KR20230067729A - Edge computing system for determining defect of goods based on image and center cloud server thereof - Google Patents

Abstract

An edge computing system for determining whether the defect of goods based on an image is disclosed. The system may include multiple edge servers and a central cloud server that manages them. Accordingly, the defect of goods can be accurately and quickly performed, and the labeling task of a neural network algorithm for determining the defect of goods can be easily performed.

Description

Edge computing system and its central cloud server for determining whether an item is defective based on image

The present invention relates to an edge computing system that quickly and accurately determines whether an item is defective based on an image, a central cloud server thereof, and an edge server.

Edge computing technology is a technology that can efficiently reduce latency and bandwidth problems by collecting, processing, and analyzing data at the end of industry and life by supporting real-time applications, and can expect fast response speed. It is evaluated as a technology that can solve situations such as the increase in the introduction of Internet of Things (IoT), traffic concentration in the central cloud, and the exponentially increasing amount of data and network traffic.

By installing and operating an artificial intelligence model in edge computing, you can reduce communication and bandwidth costs and respond with fast response speed when processing data in various areas such as security cameras, self-driving cars, drones, smart factories, and smart farms. , security can be maintained because data can be stored and processed at the end.

In particular, edge computing takes pictures of products in a smart factory environment, stores product images, processes the stored images to input them to an artificial intelligence model, and receives processed product images to create artificial intelligence models that distinguish good products from defective products. can be used

However, artificial intelligence model learning is required for image-based artificial intelligence models. When model learning is performed based on supervised learning, a lot of manpower and cost are inevitably required to label the correct answer to the image. Data labeling is to the extent that the modifier 'sticking AI eyes' is attached, so it requires a lot of time and effort.

Accordingly, there is a need for a method for accurately and quickly determining whether or not an item is defective in an edge computing environment.

On the other hand, the above information is presented only as background information to help the understanding of the present invention. No determination has been made, nor is any assertion made, as to whether any of the foregoing is applicable as prior art with respect to the present invention.

Republic of Korea Patent Publication No. 10-2021-0048058 (published date: 2021.5.3)

One object of the present invention to solve the above problems is to provide an edge system that performs necessary tasks of the system in a distributed manner by breaking away from the centralized method.

Another object of the present invention is to provide an edge computing system that determines whether a product is defective based on an image.

The problems to be solved by the present invention are not limited to the problems mentioned above, and other problems not mentioned will be clearly understood by those skilled in the art from the description below.

A central cloud server of an edge computing system for determining whether an item is defective based on an image according to an embodiment of the present invention for solving the above problems includes a communication unit communicating with a plurality of edge servers; a memory for storing a product defect prediction model predicting whether a product is defective when an product image is input and an active estimation model estimating prediction reliability of the product defect prediction model; and a controller that monitors the operation of the plurality of edge servers and determines whether to adopt the product defect prediction model for labeling based on the estimated prediction reliability.

The controller may allocate at least one work process to each of the plurality of edge servers.

The work process includes a first work process of photographing an article in the smart factory, storing and transmitting the photographed article image, a second work process of preprocessing the transmitted article image into an image to be input to the product defect prediction model, and the product defect. A third working process for executing an edge article defect prediction model corresponding to the predictive model and a fourth working process for executing an edge active estimation model corresponding to the active estimation model.

When at least one work process is assigned to each of the plurality of edge servers, the controller may be configured to provide information for performing the work process to the corresponding edge server.

The central cloud server may further include a display, and the controller may be configured to cluster product defect information of the product defect prediction model using the active estimation model and output the clustered product defect information to the display.

The controller may be configured to convert the data to 2D or 3D and output the data to the display when the dimension in which the data to be output to the display is expressed is 4-dimensional or higher.

The controller may adopt labeling information of the product defect prediction model using the active estimation model, when the prediction reliability of the active estimation model is equal to or greater than a predetermined reference value.

The controller may be configured not to adopt labeling information of the product defect prediction model when the prediction reliability of the active estimation model is less than a predetermined reference value using the active estimation model.

The controller may be configured to wait for input of labeling information for the product when the labeling information of the product defect prediction model is not adopted.

Here, some of the neural network parameters of the product defect prediction model may be transferred and applied to the active estimation model.

The controller may be configured to schedule the work process to be performed directly or by other edge servers when the work progress efficiency is below a predetermined level when a work process is performed in a predetermined edge server.

