KR102583170B1 - A method for learing model recommendation through performance simulation and a learning model recommencation device - Google Patents

Abstract

The learning model recommendation device according to an embodiment of the present invention is a learning model recommendation device that recommends a learning model calculated by machine learning or deep learning according to the request of the consumer, and includes the request of the consumer and the learning model requested by the consumer. A receiving module that receives the calculation conditions of the target learning model; A storage module that stores a storage learning model, which is a learning model created in the past; a similarity judgment module that determines the similarity between the target learning model and the stored learning model through a predetermined similarity judgment method using the calculation conditions of the target learning model; a recommendation module that calculates a recommended learning model for a learning model that satisfies predetermined recommendation conditions based on similarity among the stored learning models stored in the storage module; and an interface module that calculates an interface that allows the consumer to check and download the recommended learning model.

Description

Learning model recommendation method through performance simulation and learning model recommendation device including the same {A METHOD FOR LEARING MODEL RECOMMENDATION THROUGH PERFORMANCE SIMULATION AND A LEARNING MODEL RECOMMENCATION DEVICE}

The present invention relates to a technology for recommending a learning model, and more specifically, to a learning model recommendation device based on similarity that recommends a learning model suited to user needs and a cloud integrated operating system including the same.

Recently, as the cloud market has gradually grown, many services are being built based on the cloud, and the size of the cloud is also growing explosively. Because it is realistically difficult for an administrator to manage hundreds or thousands of virtual machines one by one, technology is being developed to automatically operate and manage them systematically. Here, the key to automatic operation is to implement the functions necessary for cloud operation using a learning model through machine learning/deep learning.

Unlike the existing method of setting up all analysis algorithms, machine learning overcomes the limitations of existing computer algorithms in that it creates a model that suits the purpose on its own based on given data. Because of these excellent functions and convenience, it is being developed and commercialized in various industrial fields.

However, machine learning or deep learning requires a lot of time and money to collect and label a lot of data, and learning models that require a lot of resources are treated as each development company's own assets and are not made public but are stored and used on their own. It is becoming. As a result, learning models with the same purpose are produced and used in many places, causing the problem of social resources being wasted.

In addition, as the number of learning models and the amount of analysis used to efficiently operate the cloud increases, the operating load increases, resulting in the problem of not being able to manage the cloud efficiently.

The present invention is intended to solve the above-mentioned problems, and includes a learning model recommendation device through performance simulation that can recommend the most similar learning model according to user needs and activate the distribution of the learning model, and a cloud integrated operating system including the same. It is about.

The learning model recommendation device according to an embodiment of the present invention is a learning model recommendation device that recommends a learning model calculated by machine learning or deep learning according to the request of the consumer, and includes the request of the consumer and the learning model requested by the consumer. A receiving module that receives the calculation conditions of the target learning model; A storage module that stores a storage learning model, which is a learning model created in the past; a similarity judgment module that determines the similarity between the target learning model and the stored learning model through a predetermined similarity judgment method using the calculation conditions of the target learning model; a recommendation module that calculates a recommended learning model for a learning model that satisfies predetermined recommendation conditions based on similarity among the stored learning models stored in the storage module; and an interface module that calculates an interface that allows the consumer to check and download the recommended learning model.

In addition, the predetermined recommendation condition is a condition that is similar to the target learning model by a first similarity or higher, and the recommendation module performs a predetermined calculation when a predetermined additional condition is satisfied even if the predetermined recommendation condition is not satisfied. The recommended learning model is calculated according to a method, and the predetermined additional condition may be a condition for judging based on a second similarity that is lower than the first similarity.

In addition, the predetermined additional condition is the stored learning model that is similar to the target learning model by more than the second similarity and less than the first similarity, and exists by more than the second similarity based on the target learning model. This may be a condition in which clusters exist below a predetermined number.

In addition, it further includes a simulation module that performs a simulation to generate a virtual learning model generated by connecting the stored learning models, wherein the recommendation module satisfies the predetermined recommendation conditions including the virtual learning model. You can judge whether it works or not.

Additionally, the simulation module can generate a simulation model that calculates the state of the final learning model when multiple learning models are input through machine learning.

In addition, the simulation module selects the storage learning model that must receive input data identical to the input data to be input to the target learning model, and outputs output data identical to the output data output from the target learning model. By selecting a learning model, the virtual learning model can be created.

In addition, in order to check whether the recommendation learning model operates normally, a simulation module that simulates the performance of the recommendation learning model may be further included.

