KR20210036226A - A distributed computing system including multiple edges and cloud, and method for providing model for using adaptive intelligence thereof - Google Patents

Abstract

According to the present invention, a method for providing a model for utilizing adaptive intelligence in a distributed computing environment including a plurality of edges and a cloud that receives data from the edges, processes the same, and delivers the processing results to the edges comprises the following steps of: generating a model for collected data in a cloud; storing the generated model in a model storage; distributing an application from the cloud to edges; and searching and downloading, by the edges, one or more models matching the distributed application from the model storage.

Description

A DISTRIBUTED COMPUTING SYSTEM INCLUDING MULTIPLE EDGES AND CLOUD, AND METHOD FOR PROVIDING MODEL FOR USING ADAPTIVE INTELLIGENCE THEREOF}

The present invention relates to a system and a method of providing an analysis model for utilizing adaptive intelligence in a distributed computing environment.

With the spread of IoT devices, it has become possible to collect various data from IoT sensors in various places.

For the collected sensor data result value, simple analysis is applied to understand the meaning according to a predetermined rule in the field, and for professional and advanced analysis, the collected data is sent to the cloud for in-depth analysis, and the result is sent back to the user. This method is commonly used.

In this cloud-based analysis, managing and analyzing data in cloud infrastructure, such as Google Big Query and Amazon Web Services, is a common technology trend at this time.

As described above, with the development of cloud computing technology, data learning and analysis using the cloud is becoming popular, but there are several problems as follows.

-Privacy invasion problem: Privacy invasion may occur as individual data needs to be transmitted to the cloud.

-Network Dependency: Sending data to the cloud means sending data to a physical/logically distant place and receiving the result, so if the network does not work properly, it is impossible to provide a service.

-Network resource consumption: Even if the performance of the network is supported, the consumption of bandwidth is large because a large amount of data needs to be transmitted, and service may become difficult if a large amount of resources is used.

-Response speed: It is difficult to use for real-time services that require fast feedback because data must be sent and received in a physically distant place.

Meanwhile, edge or fog computing is being introduced as an alternative to overcome the shortcomings of cloud analysis.

Edge was developed to overcome the physical distance and limitations of the cloud, and it is a system that can be used privately in a location close to the place where users use the service.

The edge can overcome the shortcomings of the cloud by performing the necessary functions in the middle of the cloud and the user.

However, in the existing edge-cloud collaboration analysis framework, because the analysis engine is simply deployed from the cloud to the edge, and the edge simply delegates to utilize the analysis model (AI model or machine learning model), the current edge Base analysis is an early stage of technology development and is suitable for performing general and simple tasks, but it cannot consider various individual edge characteristics and has a problem that is dependent on a specific company's cloud system.

In addition, edge processing performance is not as good as the cloud, and it is difficult to process a large amount of data, making it difficult to process complex calculations, and the number and characteristics of tasks that can be processed for each edge are very different due to the large difference in performance for each edge. It is difficult to distribute models in batches.

On the other hand, in all cases, designing and implementing an AI model requires development time and expertise. For this reason, experienced AI models should be able to be reused for a variety of purposes, such as saving to a file, comparing AI models to other AI models, testing models with new data, and so on.

In this regard, many open source machine learning libraries such as TensorFlow and Caffe provide useful classes for saving and restoring models to reduce redundant work. They can store and load trained model information, including variables, graphs, and metadata of graphs.

Meanwhile, the mechanism for converting a machine learning model into a byte stream is called serialization, and deserialization refers to the reverse process of re-creating an actual AI model in memory using a serialized format.

Recent cloud platforms provide online machine learning service hosting solutions, which can host fully trained machine learning models in a public cloud environment. The prerequisite for these services is to first register a supported serialized AI model.

However, these platforms are currently focused on registering and using common AI models for specific edge devices.

Edge devices have relatively low performance, are located in a local area, and are likely to process data containing personal information.

Therefore, cloud platforms need to be used in various business areas through a model deployment method that considers these characteristics of the edge.

In addition, a more structured model management system is required to efficiently use machine learning and deep learning models at the edge.

In an embodiment of the present invention, in order to solve the above-described problem, an edge can more flexibly select an AI model, and when a single artificial intelligence model is divided into two or more parts, it is distributed to be used by being distributed to several edges. Provides a method of providing a system and model for adaptive intelligence utilization in a computing environment.

In addition, an embodiment of the present invention provides a system and method that supports smooth communication and collaboration between an edge and a cloud or a plurality of edges, and enables distribution of a model to an edge and selection of a model by the edge.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problem as described above, and other technical problems may exist.

As a technical means for achieving the above-described technical problem, adaptation in a distributed computing environment including a plurality of edges according to the first aspect of the present invention, and a cloud that receives data from the edges and processes them and transfers the processing results to the edges. A method of providing a model for utilizing enemy intelligence comprises: generating a model for the collected data in the cloud; Storing the generated model in a model storage; Distributing an application from the cloud to the edge; And searching and downloading one or more models in which the edge matches the deployed application from the model repository.

In an embodiment, the storing of the generated model in the model storage may include metadata information required for searching of the edge in the model, and may be stored in the model storage.