The controller may be configured to reflect edge product defect prediction models of at least some of the plurality of edge servers to the product defect prediction model and to reflect an edge active estimation model to the active estimation model.

An edge computing system for determining whether an item is defective according to an embodiment of the present invention includes a central cloud server; It can include multiple edge servers.

The central cloud server communicates with a plurality of edge servers and stores an item defect prediction model for predicting whether an item is defective when an item image is input and an active estimation model for estimating the prediction reliability of the item defect prediction model, and the plurality of edge servers. The operation of the edge server may be monitored, and based on the estimated prediction reliability, whether to adopt the product defect prediction model for labeling may be determined.

The plurality of edge servers may be configured to each perform work processes allocated by the central cloud server.

Here, each of the plurality of edge servers has a first work process of photographing an article in the smart factory, storing and transmitting the photographed article image, and a second work process of pre-processing the transmitted article image into an image to be input to the product defect prediction model. , at least one of a third task process of performing an edge product defect prediction model corresponding to the product defect prediction model, and a fourth task process of performing an edge active estimation model corresponding to the active estimation model.

When at least one work process is assigned to each of a plurality of edge servers, the central cloud server may be configured to provide information for performing the work process to the corresponding edge server.

The central cloud server may cluster and display product defect information of the product defect prediction model using the active estimation model.

The central cloud server may convert the data to 2D or 3D and display the data when the dimension in which the data to be displayed is expressed is 4D or higher.

The central cloud server adopts labeling information of the product defect prediction model when the prediction reliability of the active estimation model is equal to or greater than a predetermined reference value, and labels the product defect prediction model when the prediction reliability of the active estimation model is less than a predetermined reference value. It can be configured not to adopt information.

The central cloud server may be configured to wait for input of labeling information for an article when labeling information of the product defect prediction model is not adopted.

Here, some of the neural network parameters of the product defect prediction model may be transferred and applied to the active estimation model.

The central cloud server may be configured to schedule the work process to be performed directly or by other edge servers when the work progress efficiency is below a predetermined level when a work process is performed on a predetermined edge server.

The central cloud server may be configured to reflect edge product defect prediction models of at least some of the plurality of edge servers to the product defect prediction model and to reflect an edge active estimation model to the active estimation model.

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

By providing various embodiments of the present invention as described above, the amount of labeling work of an artificial intelligence model for determining whether a product is defective based on an image is reduced, thereby improving device efficiency, and cost reduction in system configuration and operation is expected. can

The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

1 is a diagram schematically illustrating an edge computing system for determining whether a product is defective based on an image according to an embodiment of the present invention;
2 is a diagram for explaining a configuration of a central cloud server and a work process performed by a plurality of edge servers according to an embodiment of the present invention;
3 is a diagram for explaining an operating method of an edge computing system for determining whether an item is defective based on an image according to an embodiment of the present invention;
4(a) and 4(b) show exemplary results of clustering data using an active estimation model according to an embodiment of the present invention, and
5 illustrates a case of receiving a label when the labeling reliability is low according to an embodiment of the present invention.

Advantages and features of the present invention, and methods for achieving them, will become clear with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various different forms, only the present embodiments make the disclosure of the present invention complete, and those skilled in the art to which the present invention belongs It is provided to fully inform the scope of the present invention, the present invention is defined only by the scope of the claims.

Terms used in this specification are for describing embodiments and are not intended to limit the present invention. In this specification, singular forms also include plural forms unless specifically stated otherwise in a phrase. As used herein, "comprises" and/or "comprising" does not exclude the presence or addition of one or more other elements other than the recited elements. Like reference numerals throughout the specification refer to like elements, and “and/or” includes each and every combination of one or more of the recited elements. Although "first", "second", etc. are used to describe various components, these components are not limited by these terms, of course. These terms are only used to distinguish one component from another. Accordingly, it goes without saying that the first element mentioned below may also be the second element within the technical spirit of the present invention.

Unless otherwise defined, all terms (including technical and scientific terms) used in this specification may be used with meanings commonly understood by those skilled in the art to which the present invention belongs. In addition, terms defined in commonly used dictionaries are not interpreted ideally or excessively unless explicitly specifically defined.

1 is a diagram schematically illustrating an edge computing system 1000 that determines whether or not an item is defective based on an image according to an embodiment of the present invention.