Additionally, the predetermined similarity judgment method may be a method of calculating the degree of similarity by comparing the calculation conditions of the target learning model and the calculation conditions of the stored learning model.

Additionally, the calculation condition may consist of at least one of the following information: the learning purpose of the learning model, the type of output, the learning method, the type of learning model, the amount of learning data, the type of learning data, and the learning time.

The learning model recommendation method according to an embodiment of the present invention is implemented by a learning model recommendation device and recommends a learning model calculated by machine learning or deep learning according to the request of the consumer, comprising: A request receiving step in which the request of the consumer and the calculation conditions of the target learning model, which is the learning model requested by the consumer, are received by the module; A condition storage step in which a stored learning model, which is a learning model created in the past, and calculation conditions of the stored learning model are stored by a storage module; A similarity judgment step in which the similarity between the target learning model and the stored learning model is determined by a similarity judgment module based on a predetermined similarity judgment method using calculation conditions of the target learning model; A recommendation step in which a learning model that satisfies predetermined recommendation conditions is recommended to a consumer among the stored learning models stored in the storage module by a recommendation module; and an interface calculation step in which an interface through which the consumer can check and download the recommended learning model is calculated by the interface module.

A monitoring device according to an embodiment of the present invention is a monitoring device that implements a monitoring method for monitoring the operation of at least one virtual machine running on a physical server within the allocated resources of a cloud integrated operating system, wherein the virtual machine A transmission/reception module that collects operation information generated from the machine; a monitoring module that monitors whether the virtual machine operates abnormally based on the operation information collected by the transmission and reception module; A model operation module that operates a monitoring learning model that assists the monitoring function of the monitoring module, based on the request of the monitoring module; a storage module storing information necessary to implement the monitoring method; and a schedule module that determines the operation time of the surveillance learning model through a predetermined timing determination method so that the surveillance learning model can be operated within the allocated resources.

In addition, the monitoring learning model includes a first predictive monitoring learning model that predicts the load generated in the cloud integrated operation system for a predetermined future period and an error monitoring learning model that monitors whether an error occurs in the cloud service, The schedule module may determine the operation timing of the error monitoring learning model based on the load predicted by the first predictive monitoring learning model.

In addition, the monitoring learning model includes a second predictive monitoring learning model that predicts the workload generated in a virtual machine on a physical server during a predetermined future period and an error monitoring learning model that monitors whether an error occurs in a cloud service, The schedule module may determine the operation timing of the error monitoring learning model based on the load predicted by the second predictive monitoring learning model.

In addition, the monitoring learning model includes a first predictive monitoring learning model that predicts a load generated in the cloud integrated operating system for a predetermined future period, and the storage module is predicted by the first predictive monitoring learning model. A load is stored, and the schedule module can determine the operation time based on the load predicted in the first prediction monitoring model predicted in the past and stored in the storage module.

In addition, the monitoring learning model includes a second predictive monitoring learning model that predicts a workload generated in a virtual machine on a physical server during a predetermined future period, and the storage module is predicted by the second predictive monitoring learning model. One workload is stored, and the schedule module can determine the operation time based on the workload predicted by the second prediction monitoring model predicted in the past and stored in the storage module.

Additionally, the predetermined timing determination method may be a method of determining the operation timing when a load or workload is generated below a predetermined ratio based on allocated resources.

Additionally, the predetermined timing determination method may be a method of determining the operation timing at a time when the least load or workload occurs during the predetermined future period.

A cloud integrated operating system according to an embodiment of the present invention includes a management device for managing virtual machines running on a physical server; A monitoring device that monitors the operation of the virtual machine; And a learning model recommendation device that recommends a monitoring learning model necessary for the monitoring device to implement the monitoring function, wherein the monitoring device includes a transmission/reception module that collects operation information generated from the virtual machine, and the transmission/reception module collects the operation information. A monitoring module that monitors whether the virtual machine operates abnormally based on the operation information, a model operation module that operates a monitoring learning model that assists the monitoring function of the monitoring module based on a request from the monitoring module, A storage module that stores the information necessary to implement the monitoring method, a schedule module that determines the operation time of the surveillance learning model through a predetermined timing method so that the surveillance learning model can be operated within the allocated resources, and a predetermined schedule. and a model selection module that selects the monitoring and learning model to be operated in the model operation module that satisfies the operation conditions, and the model selection module makes a recommendation for the surveillance and learning model that satisfies the predetermined operation conditions. You can request a model recommendation from the device.