In an embodiment, the model stored in the model storage may include basic information for the search, model division-related information, and model training information.

In an embodiment, the basic information for the search may include one or more of model usage cost information, model performance information, device performance information required for the model, and personal information protection capability information.

In one embodiment, the model repository includes one or more of information on whether to divide the model, version information, and download file information, and searches and downloads one or more models whose edges match the deployed application from the model repository. In the step of, the edge may access and download the model based on information stored in the model repository.

In an embodiment, the information on whether to divide the model may include directory information in which all undivided models are stored and directory information in which a model divided into two or more models is stored.

In one embodiment, the step of searching and downloading one or more models in which the edge matches the deployed application from the model repository, the searched using the Rest API, as the edge searches for the matching model. The model is downloaded, but the Rest API may include name information of the searched model, version information, information about whether to be divided, and information about the order of division of the corresponding model.

In one embodiment, the method for providing a model according to the present invention includes the plurality of edges based on the downloaded model for distributing and cooperative processing of tasks between the cloud and the edge or for distributing and cooperative processing of tasks between the edges. It may further include the step of processing the task individually or in collaboration.

In addition, a distributed computing system including a plurality of edges according to the second aspect of the present invention and a cloud that receives and processes data from the edges and transmits the processing results to the edges generates a model for processing the collected data, and , The generated model is stored in a model storage, and a cloud for distributing an application to the edge, a model storage for storing and managing the generated model, and one or more models matching the deployed application are searched from the model storage and Includes Edge to download. At this time, for the distribution and cooperative processing of the task between the cloud and the edge, or for the distribution and cooperative processing of the task between the edges, the plurality of edges respectively or cooperately process the task based on the downloaded model.

In one embodiment, the cloud includes metadata information necessary for the edge search in the model and stores it in the model storage, but the metadata information includes basic information for the search, information related to model division, and a model. Training information may be included.

In an embodiment, the basic information for the search may include one or more of model usage cost information, model performance information, device performance information required for the model, and personal information protection capability information.

In one embodiment, the model storage includes at least one of information on whether to divide the model, version information, and download file information, and the edge may access and download the model based on information stored in the model storage. have.

In an embodiment, the information on whether to divide the model may include directory information in which all undivided models are stored and directory information in which a model divided into two or more models is stored.

In one embodiment, as the edge searches the matching model, the searched model is downloaded using a Rest API, but the Rest API is the name information, version information, partitioning information, and corresponding information of the searched model. It may include information about the order of division of the model.

In addition, a distributed computing system including a plurality of edges according to the third aspect of the present invention, and a cloud that receives and processes data from the edge and transmits the processing result to the edge, the cloud is used for processing the collected data. A model generation unit for generating a model, a model storage unit for storing the generated model in a model storage unit, and a distribution unit for distributing an application to the edge, wherein the edge includes at least one model matching the deployed application And a model download unit for searching and downloading from the model repository.

In addition to this, another method for implementing the present invention, another system, and a computer-readable recording medium for recording a computer program for executing the method may be further provided.

According to any one of the above-described problem solving means of the present invention, the edge is not dependent on the specificity of a fixed application and can actively selectively use a desired model through a cloud or a model storage. At this time, the edge can be selected based on various conditions such as model performance, purpose, and price.

In addition, since the model and the application are separated and provided as an edge, there is an advantage that more flexible use is possible.

In addition, even after the application is distributed, there is an advantage that it is possible to update the model according to the situation. For example, in the case of the prior art, since the model is fixed, there is a problem that it is difficult to use at a certain edge in some cases, but according to an embodiment of the present invention, the edge flexibly selects a model suitable for its own situation conditions. You will be able to.

1 is a schematic diagram illustrating a cloud and edge-based computing environment.
2 is a view for explaining the difference between the prior art and an embodiment of the present invention.
3 is a diagram illustrating a distributed computing system according to an embodiment of the present invention.
4 is a diagram illustrating an example for explaining the structure of a model storage.
5 is a diagram illustrating an example of a metadata structure.
6 is a diagram for explaining hardware for a cloud and an edge according to an embodiment of the present invention.
7 is a flowchart of a method for providing a model according to an embodiment of the present invention.
8 is a conceptual diagram of a cloud-edge distributed processing method according to the first embodiment.
9 is an exemplary configuration diagram of a cloud-edge for implementing the concept of the first embodiment shown in FIG. 8.
10 is a diagram showing a conceptual diagram of a cloud-edge distributed processing method according to a second embodiment.
11 is an exemplary configuration diagram of a cloud-edge for implementing the concept of the second embodiment shown in FIG. 10.
12 is an exemplary conceptual diagram of a cloud-edge distributed processing method according to a third embodiment.
13 is an exemplary configuration diagram of a cloud-edge for implementing the concept of the third embodiment shown in FIG. 12.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In the drawings, parts not related to the description are omitted in order to clearly describe the present invention.

In the entire specification, when a certain part "includes" a certain component, it means that other components may be further included rather than excluding other components unless otherwise stated.