The edge computing system 1000 is a kind of distributed network system, can determine whether an item is defective by using a neural network algorithm, and can include a supervised learning-based learning model to which a labeling reduction technique is applied.

The central cloud server 100 of the edge computing system 1000 is a control system and may be implemented as a cloud system. As an optional embodiment, the central cloud server 100 may be implemented as a server, PC, laptop, tablet PC, or the like.

The central cloud server 100 can manage the operation of the plurality of edge servers 200, monitor operations, and manage the computational load of the plurality of edge servers 200A to 200Z (200). In addition, the central cloud server 100 may allocate a work process to the plurality of edge servers 200A to 200Z 200 and may schedule the work process.

Here, the central cloud server 100 may request execution of a task to some of the plurality of edge servers 200A to 200Z 200 even if the task can be performed directly. For example, the central cloud server 100 may request a neural network operation to a specific edge server (eg, 200A).

The plurality of edge servers 200A to 200Z 200 may be implemented as servers or workstations for edge computing. As an optional embodiment, the plurality of edge servers 200A to 200Z 200 may be implemented as a cloud, PC, laptop, tablet PC, or the like.

The plurality of edge servers 200A to 200Z 200 may have storage modules and processors themselves, and may be driven in cooperation with the central cloud server 100 for a common goal of the edge computing system 1000. The plurality of edge servers 200A to 200Z 200 may communicate with each other within a range without problems in network communication as needed, may operate in connection with each other, and may work jointly according to a service request.

2 is a diagram for explaining a configuration of the central cloud server 100 and a work process performed by a plurality of edge servers 200A to 200D according to an embodiment of the present invention.

First, a first edge server 200A among a plurality of edge servers 200A to 200D may perform a photographing work process P1. The photographing work process P1 may be a process of photographing an article image using one or more cameras during a conveyor belt or article process in a smart factory. However, the application place is not limited to the smart factory.

The second edge server 200B may perform the preprocessing work process P2. The preprocessing work process P2 may include processes such as resizing, cropping, normalization, gray scale, and enlargement/reduction to become an item image to be input to a neural network model to be described later, but the embodiment is not limited thereto.

The second edge server 200B may directly receive an article image from the first edge server 200A. As an optional embodiment, the second edge server 200B may receive an article image from the central cloud server 100 .

The third edge server 200C may perform a work process (P3) of predicting product defects by receiving product images, and at this time, necessary information (model configuration information, labeling information, loss function) from the central cloud server 100 information), the model can be trained, and the model can be applied. A model learned by the third edge server 200C may be provided to the central cloud server 100 .

The fourth edge server 200D may perform the active estimation task process P4. The active estimation work process P4 may estimate the prediction reliability of the product defect prediction model, and may cluster data requiring labeling and provide the data to the user. A model learned by the fourth edge server 200D may be provided to the central cloud server 100 .

As an optional or additional embodiment, the central cloud server 100 may itself perform the work performed by the above-described edge servers 200A to 200D.

The central cloud server 100 may include an input unit 110 , a communication unit 120 , a display 130 , a memory 140 and a controller 190 . According to embodiments, the central cloud server 100 may include only some of the above-described components or may further include other components.

The input unit 110 includes various input means and may receive labeling information of a neural network model from a user or a robot.

The communication unit 120 is a means for communicating with a plurality of edge servers 200A to 200D, and may include a wired/wireless communication module, a mobile communication module, and a short-distance communication module, but the embodiment is not limited thereto.

The display 130 is a component for visualizing various information, and may be implemented as an LCD, LED, OLED, etc., but the embodiment is not limited thereto. The display 130 can visually output various information according to the control of the controller 190 .

The memory 140 is a module that collectively refers to various types of storage, and the memory 140 may store information required to perform calculations using artificial intelligence, machine learning, and artificial neural networks. The memory 140 may store various learning models. The learning models may be used to infer result values for new input data other than learning data, and the inferred values may be used as a basis for decision to perform a certain operation. can be used

The learning models may be trained based on label information, and a backpropagation algorithm, a cross entropy algorithm, or the like may be applied to increase learning accuracy. In the present specification, the memory 140 receives a product image and predicts whether or not the product is defective, and labels the product defect prediction model (PM) based on the predictive reliability of the product defect prediction model (PM). It may include an active estimation model (EM) necessary to determine whether to adopt for.