A monitoring method according to an embodiment of the present invention includes, through a monitoring device, monitoring the operation of at least one virtual machine running on a physical server within the allocated resources of a cloud integrated operation system, collecting operation information generated from the virtual machine; Monitoring, by a monitoring module, whether the virtual machine operates abnormally based on the operation information; A step of operating, by a model operation module, a monitoring learning model that assists the monitoring function of the monitoring module, based on a request from the monitoring module; Storing information necessary to implement the monitoring method by a storage module; And it may include a step of determining, by a schedule module, the operation time of the surveillance learning model through a predetermined timing determination method so that the surveillance learning model can be operated within the allocated resources.

A learning model recommendation device based on similarity according to the present invention and a cloud integrated operation system including the same can reduce social resource costs.

Additionally, the system can be operated efficiently.

Additionally, development time can be dramatically shortened.

Additionally, overloading of the system can be prevented.

However, the effects of the present invention are not limited to the effects described above, and effects not mentioned can be clearly understood by those skilled in the art from this specification and the attached drawings.

1 is a relationship diagram of a cloud integrated operation system according to an embodiment of the present invention.
Figure 2 is a configuration diagram of a cloud integrated operation system according to an embodiment of the present invention.
Figure 3 is a configuration relationship diagram of a monitoring device of a cloud integrated operation system according to an embodiment of the present invention.
Figure 4 is a flow chart of the learning model recommendation method implemented by the learning model recommendation device of the cloud integrated operation system according to an embodiment of the present invention.
Figure 5 is a diagram illustrating predetermined recommendation conditions and predetermined additional conditions among the learning model recommendation method according to an embodiment of the present invention.
Figure 6 is a diagram illustrating a learning model mediation interface screen provided by the interface module of the learning model recommendation device according to an embodiment of the present invention.
Figure 7 is a flowchart illustrating a process in which a learning model recommendation device selects a learning model to recommend to a monitoring device according to an embodiment of the present invention.
Figure 8 is a diagram illustrating how the model selection module of the monitoring device calculates the first operation time and the second operation time according to an embodiment of the present invention.

Hereinafter, specific embodiments of the present invention will be described in detail with reference to the drawings. However, the spirit of the present invention is not limited to the presented embodiments, and those skilled in the art who understand the spirit of the present invention may add, change, or delete other components within the scope of the same spirit, or create other degenerative inventions or this invention. Other embodiments that are included within the scope of the invention can be easily proposed, but this will also be said to be included within the scope of the invention of the present application.

Figure 1 is a relationship diagram of a cloud integrated operation system according to an embodiment of the present invention.

Referring to FIG. 1, the cloud integrated operation system 10 according to an embodiment of the present invention manages and operates virtual machines running on a physical server, and sells, rents, and provides cloud services for learning models needed by cloud users (consumers). It may be a system that provides.

To this end, the cloud integrated operation system 10 is connected to the consumer (M10) and the physical server (S10) through a wired/wireless network and can transmit and receive necessary information.

For example, the consumer (M10) may be a person who receives a virtualized cloud service.

For example, the consumer (M100) can send the conditions and request for the learning model he or she wants to purchase or rent to the learning model recommendation device, and the learning model recommendation device is an interface that can request and transmit and receive a learning model or learning model from the consumer. etc. can be provided.

When there is a request from a consumer (client), the cloud integrated operation system 10 may be a system that recommends the necessary cloud service or learning model according to the request.

Additionally, the learning model recommendation method can provide an interface so that the learning model can be used in a cloud environment.

The network in the present invention may be a core network integrated with a wired public network, wireless mobile communication network, or mobile Internet, etc., and may include TCP/IP protocols and various services existing at the upper layer thereof, such as HTTP (Hyper Text Transfer Protocol) and HTTPS. (Hyper Text Transfer Protocol Secure), Telnet, FTP (File Transfer Protocol), DNS (Domain Name System), SMTP (Simple Mail Transfer Protocol), etc., which may refer to a worldwide open computer network structure, examples of which include It comprehensively refers to a data communication network that can transmit and receive data in various forms without limitation.

Below, the cloud integrated operation system will be described in detail.

Figure 2 is a configuration diagram of a cloud integrated operation system according to an embodiment of the present invention, and Figure 3 is a configuration diagram of a monitoring device of a cloud integrated operation system according to an embodiment of the present invention.

Referring to Figures 2 and 3, the cloud integrated operation system according to an embodiment of the present invention includes a learning model recommendation device 100 that recommends a learning model according to the request of the consumer, and a learning model recommendation device that manages virtual machines running on a physical server. It may include a management device and a monitoring device that monitors the operation of the virtual machine.