Hereinafter, in the description of the present invention, first, a cloud and edge-based computing environment will be outlined, and an embodiment of the present invention in which a model stored in a model storage at an edge is selected in order to solve the problems and problems of the current technology will be described. do. And finally, how to distribute the model from the cloud server to the edge will be additionally explained.

1 is a schematic diagram illustrating a cloud and edge-based computing environment.

Distributed computing system according to an embodiment of the present invention is connected to the things (Things) 30, which is a comprehensive component including a plurality of IoT sensors (not shown) and a plurality of actuators (not shown), and various Multiple Edges (20: 20a, 20b, 20c, ...) that execute edge computing, such as performing processing to receive data and send control signals to it, and execute cloud computing by receiving data from the edge. It consists of a cloud (10) that delivers the execution result to the edge.

For example, IoT sensors, applications, and user devices of the object side 30 produce a lot of data (eg, big data).

The edge 20 basically collects data (big data) from the object side 30; Pre-processing such as data purification for the use of big data, cleaning, sampling, and combining for the main processing of big data is performed, and the result is delivered to the cloud 30.

The functions of the edge 20 can be designed in various ways, for example, it can be designed to process itself without sending it to the cloud 30, performing only the basic functions and passing all of the core tasks to the cloud 30 It can also be designed to be

The cloud 30 mainly performs deep learning such as deep learning and related analysis, inference, etc., comprehensively performing the task(s) handed over from the edge 20, or distributing some of the tasks to a specific edge 20, etc. , Carry out the core processing.

The result processed by the cloud 30 (or the result processed by the edge 20) is applied to the actuator of the object side 30 to perform IoT control based on the situation and judgment.

In order to implement the distributed processing method between the cloud and the edge as in the present invention, basically, a distribution method or distribution rule for performing an operation allot between the edge 20 and the cloud 30 is required. According to the design of this distribution rule, the cloud-edge distributed processing method of the present invention may be implemented in various ways.

Meanwhile, an example of the prior art for generating and distributing a model in a distributed computing environment including a cloud and an edge is as follows.

Amazon announced'AWS DeepLense', a tool that makes'machine learning' easier to use. In addition, Amazon introduced a technology to eliminate the difficulties that arise when using machine learning on a local device as a service that can easily create, train, and deploy machine learning models called'Amazon SageMaker'.

In addition, Caffe plays a role of storing and providing models working with various architectures and data created by many engineers in the Model Zoo.

In addition,'Amazon Azure machine learning service' proposes a method for deploying, running, and managing models that have already been created on devices, and provides functions that make them available at the edge.

Large companies such as Amazon are taking a method of distributing trained AI models to devices in order to use them by giving intelligence to the edge.

As an example, the above technologies generally follow the following procedure.

1.Register the model

2. Prepare to deploy (specify assets, usage, compute target)

3. Deploy the mode to the compute target

4. Test the deployed model, also called web site

In such a conventional technique, first, in order to run an AI application program at the edge, it is common to deploy a container including a specific AI model in the cloud to the edge together.

In other words, to run an AI application at the edge, a container containing a specific AI model must be deployed from the cloud to the edge.

To do this, first register the model, containerize the application with the registered model, and then deploy the container to the target edge.

Containerized applications like this are characterized in that they are different depending on the specific AI model. For example, it is arranged in one direction to the edge, depends on the initial artificial intelligence model, and it is not easy to correct it in consideration of changes in the situation of the edge in real time. Also, the AI model of the container device cannot be changed depending on the performance and purpose of the edge.

And since each edge has a variety of performance and purpose, the edge must be able to dynamically select the AI model in order to flexibly use the AI learning model at this edge.

To this end, an embodiment of the present invention may provide a function to dynamically select the correct intelligence in terms of the edge by separating the AI model and the application program container.

In addition, according to an embodiment of the present invention, the additional model storage and the edge communicate with each other, so that the optimal model can be downloaded directly through the application program container.

Second, in the prior art, there is a case in which a method considering the case of connecting and using two or more application containers is required.

In other words, a single device typically uses a single independent AI model, but not all devices are capable of running a specific AI application due to performance limitations in terms of inference at the edges with varying performance. Therefore, it is necessary to consider another method of dividing an AI application program and distributing it across multiple edges.

For example, assume two edges are A and B. A is an authenticated device, but its performance is poor, while B is a high-performance device that is not authenticated to the user. If the user is allowed to store data only in A and process the original data due to privacy concerns, there are cases in which the task cannot be performed properly due to limited performance.

The best way to solve this problem is collaboration between A and B. Filtered functional information from the original data without personal information is sent from A to B, and after sending the selected information, B can complete the work with this partial information. Through these collaborative scenarios, it will be possible to create and run flexible AI applications at various edges.

2 is a view for explaining the difference between the prior art and an embodiment of the present invention.

In the method for providing a system and model for adaptive intelligence utilization in a distributed computing environment according to an embodiment of the present invention, a containerized application to be deployed at the edge does not include a main AI model fixedly, and the edge is a model. It is characterized by directly fetching the required model from the repository.

In other words, Edge may receive a fixed AI model from a cloud server, but it is characterized in that it can search for and download and use an appropriate AI model from the model repository according to the real-time needs of the service.