Here, the active estimation model (EM) may receive labeling but unlearned data, compare the predicted value of the product defect prediction model (PM) with the correct answer value, and build a correct answer data set for learning.

In addition, the active estimation model (EM) uses a Softmax function or a function capable of determining reliability (eg, a function using a mean map, a variance map, etc.) to predict the probability of predicting a correct label and the predictive value of an incorrect label. The probability can be calculated, and if the probability of predicting with the correct label is high, the item defect prediction model (PM) performs the labeling task with the predicted value, and if the probability of predicting with the incorrect label is high, the user or AI-based robot It can be implemented to directly label.

Here, the softmax function or the function capable of determining reliability may be disposed at the end of the fully connected network, but the embodiment is not limited thereto.

The controller 190 monitors the operation of the plurality of edge servers 200A to 200D, and determines whether or not to adopt the product defect prediction model (PM) for labeling based on the estimated prediction reliability of the active estimation model (EM). can

As described above, the controller 190 may allocate various work processes to the plurality of edge servers 200A to 200D, monitor the work process, and when it is determined that it is difficult to perform the work process (eg, network traffic If this value exceeds the predetermined value or the operation rate of the processor of the edge server exceeds the predetermined value), the corresponding work process may be requested to be performed by another edge server or the corresponding work process may be performed by itself.

The controller 190 may include one or more processors, and even when expressed in the singular number, it may be regarded as plural. The controller 190 is a module that controls components of the central cloud server 100, and the controller 190 is embedded in hardware, having a physically structured circuit to perform functions expressed by codes or commands included in a program. may mean a data processing device. As an example of such a data processing device built into hardware, a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated (ASIC) circuit), field programmable gate array (FPGA), etc., but the scope of the present invention is not limited thereto. The controller 190 may include a separate running processor for performing artificial intelligence calculations or may include a running processor itself.

Specifically, the controller 190 may allocate at least one work process to each of the plurality of edge servers 200A to 200D. The work process includes a first work process of photographing an item in the smart factory, storing and transmitting the photographed item image, a second work process of preprocessing the transmitted item image into an image to be input to the product defect prediction model (PM), and the product It may include a third work process for executing the edge product defect prediction model corresponding to the defect prediction model (PM) and a fourth work process for executing the edge active estimation model corresponding to the active estimation model (EM). is not limited to the above-described work process.

When assigning at least one work process to each of the plurality of edge servers 200A to 200D, the controller 190 may provide information for performing the work process to the corresponding edge server through the communication unit 120 .

The controller 190 may cluster product defect information of the product defect prediction model (PM) using the active estimation model (EM) and output the clustered product to the display 130 .

In this case, if the dimension in which the data to be output to the display 130 is expressed is 4-dimensional or higher, the controller 190 may convert the data into 2-dimensional or 3-dimensional data and output the data to the display 130 . If it is 4-dimensional or more, it is difficult to visually express it, so it can be changed to 2-dimensional or 3-dimensional. Here, the dimension may vary according to various characteristic parameters, and high-dimensional data may be converted into 2-dimensional or 3-dimensional data and displayed on the display 130 .

The controller 190 may adopt the labeling information of the product defect prediction model (PM) when the prediction reliability of the active estimation model (EM) is greater than or equal to a predetermined reference value by using the active estimation model (EM).

The controller 190 may not adopt the labeling information of the product defect prediction model (PM) when the prediction reliability of the active estimation model (EM) is less than a predetermined reference value.

When the labeling information of the product defect prediction model is not adopted, the controller 190 may wait for input of labeling information for the product. In this case, labeling information may be input by a user, or may be input using a pre-learned model for automatic labeling.

Here, some of the neural network parameters of the product defect prediction model (PM) may be transferred and applied to the active estimation model (EM). The transition target neural network parameters may correspond to networks other than the distal part of the model. For example, when a Convolution Neural Network (CNN) is applied, it may be a neural network parameter used for convolution operation, max pooling operation, etc., but the embodiment is not limited thereto.

The controller 190 may schedule the work process to be performed directly or by another edge server when the work progress efficiency is below a certain level when a work process is performed on a predetermined edge server.