Below, each device will be described in detail.

The learning model recommendation device 100 according to an embodiment of the present invention is a learning model recommendation device 100 that recommends a learning model calculated by machine learning or deep learning according to the request of the consumer, and A receiving module (110) that receives the calculation conditions of the target learning model, which is a learning model requested by the consumer, a stored learning model that is a learning model created in the past, and a first storage module (120) that stores the calculation conditions of the stored learning model. ), a similarity judgment module 130 that determines the similarity between the target learning model and the stored learning model and a learning model that satisfies predetermined recommendation conditions among the stored learning models stored in the first storage module 120 are recommended for learning. It may include a recommendation module 140 that recommends the model to the consumer.

In addition, the learning model recommendation device 100 may further include a simulation module 150 that simulates the state of the final learning model when connecting arbitrary learning models.

In addition, the learning model recommendation device 100 may further include a learning suggestion module 160 that suggests creating a learning model to the administrator when a predetermined learning condition is satisfied.

In addition, the learning model recommendation device 100 may further include a price calculation module 170 that calculates the cost to be charged while recommending the stored learning model.

In addition, the learning model recommendation device 100 may further include an interface module 180 that calculates and transmits an interface that can display and collect information necessary to implement the learning model recommendation method on the computing device of the manager or consumer. there is.

The receiving module 110 can collect information necessary to implement a learning model recommendation method.

For example, the consumer may receive the request information and the calculation conditions for the learning model that the consumer needs.

Consumers can input learning model requests and calculation conditions through the interface provided by the interface module 180 through their computing devices.

Calculation conditions may include the learning purpose of the learning model, type of output, learning method, type of learning model, amount of learning data, type of learning data, learning time, etc.

Learning purpose may refer to the purpose or task that the learning model seeks to solve.

For example, the learning purpose may be to identify a person based on image data or to identify lung cancer based on X-RAY data.

However, it is not limited to this, and the types of learning objectives can be modified in various ways at a level that is obvious to those skilled in the art.

The type of output may refer to the specific meaning of the result.

For example, it can indicate ‘whether it is a person or not a person’, ‘whether it is lung cancer or not’, ‘a certain treatment method’, etc.

However, it is not limited to this, and the type of output can be modified in various ways at a level that is obvious to those skilled in the art.

The learning method may refer to a method of machine learning and/or deep learning a learning model.

For example, the learning method can be classified as deep learning or machine learning.

Additionally, learning methods can be divided into supervised learning, unsupervised learning, and reinforcement learning.

The learning model type may refer to the specific algorithm type used for learning.

For example, types of learning models can be classified into k-nearest neighbor algorithm (K-NN), support vector machine algorithm (SVM), decision tree algorithm, random forest algorithm, etc.

For example, learning model types can be divided into ANN, CNN, DNN, GAN, and RNN.

However, it is not limited to this, and the types of learning models can be modified in various ways at a level that is obvious to those skilled in the art.

The amount of learning data may refer to the amount of data required to create a learning model.

For example, the amount of learning data may be expressed in numbers and units expressing the amount of data, such as GB or TB.

The type of learning data may be classified into image form, text form, or video form.

Learning time may refer to the time required for machine learning/deep learning.

However, it is not limited to this, and the above calculation conditions can be modified in various ways at a level that is obvious to those skilled in the art.

The receiving module 110 may transmit the received information to other modules, including the similarity determination module 130.

The receiving module 110 can receive information necessary to operate the learning model recommendation method and the learning model mediation platform.

The first storage module 120 may store all information necessary to implement the learning model recommendation method.

In the first storage module 120, the stored learning model, which is a learning model learned in the past, and the calculation conditions of the stored learning model may be matched and stored.

The first storage module may include internal memory and/or external memory.

For example, internal memory may include volatile memory (e.g., DRAM, SRAM, or SDRAM), non-volatile memory (e.g., one time programmable ROM (OTPROM), PROM, EPROM, EEPROM, mask ROM, flash ROM, flash memory, hard memory, etc. may include at least one of a drive, or solid state drive (SSD)

External memory may be a flash drive, such as compact flash (CF), secure digital (SD), Micro-SD, Mini-SD, extreme digital (xD), multi-media card (MMC), or memory stick. It may include etc.

The similarity judgment module 130 may be a model that determines the degree of similarity between the target learning model and the stored learning model.

The similarity judgment module 130 can determine the degree of similarity between the target learning model and the stored learning model through a predetermined similarity judgment method.