In addition, an embodiment of the present invention is characterized in that intelligence is shared between two or more edges. In other words, in order to actively use the device in an evolving edge environment, intelligence can be shared between each edge device. Accordingly, an embodiment of the present invention may share AI data (AI model) and an application between a plurality of edges through a model storage.

Referring to FIG. 2, in the case of the prior art (a), an application including a model is unilaterally distributed to the edge, but according to an embodiment (b) of the present invention, an application distributed to the edge is required to determine the required model in consideration of the situation It is characterized by being searched and provided in the model repository.

3 is a diagram illustrating a distributed computing system 100 according to an embodiment of the present invention.

The distributed computing system 100 according to an embodiment of the present invention includes a cloud 10, a model storage A, and a plurality of edges 20.

The cloud 10 receives data from the edge 20, processes it, and transmits the processing result to the edge 20. In addition, the cloud 10 generates a model for processing the collected data, stores the generated model in the model storage A, and distributes an application without a model to the edge 20.

That is, in the case of the prior art, the cloud was to distribute the application including the AI model, but in an embodiment of the present invention, the cloud 10 distributes the application without the AI model to the edge 20, and (20) is characterized by self-searching a suitable AI model through the cloud 10 or the model storage (A) in the future.

The model storage A stores and manages the model generated by the cloud 10.

4 is a diagram showing an example for explaining the structure of the model storage (A).

The model storage A is configured to include one or more of information on whether to divide a model, information on a version of a model, and information on a download file.

As an example, FIG. 4 shows a structure in which several modified forms of the modified VGG16 model are stored, and the structure of each model may be divided into a full and a partial.

Full refers to directory information in which general undivided models are stored, and Partial refers to directory information in which two or more divided models are stored.

In one embodiment, each of these models may be managed for each version.

In addition, the subfolder contains the metadata of the saved model and a JSON file that describes the file to be downloaded from the edge.

Based on this structure, the edge 20 in an embodiment of the present invention can access and download a model based on information stored in the model storage A.

Referring back to FIG. 3, the edge 20 searches and downloads one or more models matching the application distributed from the cloud 10 from the model storage A.

In this case, the edge 20 may search and download only one model suitable for a specific application, as well as search and download a plurality of models. In addition, in the case of a plurality of models, the edge 20 may find and download the most suitable model.

Meanwhile, in an embodiment, the cloud 10 may store metadata information necessary for model search of the edge 20 in the model to be included in the model storage A.

That is, in order for the edge 20 to search for a model stored in the model storage A, each model must include metadata information necessary for the search.

5 is a diagram illustrating an example of a metadata structure.

The metadata information may include basic information for searching (model_information), model division related information (partial_information), and model training information (training_information), as in the example illustrated in FIG. 5.

Here, the basic information for the search (model_information) includes basic information for the edge 20 to perform a model search. As an example, it may include one or more of model usage cost information, model performance information, device performance information required for the model, and personal information protection capability information, and based on this, the edge 20 searches for a model suitable for itself. I can.

Specifically, basic information (model_information) for searching may include the following information related to searching.

-Minimum specifications for model operation (GPU availability/specifications, CPU specifications, minimum memory size, etc.), model price, model performance, model purpose, model input and output types, model supplier information, model processing speed , Whether cooperation between devices is possible and its information (Full, Partial)

Meanwhile, according to an exemplary embodiment of the present invention, as the edge 20 searches for a matching model, the searched model may be downloaded using the Rest API.

In this case, the Rest API may include name information of the searched model, version information, information on whether to be divided, and information on the order of division of the corresponding model.

E.g,

In the form of http://server.com/model?model_name=VGG16&model_version=[v1.0]&partial_mode=Partial&partial_num=1, VGG is 16, model version is 1, partial_mode is divided, and the first divided model It can be used as a Rest API to download.

Name Required/
Optional Type Description Example Model_name Required String Provides multiple AI models that can be classified and predict future conditions
-First of all, Edge must know the correct model name to download, and the application container must be able to download and manage the model based on the described information. model_name=Yolo_v3
Model_version Optional String Versioning of the AI model Must maintain a history of changes as the AI model can be modified or improved
-The model version can be downloaded and used by the method proposed in the present invention. model_version=[v1.0] Partial_mode Required String There are two types of models: Full and Partial. According to the method proposed in the present invention, tasks can be performed independently. Can support cloud/edge or edge/edge working together
-The model repository is characterized by being able to provide a single AI model or a divided AI model. partial_mode=Full Partial_number Optional integer Segmented order information is required to download a model suitable for the edge. For full models, the part number is always 0.
-If a model divided into N is used, model information set from 1 to N is required. partial_number=1

In this way, when the edge 20 downloads an application and one or more models matching it, for the distribution and collaboration of tasks between the cloud 10 and the edge 20, or the distribution of tasks between a plurality of edges 20 and For collaborative processing, a plurality of edges 20 cooperate with each other or each other to process a task based on the downloaded model. On the other hand, a sub-directory called Partial stores the data of the divided model, and the model has multiple edges. (20) It is divided into several sub-models for collaboration and then saved. Here, subdivisions of the partial directories P1 and P2 may be grouped by version. In addition, the sub-directory may include order information of the divided models.