The controller 190 may be configured to reflect the edge product defect prediction model of at least some of the plurality of edge servers 200A to 200D to the product defect prediction model (PM) and to reflect the edge active estimation model to the active estimation model (EM). can In addition, the controller 190 may reflect a product defect prediction model (PM) or an active estimation model (EM) to the edge server.

3 is a diagram for explaining an operating method of an edge computing system for determining whether an item is defective based on an image according to an embodiment of the present invention.

First, the product defect prediction model (PM) receives the product image (GI) and predicts whether the corresponding product is good or defective (310).

The active estimation model (EM) may receive at least one product image and labeling information (LI) including defect prediction information as a set, and output the estimated prediction reliability 320 of the product defect prediction model (PM). .

Here, the estimated prediction reliability 320 is a numerical value representing the prediction accuracy of the product defect prediction model (PM), and it can be determined whether the product defect prediction model (PM) is a reliable model.

The active estimation model (EM) may be used by freezing most of the operation parameters of the product defect prediction model (PM) (330), and may be configured by changing only the terminal network, but the embodiment is not limited thereto.

The active estimation model (EM) may provide parameters and related information of the algorithm when retraining the product defect prediction model (PM).

The controller 190 may be configured to use 340 as labeling information if the estimated reliability 320 of the product defect prediction model is reliable, and to wait for labeling information input if not reliable.

4(a) and 4(b) show clustering of data using an active estimation model according to an embodiment of the present invention.

When the performance of the active estimation model (EM) is low (Fig. 4(a)), the distinction between good/defective products is not clear, and when the performance is high (Fig. 4(b)), good/defective goods are classified. this may be clear.

When the performance of the active estimation model (EM) is low, the product defect prediction model (PM) may be re-trained, and during re-learning, the operation information and parameter values of the active estimation model (EM) are changed to the product defect prediction model (PM). ) can be provided.

5 illustrates a case of receiving a label when labeling reliability is low according to an embodiment of the present invention.

Referring to FIG. 5 , the central cloud server 100 may receive user labeling information through the input unit 110 or the communication unit 120 (510) and use it as labeling information.

In addition, the central cloud server 100 may create a label by setting at least one product image and correct answer information configured as a pair using a pre-learned model for generating labeling information.

When a user directly inputs labeling, the central cloud server 100 may store the corresponding labeling in a file format such as JSON or CSV.

The operating method according to an embodiment of the present invention described above may be implemented as a program (or application) to be executed in combination with a computer, which is hardware, and stored in a medium. Here, the computer may be the central cloud server 100 described above.

The above-described program is C, C++, Python, JAVA, C, C++, Python, JAVA, which can be read by a processor (CPU) of the computer through a device interface of the computer so that the computer reads the program and executes the methods implemented as a program. It may include a code coded in a computer language such as machine language. These codes may include functional codes related to functions defining necessary functions for executing the methods, and include control codes related to execution procedures necessary for the processor of the computer to execute the functions according to a predetermined procedure. can do. In addition, these codes may further include memory reference related codes for which location (address address) of the computer's internal or external memory should be referenced for additional information or media required for the computer's processor to execute the functions. there is. In addition, when the processor of the computer needs to communicate with any other remote computer or server in order to execute the functions, the code uses the computer's communication module to determine how to communicate with any other remote computer or server. It may further include communication-related codes for whether to communicate, what kind of information or media to transmit/receive during communication, and the like.

Steps of a method or algorithm described in relation to an embodiment of the present invention may be directly implemented as hardware, implemented as a software module executed by hardware, or implemented by a combination thereof. A software module may include random access memory (RAM), read only memory (ROM), erasable programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, hard disk, removable disk, CD-ROM, or It may reside in any form of computer readable recording medium well known in the art to which the present invention pertains.

Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art to which the present invention pertains can be implemented in other specific forms without changing the technical spirit or essential features of the present invention. you will be able to understand Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive.