The predetermined similarity judgment method may be a method of calculating the degree of similarity by comparing the calculation conditions of the target learning model and the calculation conditions of the stored learning model, respectively.

The similarity judgment module 130 can preprocess and vectorize the calculation conditions of the storage learning model. The vectorization method utilizes known techniques, so detailed descriptions thereof may be omitted.

Likewise, the similarity judgment module 130 can preprocess and vectorize the calculation conditions of the target learning model.

The predetermined similarity judgment method may be a method of calculating the degree of similarity by comparing the calculation conditions of the target learning model with the calculation conditions of the target learning model.

For example, the similarity judgment module 130 may calculate the degree of similarity based on the distance between the calculation condition vector information of the target learning model and the calculation condition vector information of the stored learning model.

However, it is not limited to this, and the similarity determination method of the similarity determination module 130 can be modified in various ways at a level that is obvious to those skilled in the art.

The recommendation module 140 may select a recommended learning model by determining whether a predetermined recommendation condition, a predetermined additional condition, and a predetermined merge condition are satisfied.

The predetermined recommendation condition may be a condition that is similar to the target learning model by a first degree of similarity or higher.

When the distance to the target learning model is less than or equal to the first distance, it may be judged to be similar to or greater than the first similarity.

A predetermined additional condition is that the cluster of stored learning models that are similar to the target learning model by more than the second similarity and less than the first similarity, and exist by the second similarity or more based on the target learning model, is predetermined. It may be a condition that exists in numbers or less.

When the distance from the target learning model is greater than the first distance but less than or equal to the second distance, it may be determined to be similar to the second similarity or greater but less than the first similarity.

Here, the second distance may mean a longer distance than the first distance.

Here, the predetermined number of clusters may be three.

However, it is not limited to this, and the predetermined number of clusters can be modified in various ways at a level that is obvious to those skilled in the art.

The predetermined merging condition may be a condition in which the similarity with the target learning model is greater than or equal to the first similarity when the stored learning models that are similar to the target learning model and less than the second similarity are connected.

When the distance from the target learning model exceeds the second distance, it may be judged to be similar to less than the second similarity degree.

The stored learning model itself is similar to the target learning model at the second similarity level, but when multiple stored learning models are combined, there may be cases where the first similarity level is higher than the target learning model.

For example, it can be assumed that the target learning model is a model that detects whether an intruder is an intruder based on video data and calculates a response plan according to the type of intrusion. The first stored learning model may be a learning model that determines whether an intruder is an intruder based on video data, the second stored learning model may be a learning model that determines the type of intrusion based on video data, and the third stored learning model The model may be a model that calculates a response plan based on the type and form of the intrusion. Each of the first to third stored learning models is similar to the target learning model by less than the second similarity, but when combining the first to third stored learning models, it is similar to the target learning model by more than the first similarity. can do.

Through this, it is possible to recommend a learning model that suits the needs of the user by making full use of the storage learning module stored in the storage module 120.

The simulation module 150 can create a virtual learning model by connecting the storage learning modules stored in the storage module 120 with each other.

If the above-mentioned predetermined recommendation conditions and predetermined additional conditions are not satisfied, the simulation module 150 receives the calculation conditions of the target learning model from the receiving module 110 and provides information about the stored learning model from the storage module 120. All information can be delivered.

The simulation module 150 can combine stored learning modules based on data input to the target learning model and data calculated from the target learning model.

As a specific example, the simulation module 150 may select a stored learning model (starting stored learning model) in which the input data to be input to the target learning model matches the input data to be input to any stored learning model.

Additionally, the simulation module 150 may select a stored learning model (final stored learning model) in which the output data to be output from the target learning model matches the output data to be output from any stored learning model.

Here, matching may mean that the input data type and data content match at least partially.

For example, the data format may mean text format, video format, image format, etc.

For example, data content may refer to qualitative content about data such as parking lot images, vehicle images, and forest fire images.

The method of determining the identity of the meaning of the content can be done through natural language reasoning, and detailed explanations thereof can be omitted within the scope of known technology.

In addition, the simulation module 150 uses data matching the output data of the starting storage learning model as input data, and data matching the input data of the ending storage learning model as output data of the storage learning model (mediate storage model). learning model) can be selected.