Meanwhile, in FIG. 3 in which the cloud 10 and the model storage A are illustrated in an exemplary embodiment of the present invention, they are illustrated as being independent configurations, but are not limited thereto. That is, each of the above functions may be provided through a plurality of programs in one server, and the servers composed of a plurality of hardware may each be configured as independent server computers and implemented in a form in which they are interoperable.

6 is a diagram for explaining hardware of the cloud 10 and the edge 20 according to an embodiment of the present invention.

In addition, in an embodiment of the present invention, the cloud 10 and the edge 20 are a communication module 110, a memory 120, and a processor that executes a program stored in the memory 120 as shown in FIG. 130).

The communication module 110 is preferably composed of a wireless communication module, but this does not exclude a wired communication module. The wireless communication module may be implemented with WLAN (wireless LAN), Bluetooth, HDR WPAN, UWB, ZigBee, Impulse Radio, 60GHz WPAN, Binary-CDMA, wireless USB technology, and wireless HDMI technology. In addition, the wired communication module may be implemented as a power line communication device, a telephone line communication device, a cable home (MoCA), Ethernet, IEEE1294, an integrated wired home network, and an RS-485 control device.

The memory 120 collectively refers to a nonvolatile storage device and a volatile storage device that continuously maintains stored information even when power is not supplied. For example, the memory 120 may be a compact flash (CF) card, a secure digital (SD) card, a memory stick, a solid-state drive (SSD), and a micro SD card. Magnetic computer storage devices such as NAND flash memory, hard disk drives (HDDs) such as cards, and optical disc drives such as CD-ROMs, DVD-ROMs, etc. I can.

For reference, the components shown in FIGS. 1 to 6 according to an embodiment of the present invention may be implemented in software or in hardware such as a field programmable gate array (FPGA) or an application specific integrated circuit (ASIC). Can play roles.

However,'components' are not meant to be limited to software or hardware, and each component may be configured to be in an addressable storage medium or configured to reproduce one or more processors.

Thus, as an example, components are components such as software components, object-oriented software components, class components, and task components, processes, functions, properties, procedures, and sub- Includes routines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays and variables.

Components and functions provided within the components can be combined into a smaller number of components or further divided into additional components.

Hereinafter, a method of providing a model for utilizing adaptive intelligence in the distributed computing system 100 according to an embodiment of the present invention will be described with reference to FIG. 7.

7 is a flowchart of a method for providing a model according to an embodiment of the present invention.

In the method for providing a model according to an embodiment of the present invention, first, a model for data collected from the cloud is generated (S110).

Next, the generated model is stored in the model storage (S120).

Next, the application is distributed from the cloud to the edge (S130).

Next, one or more models matching the application to which the edge is deployed are searched and downloaded from the model repository (S140).

In the above description, steps S110 to S140 may be further divided into additional steps or may be combined into fewer steps, according to an embodiment of the present invention. In addition, some steps may be omitted as necessary, or the order between steps may be changed. In addition, even if other contents are omitted, the contents already described with respect to the distributed computing system 100 in FIGS. 1 to 6 are also applied to the method of providing a model in FIG. 7.

Hereinafter, a method of distributing a model from the cloud 10 to the edge 20 will be additionally described with reference to FIGS. 8 to 13.

In order to implement the distributed processing method between the cloud 10 and the edge 20 as in the present invention, a distribution method or a distribution rule is basically required to perform an operation allot between the edge 20 and the cloud 10. Do. According to the design of this distribution rule, the cloud-edge distributed processing method of the present invention may be implemented in various ways.

Meanwhile, in the description of the present invention, the term'operation' has been used to encompass all actions such as data collection, processing, learning, analysis, and reasoning performed in an edge-cloud based IoT computing environment. ,'Model' or'engine' was used to encompass learning models (engines), analytic models (engines), inference models (engines), and AI models required for IoT processing.

In the first embodiment, a plurality of models are generated in the cloud 10 and some of the generated models are distributed to the edge 20 to perform'lightweight' processing at the edge 20 to perform edge-cloud distributed processing. This is the way to do it. According to this concept, for convenience, the first embodiment is referred to as'selective model distribution technique for utilizing a lightweight model' or model weight reduction.

8 is a conceptual diagram of a cloud-edge distributed processing method according to the first embodiment.

According to FIG. 8, models or engines 11a, 11b, 11c, 11d, 11e, ...) are generated in the cloud 10. In the IoT environment, when it is difficult to generate and learn/analyze such models at the edges (20a, b, c, ...), a number of models (engines) for the main operation are used in the cloud 10. It is generated, and processed by importing (or distributing) only a model that can be processed by itself at the edge 20.

Model allocation from the cloud 10 may be led from the edge 20 side or the cloud 10 side according to the situation of the edge (processing performance, memory capacity, etc.).

Depending on the edge's performance, the type and number of models distributed to the edge may vary.

In FIG. 8, one model 11a is distributed to the first edge 20a, two models 11b and 11e are distributed to the second edge 20b, and one model 11c is distributed to the third edge 20c. Is shown.