1000: edge system,
100: central cloud server, 200: edge server

Claims (20)

As a central cloud server of an edge computing system that determines whether an item is defective based on an image,
A communication unit communicating with a plurality of edge servers;
a memory for storing a product defect prediction model predicting whether a product is defective when an product image is input and an active estimation model estimating prediction reliability of the product defect prediction model; and
And a controller for monitoring the operation of the plurality of edge servers and determining whether to adopt the labeling of the product defect prediction model based on the estimated prediction reliability. According to claim 1,
The controller assigns at least one work process to each of the plurality of edge servers;
The work process is
A first work process of photographing an item in the smart factory, storing and transmitting the photographed item image, a second work process of preprocessing the transmitted item image into an image to be input to the product defect prediction model, and corresponding to the product defect prediction model A central cloud server comprising: a third work process for executing an edge product defect prediction model for performing an edge product defect prediction model and a fourth work process for executing an edge active estimation model corresponding to the active estimation model. According to claim 2,
The controller,
When at least one work process is allocated to each of the plurality of edge servers, the central cloud server configured to provide information for performing the work process to the corresponding edge server. According to claim 3,
further comprising a display;
The controller uses the active estimation model,
A central cloud server configured to cluster product defect information of the product defect prediction model and output the clustered product to the display. According to claim 4,
The controller,
The central cloud server configured to convert the data to 2-dimensional or 3-dimensional data and output the data to the display when the dimension in which the data to be output to the display is expressed is 4-dimensional or higher. According to claim 5,
The controller uses the active estimation model,
When the prediction reliability of the active estimation model is greater than or equal to a predetermined reference value, labeling information of the product defect prediction model is adopted;
Wherein the central cloud server is configured not to adopt the labeling information of the product defect prediction model when the prediction reliability of the active estimation model is less than a predetermined reference value. According to claim 6,
The controller,
The central cloud server configured to wait for input of labeling information for an article when labeling information of the product defect prediction model is not adopted. According to claim 7,
A central cloud server in which some of the neural network parameters of the product defect prediction model are transferred and applied to the active estimation model. According to claim 8,
The controller,
A central cloud server configured to schedule the work process to be performed directly or by another edge server when the work progress efficiency is below a predetermined level when the work process is performed on a predetermined edge server. According to claim 9,
The controller,
A central cloud server configured to reflect edge product defect prediction models of at least some of the plurality of edge servers to the product defect prediction model and to reflect an edge active estimation model to the active estimation model. As an edge computing system for determining whether an item is defective,
central cloud server;
Includes multiple edge servers,
The central cloud server,
Communicate with a plurality of edge servers and store an item defect prediction model that predicts whether an item is defective when an item image is input and an active estimation model that estimates the prediction reliability of the product defect prediction model,
Monitoring the operation of the plurality of edge servers, and based on the estimated prediction reliability, determining whether to adopt the labeling of the product defect prediction model;
The plurality of edge servers,
An edge computing system configured to perform each of the work processes assigned by the central cloud server. According to claim 11,
Each of the plurality of edge servers,
A first work process of photographing an item in the smart factory, storing and transmitting the photographed item image, a second work process of preprocessing the transmitted item image into an image to be input to the product defect prediction model, and corresponding to the product defect prediction model An edge computing system that performs at least one of a third work process for performing an edge article defect prediction model for performing an edge product defect prediction model, and a fourth work process for performing an edge active estimation model corresponding to the active estimation model. According to claim 12,
The central cloud server,
When at least one work process is assigned to each of a plurality of edge servers, the edge computing system configured to provide information for performing the work process to the corresponding edge server. According to claim 13,
The central cloud server,
An edge computing system for clustering and displaying product defect information of the product defect prediction model using the active estimation model. According to claim 14,
The central cloud server,
An edge computing system that converts the data to 2D or 3D and displays the data when the dimension in which the display target data is expressed is 4D or higher. According to claim 15,
The central cloud server,
When the prediction reliability of the active estimation model is greater than or equal to a predetermined reference value, labeling information of the product defect prediction model is adopted, and when the prediction reliability of the active estimation model is less than a predetermined reference value, labeling information of the product defect prediction model is not adopted. Configured, edge computing system. According to claim 16,
The central cloud server,
Edge computing system configured to wait for input of labeling information for an article when labeling information of the article defect prediction model is not adopted. According to claim 17,
Edge computing system wherein some of the neural network parameters of the product defect prediction model are transferred and applied to the active estimation model. According to claim 18,
The central cloud server,
An edge computing system configured to schedule the work process to be performed directly or by another edge server when the work progress efficiency is below a predetermined level when the work process is performed on a predetermined edge server. According to claim 19,
The central cloud server,
The edge computing system configured to reflect the edge product defect prediction model of at least some of the plurality of edge servers to the product defect prediction model and to reflect the edge active estimation model to the active estimation model.

Images