and the main model. The auxiliary model is at least one of the alternative model and other versions of the main model.

[0084] The main model is a model used to provide inference results for business scenarios. The auxiliary model is a historical version model or a comparison model that users want to compare horizontally. The auxiliary model does not directly participate in business work. By comparing the main model horizontally and vertically with the auxiliary model, it can reduce the possibility of incorrect guidance results in the production environment due to performance issues after the main model is officially launched. At the same time, it ensures that when the distribution of real business data features changes, the intelligent edge terminal can quickly find the optimal model to guide production business work.

[0085] Exemplarily, the main model and N auxiliary models in the model operation unit are run by model calling, where N is a positive integer. Exemplarily, the same feature data set or different feature data sets are input to the main model and the multiple auxiliary models. The inference data generated when calling the main model and the multiple auxiliary models are automatically stored in the storage area 1133. The model performance evaluation loss function configuration unit performs model performance evaluation based on the inference result data set and the real result data set of the main model and the multiple auxiliary models, and transmits the obtained performance evaluation results of the multiple auxiliary models and the main model to the model performance display unit. Afterwards, the model performance display unit instructs the model operation unit whether to replace the main model based on the performance evaluation results of the multiple auxiliary models and the main model.

[0086] like Figure 6 As shown, the process of automatically sharing feature information includes the following steps.

[0087] The feature management unit in the management module 113 has preset SHAP / Permutation / Boruta / customized feature importance analysis tasks. The feature importance analysis tasks are run in the model running unit at a set time (e.g., daily or weekly) to calculate the feature importance of the main model. The important feature results are then combined with the metadata database of the feature engineering to transmit the feature information containing the important features to the feature management unit.

[0088] The feature management unit then shares the feature information with the user 130 (eg, a designated business person or algorithm person).

[0089] This embodiment can obtain the feature importance ranking results in the business scenario by periodically and automatically running the feature importance extraction task, and at the same time share the important feature results to the feature management unit, and push them to designated users regularly to facilitate the subsequent update and tuning of the model, and also facilitate the rapid construction of feature engineering in similar scenarios.

[0090] See also Figure 7 ,The process of interaction between the intelligent edge device and the user includes the ,following steps.

[0091] The communication unit in the management module 113 includes a data communication unit, an instruction generation unit, and an instruction analysis unit.

[0092] First, the model performance evaluation result is pushed to the designated user 130 through the data communication unit. Exemplarily, the model performance evaluation result in the model performance display unit is transmitted to the data communication unit, and then the data communication unit pushes it to the designated user 130 through the official account applet / text message. After receiving the model performance evaluation result report, the designated user 130 can choose whether to perform a model tuning and upgrading operation, or choose whether to perform a model replacement operation.

[0093] The user 130 transmits the instruction back to the instruction analysis unit. The instruction analysis unit analyzes the user's instruction. The instruction generation unit generates corresponding instructions and executes the instructions in the model operation unit to implement automatic replacement operation or automatic update operation or customized optimization model online operation.

[0094] This embodiment regularly pushes the actual performance evaluation results of the model to relevant business personnel through the WeChat official account app / text message. After understanding the model performance, the business personnel can trigger the same model tuning / different model replacement and other instruction operations to enable the business personnel to lead the main model update or replacement operation.

[0095] The intelligent edge device proposed in this disclosure realizes the ability of edge reasoning models to learn while using them in the actual business operation process through the combination of software and hardware, cloud-edge collaboration, and human-computer interaction.

[0096] The present disclosure is aimed at smart industrial scenarios and can realize truly fully automated edge inference model updates. For business personnel, it can achieve the purpose of triggering model updates according to the actual needs of business scenarios. For algorithm personnel, it can realize automatic acquisition of important features, reduce the input cost of feature engineering for algorithm personnel, and thus continuously improve model performance. For smart edge terminals, it can realize online vertical comparison between different versions of the same model, as well as online horizontal comparison between different models, breaking the current situation where only offline horizontal / vertical comparison can be performed.

[0097] It should be noted that the numerical values ​​in this article are only used for exemplary description. In other embodiments of the present application, other numerical values ​​can also be sampled to implement this solution. The specific settings should be reasonable according to the actual situation, and this application does not limit this.

[0098] Finally, it should be noted that: Obviously, the above embodiments are only examples for clearly explaining the present application, and are not intended to limit the implementation methods. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from this are still within the scope of protection of the present application.