Since there are a variety of models created in the cloud 10 such as large, small, high-performance, and low-performance, among these models, select the one that suits the performance or capacity of each edge 20 so that appropriate processing results can be produced. By doing so, it can maximize the efficiency of the entire IoT computing system.

In other words, in order for each edge to produce the best performance it can achieve, the model is shared among multiple edges and processed (distributed operation).

9 shows a configuration diagram of a cloud 10-an edge 20 for implementing the concept of the first embodiment shown in FIG. 8. Basically, there is a cloud computing unit 12 in the cloud 10 to perform the original cloud operation (processing), and the model generation unit 14 generates a model.

That is, the cloud 10 stores large amounts of data based on high processing power and storage capacity, and creates one or more models suitable for various situations. In addition, the edge 20 includes an edge computing unit 21 to perform operations (processing) required for the edge 20.

In order to implement the edge-cloud distributed processing according to the first embodiment, a pre-determined model distribution rule (13, 22) is added to the cloud 10 and the edge 20, and the first implementation is carried out based on this rule. You can proceed with courtesy techniques.

The model selection unit 23 of the edge 20 uses the model distribution rule 22 to determine which model generated in the cloud 10 and how to use it according to its dynamically changing situation and selects a model.

Model selection at the edge 20 may determine the number of models and the number of models that can be processed in consideration of two things: 1) fixed process processing capability of the edge 20, and 2) process resources available in real time. .

When the edge 20 selects a model (23), the model distribution unit 15 of the cloud 10 corresponds to the selected model according to the model distribution rule 13 of the cloud linked to the model distribution rule 22 of the edge. Distribute to the edge (20). The edge computing unit 21 that performs the original edge operation (processing) operates on the edge 20 to perform processing using the distributed AI model.

In this way, by dividing one or more models according to distributed processing rules, each edge 20 selectively uses a predictable lightweight model, thereby increasing the prediction rate of the entire IoT system and flexibly responding to all situations.

In the above and below, the components of the cloud 10 are converted to the cloud computing unit 12, the model generation unit 14, the model distribution unit 15, and the components of the edge 20 are edge computing. The part 21 and the model selection part 23 were named as'~ part', but this is only for convenience for clarification. Although these names may imply a physical existence or a hardware device, it is obvious to those skilled in the art that they mean not only a hardware device but also a non-physical entity or a function, routine, processing unit, etc. of software. Hereinafter, the same applies to the other second and third embodiments.

The second embodiment is based on exchanging intermediate data processed by the cloud and the edge. According to this concept, the second embodiment may be referred to as a'processing enhancement technique based on organic cooperation between cloud and edge'.

There are cases in which things should not send critical information (for example, information that needs privacy protection such as personal information, photos, voices, etc.) to the cloud. However, it is often difficult for the edge to process information due to problems such as the size and type of the data. This is because the edge is generally weak in processing power.

To solve this problem, the cloud gradually optimizes the analysis engine (model) through learning and distributes it to the edge.

Edge separates the collected data from the object side and immediately performs primary learning/analysis or pre-processing on a part and sends selective data to the cloud.

The cloud processes this process based on the data received from the edge, and quickly delivers the decision-making results to the edge.

In this way, the original data from the edge is transmitted to the cloud in a state of primary learning or pre-processing and transmitted in a form that cannot be reverse-analyzed, thereby protecting personal information (privacy), while sending the primary learning/analysis results to the cloud, and more than the original data. Since it is possible to send less data, it can save transmission bandwidth.

10 is a conceptual diagram of a cloud-edge distributed processing method according to a second embodiment.

(31)를 보내면 엣지(20)는 이 정보 중에서 A와 B를 분리한다. Critical information from the object side (30) to the edge (20)

Sending (31) the edge 20 separates A and B from this information.

(32)는 그대로 클라우드(10)로 보내고 엣지(20)에서는 다른 일부분(A 부분)만을 처리하고 그 처리결과인

(33)를 클라우드(10)로 보낸다. Separate parts, for example

(32) is sent to the cloud 10 as it is, and at the edge 20, only the other part (part A) is processed, and the result of the processing is

Send (33) to the cloud (10).

(32)와

(33)를 기반으로 처리한 완전한 처리 결과

(34)를 엣지(20)로 보내 준다. Cloud (10) is received separately

(32) and

(33) Complete treatment results based on

Send (34) to the edge (20).

The edge 20 receives this and transmits it to the object side 30 as final result data. To this end, in the cloud 10, a model 16 that can process the entire critical information 31 in advance, a model that can process only part A (16'), and a model that can process only part B (16") And the model 16' is distributed to the edge 20 so that the edge 20 can process only part A.

That is, the cloud 10 is to give the edge 20 a role appropriate to the processing capability of the edge 20, and this role assignment is preferably set to a range so that only processing limited to the edge 20.

In this way, the edge 20 does not send the entire critical information 31 to the cloud 10, and separates it and collaborates to process the edge 20 and the cloud 10 separately, thereby eliminating the concern of privacy invasion. have. Therefore, the separation of data should be performed to the extent that the privacy invasion problem does not occur only with the separated parts A and B.