The intermediate storage learning model may not exist, and may be one or multiple. If there are multiple intermediate storage learning models, the output data of the preceding intermediate storage learning model and the input data of the succeeding intermediate storage learning model can be aligned to match each other, and the input data of the most preceding intermediate storage learning model is the starting storage. The simulation module 150 may select the intermediate storage learning model so that it matches the output data of the learning model, and the output data of the most recent intermediate storage learning model matches the input data of the final storage learning model.

When multiple learning models are input, the simulation module 150 can generate a simulation model that calculates the state of the final learning model through machine learning/deep learning.

Machine learning may use a back propagation algorithm, which is an algorithm that updates the weights of a neural network using labeled data of the output layer, but is not limited to this.

In addition, since the deep neural network and back propagation algorithm are known in the art, detailed description thereof may be omitted.

A virtual learning model can be created by inputting the starting stored learning model, ending stored learning model, and/or intermediate stored learning model selected by the simulation module 150 into the simulation model.

Additionally, the simulation module 150 can calculate the calculation conditions of the virtual learning model based on the stored learning models input to the simulation model.

Here, the simulation module 150 can simulate generating a virtual learning model by combining all the numbers of cases so that the calculation conditions of the virtual learning model are as similar as possible to the calculation conditions of the target learning model.

For example, the learning time can be calculated as the sum of the learning times of the stored learning models input to the simulation model.

For example, the amount of learning data can be calculated as the sum of stored learning models input to the simulation model.

The simulation module 150 may not be able to calculate the calculation conditions of a virtual learning model that cannot be estimated as stored learning models.

The similarity judgment module 130 may determine the similarity based only on the calculation conditions calculated when determining the similarity between the virtual learning model and the target learning model.

The simulation module 150 can simulate the performance of the recommendation learning model to determine whether the recommendation learning model operates normally.

As a specific example, the simulation module 150 can input virtual data suitable for the recommendation learning model and check the generated data to simulate whether the recommendation learning model operates normally and whether the result value is normally derived.

If, as a result of the performance simulation of the simulation module 150, the recommended learning model does not operate normally or does not produce normal result values, the recommendation module does not recommend the recommended learning model to the consumer and excludes the stored learning model corresponding to the recommended learning model. can do. Afterwards, the recommendation module can select a learning model based on newly predetermined recommendation conditions as a recommended learning model and recommend it to the consumer.

The learning suggestion module 160 may suggest creating a learning model to the administrator when predetermined learning conditions are satisfied.

The predetermined learning condition may be a condition under which the predetermined recommendation condition is not satisfied.

As a specific example, the predetermined learning condition may be a condition in which there are a predetermined number or more similar target learning models for which the predetermined recommendation condition, the predetermined additional condition, and the predetermined merge condition are not satisfied.

The learning suggestion module 160 calculates conditions for a target learning model (hereinafter referred to as a rejection learning model) for which the predetermined recommendation condition, the predetermined additional condition, and the predetermined merge condition are not satisfied from the recommendation module 140. can be delivered and stored.

The learning proposal module 160 can vectorize the calculation conditions of the rejection learning model and cluster them.

Since the method of clustering data uses known techniques, detailed description thereof may be omitted.

If one clustered group consists of a predetermined number of rejection learning models, the learning limit module can suggest to the administrator to create a learning model using the calculation conditions calculated by reconstructing the vector value of the average of the group.

To this end, the interface module 180 can produce an interface that displays calculation conditions suggesting learning to the administrator and inquires about acceptance.

The price calculation module 170 can calculate the cost of using, selling, or licensing the stored learning model in a cloud environment.

The price calculation module 170 can set the cost according to the production difficulty of the storage learning model.

For example, the price calculation module 170 can set a higher cost as the manufacturing difficulty increases. Conversely, the price calculation module 170 can set a lower cost as the manufacturing difficulty decreases.

The price calculation module 170 may calculate the cost to be charged based on the degree of similarity between the target learning model and the recommended learning model to be recommended.

Here, the degree of similarity can be calculated based on the distance between the two models.

As a result, consumers can be compensated for the disadvantages caused by not being able to use the desired learning model.

As the learning model recommendation method is implemented, the interface module 180 calculates an interface that can display or collect information necessary for the manager or consumer, and provides it to the manager or consumer's computing device.

The interface module 180 can calculate an interface for a learning model brokerage platform that can buy and sell learning models and provide the interface to buyers or sellers.

A detailed explanation of the learning model brokerage platform will be provided later.

The interface module 180 can transmit the learning model recommendation method and the information necessary to operate the learning model brokerage platform to the consumer.

The interface module 180 can calculate an interface through which the consumer can check and download the recommended learning model and transmit it to the consumer.