[0099] It should also be understood that the terms and expressions used herein are for descriptive purposes only, and one or more embodiments of this specification should not be limited to these terms and expressions. The use of these terms and expressions does not mean to exclude any equivalent features of the illustrations and descriptions (or parts thereof), and it should be recognized that various modifications that may exist should also be included in the scope of the claims. Other modifications, changes and substitutions may also exist. Accordingly, the claims should be deemed to cover all such equivalents.

Claims

1. A method for operating an intelligent edge device, wherein the intelligent edge device is deployed with a main model having an inference function, wherein: The method comprises: Calling the main model to perform reasoning based on the acquired feature data set to obtain a reasoning result data set; Automatically transmit and store the inference data when calling the main model for inference, the inference data at least including the feature data set and the inference result data set; An incremental dataset is created based on the collected real result dataset and the feature dataset for updating the main model.

2. The method for operating the intelligent edge device according to claim 1, wherein: Also includes: An interface is defined to automatically transfer and store said real result data set.

3. The method for operating the intelligent edge device according to claim 1, wherein: Also includes: Triggering updating of the main model according to a set time or according to a performance evaluation result of the main model; outputting the incremental data set when triggering updating of the main model; as well as Receive and store the optimized model updated based on the incremental data set.

4. The method for operating the intelligent edge device according to claim 3, wherein: Triggering updating of the main model according to the performance evaluation result of the main model includes: Evaluate the performance of the main model based on the inference result dataset and the real result dataset; When the performance evaluation result exceeds a performance decay threshold, updating of the main model is triggered.

5. The method for operating the intelligent edge device according to claim 3, wherein: Also includes: Automatically replacing the optimized model with the main model; Alternatively, the optimization model is replaced by the main model according to multiple performance evaluation results of the main model and the optimization model.

6. The method for operating the intelligent edge device according to claim 1, wherein: Also includes: Performing performance evaluation on the main model and a plurality of stored auxiliary models, wherein the auxiliary model is at least one of an alternative model and other versions of the main model; Whether to replace the main model is determined according to performance evaluation results of the plurality of auxiliary models and the main model.

7. The method for operating the intelligent edge device according to claim 6, wherein: The performance evaluation of the main model and the stored plurality of auxiliary models includes: The same feature data set or different feature data sets are input into the main model and the stored multiple auxiliary models, and the performance evaluation results of the multiple auxiliary models and the main model are obtained based on the inference result data sets output by the main model and the multiple auxiliary models and the true result data sets.

8. The method for operating an intelligent edge device according to any one of claims 3 to 7, wherein: Also includes: Pushing the performance evaluation result to the user; receiving instructions generated based on the instructions sent back by the user; as well as The instructions are executed to trigger updating of the main model or replacing of the main model.

9. An intelligent edge device, wherein: include: The reasoning module deploys the main model with reasoning capabilities; An application module calls the main model to perform reasoning based on the acquired feature data set to obtain a reasoning result data set, and obtains a real result data set; as well as A management module automatically transmits and stores inference data when calling the main model for inference, wherein the inference data at least includes the feature data set and the inference result data set, and creates an incremental data set based on the real result data set and the feature data set for updating the main model.

10. The intelligent edge device according to claim 9, wherein: The management module also defines an interface to automatically transmit and store the real result dataset, and creates the incremental dataset based on the stored real result dataset and the feature dataset.

11. The intelligent edge device according to claim 9, wherein: The management module also triggers updating of the main model according to a set time or according to a performance evaluation result of the main model, outputs the incremental data set when updating of the main model is triggered, and receives and stores an optimized model updated based on the incremental data set.

12. The intelligent edge device according to claim 9, wherein: The management module also performs performance evaluation on the main model and multiple stored auxiliary models, and determines whether to replace the main model based on the performance evaluation results of the multiple auxiliary models and the main model, wherein the auxiliary model is at least one of an alternative model and other versions of the main model.

13. The intelligent edge device according to any one of claims 11 to 12, wherein: The management module also pushes the performance evaluation result to the user, receives instructions generated based on the instructions sent back by the user, and executes the instructions to trigger the update of the main model or the replacement of the main model.

14. An operating system for an intelligent edge device, wherein: include: An intelligent edge device as described in any one of claims 9 to 13 or an intelligent edge device performing the method as described in any one of claims 1 to 8; as well as The cloud service center provides the master model to the smart edge device, and updates the master model based on the received incremental data set to provide the updated optimization model to the smart edge device.

Images