FIG. 11 shows an exemplary configuration diagram of a cloud 10-an edge 20 for implementing the concept of the second embodiment shown in FIG. 10.

Basically, there is a cloud computing unit 12 in the cloud 10 to perform the original cloud operation (processing), and the model generation unit 14 generates a model.

That is, the cloud 10 stores large amounts of data based on high processing power and storage capacity, and creates one or more models suitable for various situations. In addition, the edge 20 includes an edge computing unit 21 to perform operations (processing) required for the edge 20.

In order to implement the collaboration technique according to the second embodiment, the cloud 10 generates an AI model (14) and distributes it to the edge 20 (15). When the edge 20 receives information from the object side, it separates it into a range where the privacy problem does not occur in the data separation unit 26, and performs immediate primary learning/analysis (25) and/or preprocessing (24) for some. Carry out.

Then, the immediate primary learning/analysis and/or preprocessed partial data is sent to the cloud 10. In the cloud 10, a main operation for this data, for example, inference, is performed using the generated model (17).

Here, immediate learning (25), preprocessing (24), and main processing (17) are operations performed by the edge computing unit 21 and the cloud computing unit 12 at the edge 20 and the cloud 10, respectively. Indicate that it shows an example of.

In the description of this embodiment, as mentioned above, the description with the name'~ unit' means not only the function of the hardware device, but also the function of the software, a routine, a processing unit, and the like.

In the third embodiment, a number of tasks to be processed in the cloud are partially processed by several edges, and may be referred to as'processing modularization capable of independent/organic processing at the edge'.

The cloud creates models responsible for several different functions. Edge can perform processing by using various models received from the cloud in combination according to its own state.

12 is a conceptual diagram of a cloud-edge distributed processing method according to a third embodiment.

There are several tasks to be processed in the cloud 10 (a, b, c). For example, suppose that when photo information comes in from the object side 30, the task is to identify a person from the photo (a), a task to determine gender (b), and a task to estimate age (c).

There may be models for each task. Since all of these tasks cannot be processed due to the capabilities of the edge 20, the necessary models are distributed so that only necessary or appropriate tasks are processed for each edge 20, and the tasks to be performed by each model are distributed.

For example, in FIG. 12, each edge 20 so that the first edge 20a processes only task a, the second edge 20b processes only tasks a and b, and the third edge 20a processes only tasks a and c. ) And the cloud 10 allow each edge to process a task distributed to each of them according to a task distribution rule (see FIG. 13).

13 shows an exemplary configuration diagram of a cloud 10-an edge 20 for implementing the concept of the third embodiment shown in FIG. 12.

Basically, there is a cloud computing unit 12 in the cloud 10 to perform the original cloud operation (processing), and the model generation unit 14 generates a model. That is, the cloud 10 stores large amounts of data based on high processing power and storage capacity, and creates one or more models suitable for various situations. In addition, the edge 20 includes an edge computing unit 21 to perform operations (processing) required for the edge 20.

In order to implement the third embodiment, by adding the task distribution rules 19 and 27 previously determined to the cloud 10 and the edge 20, the technique of the third embodiment can be carried out based on this rule. The edge 20 applies the task distribution rule 27 to select a task created in the cloud 10 (eventually, a task processing model to be used) and select a task from the cloud 10 according to its own situation. Is distributed (18).

More specifically, when the edge 20 selects a task, the task distribution unit 18 of the cloud 10 according to the task distribution rule 19 of the cloud that is linked to the task distribution rule 27 of the edge, the selected task ( And the corresponding model) is distributed to the corresponding edge 20.

The edge computing unit 21 of the edge 20 performs the task processing using the distributed model.

The difference between this third embodiment and the first embodiment is that data processing is not performed using multiple engines (models) for the same purpose in parallel, but multiple models that perform tasks for different purposes are serialized. It means that the task is shared by using it properly. Of course, the task's shared processing power and scope depend on the state of the edge (eg, the decision-making situation).

In the description of this embodiment, as mentioned above, what is described by the name'~ unit' means not only the functions of the hardware device but also the functions, routines, and processing units of software.

According to the above-described embodiment of the present invention, the following effects can be expected.

Edge-based distributed intelligence framework: Delegating/sharing the analysis considering the fixed/fluid characteristics of the cloud and various edges, cooperating organically to derive the optimal analysis result, and providing a framework that considers the application and application of the industry. It becomes possible.

Analysis optimization considering the resources and characteristics of the edge: The analysis model, performance, and analysis range are adaptively determined by considering the characteristics of the edge with a closed network environment and domain-dependent, fixed/fluid capabilities, and changed according to the situation. It is possible to utilize the analysis performance of

Increasing the usability of edge analysis: The usability of edge analysis is increased by modularizing the main functions so that the edge-based analysis engine can be easily utilized and deployed.

Organic collaboration between edge and cloud: In the past, the cloud was an infrastructure for big data processing, and the edge was a terminal that ran simple services, each independently performed tasks, and the analysis collaboration method between edge and cloud was 1) unilaterally in the cloud. Although it was in the early stages of technology development in progress by delegating/distributing inference engines to the edge, or 2) locally optimizing the edge learning model with relatively simple user data, according to the present invention, the edge and the cloud complement the other device. Or it will evolve into a collaborative medium, and through collaboration, the strengths of each device can be maximized and the weaknesses can be compensated, thereby gradually improving the analysis performance.