The interface module 180 can transmit calculation conditions that were requested to be learned by the administrator but were rejected to the seller of the learning model.

The seller list of the learning model may consist of sellers who use the learning mediation platform.

Through this, the learning model platform and learning model sales market can be activated and the needs of consumers can be met.

The learning model recommendation device not only recommends learning models needed internally to other devices within the system, but can also recommend appropriate learning models to cloud service users or individual Internet users.

The management device 200 according to an embodiment of the present invention includes an operation module 210 that operates a virtual machine on a physical server, a repair module 220 that corrects errors when the virtual machine operates abnormally, and a virtual machine on the physical server. It may be provided with a migration module 230 that calculates and migrates the optimal arrangement of machines, and a mediation module 240 that mediates the learning model recommendation device and the monitoring device 300 with each other.

The operation module 210 can perform operations such as creating and deleting virtual machines according to user requests.

The repair module 220 can repair the virtual machine when an error occurs in the virtual machine according to a predetermined repair method.

In the repair module 220, response methods according to errors may be stored in a database. The repair module 220 may receive logs and metadata of a virtual machine that generates an error from the monitoring device 300 and correct the error of the virtual machine according to a pre-stored response method.

However, it is not limited to this, and the repair method of the repair module 220 can be modified in various ways at a level that is obvious to an ordinary technician.

For example, the repair module 220 may request the model operation module 330 to calculate an appropriate repair method according to the error. Accordingly, one monitoring learning model stored in the model operation module 330 is operated, and an appropriate repair method according to the error can be calculated, and the calculated repair method is transmitted from the model operation module 330 to the repair module 220. It can be.

The migration module 230 can migrate virtual common people to place a virtual machine on a physical machine.

For example, the migration module 230 may request the model operation module 330 to predict the workload of the virtual machine, and predict that one monitoring learning model stored in the model operation module 330 will be operated and generated in the cloud service. A predicted workload can be calculated. The workload of the virtual machine predicted by the model operation module 330 can be transferred to the migration module 230, and the migration module 230 provides a virtual machine whose pre-specified objective function is minimized based on the workload of the predicted virtual machine. The optimal placement of machines can be calculated, and virtual machines can be placed on physical servers accordingly.

Accordingly, detailed descriptions may be omitted within the scope of known technology.

The mediation module 240 can mediate the monitoring device 300 and the learning model recommendation device so that the monitoring device 300 can appropriately recommend the required learning model from the learning model recommendation device.

The mediation module 240 can receive information necessary for the operation module 210, repair module 220, and migration module 230 from the monitoring device 300 in connection with the monitoring device 300.

However, without being limited to this, the monitoring device 300 may directly communicate with the learning model recommendation device and exchange information necessary for recommending the learning model.

The monitoring device 300 according to an embodiment of the present invention monitors the operation of at least one virtual machine running on a physical server within the allocated resources of the cloud server integrated management system, A transmission and reception module 310 that collects operation information generated from the virtual machine, a monitoring module 320 that monitors whether the virtual machine operates abnormally based on the operation information collected by the transmission and reception module 310, Based on the request of the monitoring module 320, a model operation module 330 operates a surveillance learning model that assists the monitoring function of the monitoring module 320, and allows the surveillance learning model to be operated within the allocated resources. , It may include a schedule module 340 that determines the operation timing of the monitoring learning model through a predetermined timing determination method.

Additionally, the cloud server integrated management system may further include a model selection module 350 that selects the monitoring learning model that satisfies predetermined operation conditions.

Additionally, the cloud server integrated management system may further include a second storage module 360 that stores all information necessary to implement the monitoring method.

The transmitting and receiving module 310 can collect operation information generated from the virtual machine.

In addition, the transmitting and receiving module 310 can transmit and receive information necessary to implement the monitoring method with external servers, clients, and internal components of the cloud server integrated management system.

For example, operation information may include both metric data and log data generated from a virtual machine.

The monitoring module 320 can monitor in real time whether the virtual machine is operating abnormally based on the operation information.

For this purpose, the monitoring module 320 may store error patterns of operation information that can be judged as abnormal, and if a similar error pattern is detected, it may be determined that an error has occurred in the corresponding virtual machine.

The monitoring module 320 can transmit information about a virtual machine in which an error occurred to the management device 200.

The monitoring module 320 can produce an interface through which information about the load generated by the virtual machine, the load status of the physical server, the list of physical servers in operation, etc. can be displayed to the cloud server integrated management system administrator.