Edge-to-edge extension: The technology of the present invention can be extended through collaboration such as delegation, selection, and distribution of analysis between edge and edge as well as edge-cloud. For example, between the collaborating edges, an arbitrary edge may directly access the model repository by the surrounding edge and download a segmented AI model or a single AI model from the surrounding edge. In other words, an arbitrary edge can directly import a segmented AI model from the model repository, but if its own resources increase, the appropriate segmentation takes into account the AI model's name, version, whether it is segmented, and the segmented order information of the model. The created AI model can be searched from the surrounding edge and downloaded and used. In this case, the edges can share the AI model name, version, partitioning status, and partitioning order information of the model that they directly brought from the model repository.

An embodiment of the present invention may be implemented in the form of a computer program stored in a medium executed by a computer or a recording medium including instructions executable by a computer. Computer-readable media can be any available media that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. Further, the computer-readable medium may include both computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Communication media typically includes computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave, or other transmission mechanism, and includes any information delivery media.

Although the methods and systems of the present invention have been described in connection with specific embodiments, some or all of their components or operations may be implemented using a computer system having a general-purpose hardware architecture.

The above description of the present invention is for illustrative purposes only, and those of ordinary skill in the art to which the present invention pertains will be able to understand that other specific forms can be easily modified without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative and non-limiting in all respects. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as being distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the claims to be described later rather than the detailed description, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. do.

10: the cloud
20: edge
30: things
A: Model repository
100: distributed computing system
110: communication module
120: memory
130: processor

Claims (15)

In a method of providing an analysis model for adaptive intelligence utilization in a distributed computing environment,
Generating an analysis model in a cloud server;
Storing the generated analysis model in a model storage;
Distributing an application from the cloud server to an edge device; And
And searching and downloading one or more analysis models matching the deployed application by the edge device from the model storage.
The method of claim 1,
Storing the generated analysis model in a model storage comprises:
The method for providing an analysis model to include metadata information necessary for searching of the edge device in the analysis model and storing it in the model storage.
The method of claim 2,
The analysis model stored in the model repository includes basic information for the search, model division-related information, and model training information.
The method of claim 3,
The basic information for the search includes at least one of model usage cost information, model performance information, device performance information required for the model, and personal information protection capability information.
The method of claim 1,
The model storage includes at least one of information on whether to divide the analysis model, version information, and download file information,
The step of searching and downloading one or more analysis models matching the deployed application by the edge device from the model storage includes the edge device accessing and downloading the analysis model based on information stored in the model storage. How to provide an analysis model.
The method of claim 5,
The analysis model providing method, wherein the information on whether the analysis model is divided includes directory information in which the entire analysis model is not divided and directory information in which the analysis model divided into two or more is stored.
The method of claim 1,
The step of searching and downloading one or more analysis models matching the deployed application by the edge device from the model repository,
As the edge device searches for the matched analysis model, the searched model is downloaded using Rest API,
The Rest API is an analysis model providing method that includes name information, version information, information about whether to be divided, and information about the order of division of the searched analysis model.
The method of claim 1,
For distributing and cooperative processing of tasks between the cloud server and the edge device, or for distributing and cooperative processing of tasks between the edge devices, two or more edge devices process tasks individually or in collaboration based on the downloaded analysis model Analysis model providing method further comprising the step.
A cloud server that creates an analysis model for processing the collected data, stores the generated analysis model in a model storage, and distributes an application to an edge device,
A model repository for storing and managing the generated analysis model,
Distributed computing system comprising an edge device for searching and downloading one or more analysis models matching the deployed application from the model storage.
The method of claim 9,
The edge device includes two or more edge devices,
The distributed computing system wherein the edge devices respectively or cooperately process the task based on the downloaded analysis model.
The method of claim 9,
The cloud server stores metadata information necessary for searching of the edge device in the model to be included in the model and stored in the model storage,
The information of the metadata includes basic information for the search, information related to model division, and model training information,
The basic information for the search includes at least one of model usage cost information, model performance information, device performance information required for the model, and personal information protection capability information.
The method of claim 9,
The model storage includes at least one of information on whether to divide the analysis model, version information, and download file information,
Wherein the edge device accesses and downloads the analysis model based on information stored in the model storage.
The method of claim 12,
The information on whether the analysis model is divided includes directory information in which all undivided models are stored and directory information in which two or more divided analysis models are stored.
The method of claim 9,
As the edge device searches for the matched analysis model, it downloads the searched analysis model using Rest API,
The Rest API is a distributed computing system that includes name information, version information, information about whether to be divided, and information about the order of division of the searched analysis model.
In a distributed computing system comprising one or more edge devices and a cloud server,
The cloud server includes a model generation unit for generating an analysis model for processing data, and a distribution unit for distributing an application to the edge device,
The edge device is a distributed computing system including a model download unit for searching and downloading one or more analysis models matching the distributed application.

Images