Additionally, the monitoring module 320 can produce an interface through which recipients can check the usage status of the virtual machine, incurred cloud usage costs, etc.

Here, the interface can be expressed through visual images such as graphs.

The monitoring function in the present invention calculates all information generated in the cloud service, such as detecting abnormal signs of virtual machines, predicting virtual machine load, predicting load of the cloud integrated operation system, amount of power generated by the physical server, and detecting abnormal signs of physical machines. It may include detection, monitoring, and prediction functions.

The model operation module 330 can operate the monitoring learning model, which is a learning model that helps the management device 200 operate or the monitoring function of the monitoring module 320.

For example, the monitoring learning model may further include a first predictive monitoring learning model that predicts the load generated in the cloud server integrated management system for a predetermined future period.

For example, the surveillance learning model may further include a second predictive surveillance learning model that predicts the workload generated in a virtual machine on a physical server for a predetermined future period.

In one example, the predetermined future period may be 30 days.

However, it is not limited to this, and the predetermined future period can be modified in various ways at a level that is obvious to those skilled in the art.

For example, the monitoring learning model may further include an error monitoring learning model that monitors whether an error occurs in the cloud service.

The error monitoring learning model can determine whether there is an error in the virtual machine through operation information and calculate a solution to resolve the error.

A monitoring learning model is stored in the model operation module 330.

The second predictive monitoring learning model can calculate the workload predicted to occur in all virtual machines installed on the physical server.

The model selection module 350 may select the monitoring learning model that satisfies predetermined operation conditions, and transmit the selected monitoring learning model to the model operation module 330.

The predetermined operating condition may be a condition in which a load lower than a predetermined threshold is applied.

The model selection module 350 may change the predetermined threshold based on the load predicted by the first predictive monitoring learning model.

The model selection module 350 can receive information about the resources used among all resources to the cloud integrated operation system for a predetermined period stored in the storage module in order to calculate a predetermined threshold.

In contrast, the model selection module 350 can receive information about the spare resources of the cloud integrated operation system for a predetermined future period through the first predictive monitoring learning model of the model operation module 330.

Here, free resources may mean the value obtained by subtracting used resources from total resources.

For example, the predetermined threshold may mean 70% of average free resources over a future period or a predetermined period.

However, it is not limited to this, and the exact value of the predetermined threshold can be modified in various ways at a level that is obvious to those skilled in the art.

The model selection module 350 may request a learning model recommendation device to recommend a monitoring learning model that satisfies the predetermined operation conditions.

The model selection module 350 may transmit the calculation conditions of the monitored learning model together when requesting a monitored learning model required for the learning model recommendation device.

The model selection module 350 receives a monitored learning model (recommended learning model) from the learning model recommendation device and transmits it to the model operation module 330, and the model operation module 330 can store the monitored learning model.

The schedule module 340 may determine the operation timing of the error monitoring learning model based on the load predicted by the first predictive monitoring learning model.

Here, the operation time may be a term that includes not only the starting point but also the progressing time.

For example, the predetermined timing determination method may be a method of determining the operation timing at a time when the load and/or workload is generated below a predetermined ratio based on the resources allocated to the cloud integrated operation system.

As a specific example, the schedule module 340 may divide a predetermined future period into predetermined division periods.

Here, the partition period may be in units of hours, days, or weeks.

Here, the schedule module 340 can calculate the average used resources predicted to be generated in the cloud integrated operation system for each partition period.

The schedule module 340 may determine the operation period to be a period when the used resources calculated for each section period are below a predetermined ratio based on the total resources.

Here, the predetermined ratio (first ratio) may be 10%.

However, it is not limited to this, and the predetermined ratio can be modified in various ways at a level that is obvious to those skilled in the art.

The schedule module 340 can determine the operation time of the monitoring learning model differently from the method described above.

For example, the schedule module 340 may determine the operation time of the error monitoring learning model based on the load predicted by the second predictive monitoring learning model.

For example, the schedule module 340 may determine the operation time of the error monitoring learning model by considering both the load predicted by the first predictive monitoring learning model and the workload predicted by the second predictive monitoring learning model.

For example, the schedule module 340 may determine the operation time of the current monitoring learning model based on the load previously predicted by the first predictive monitoring learning model stored in the second storage module.

For example, the schedule module 340 may determine the operation time of the current surveillance learning model based on the workload previously predicted by the second predictive surveillance learning model stored in the second storage module.

For example, the schedule module 340 considers both the workload previously predicted by the first predictive monitoring learning model and the workload previously predicted by the second predictive monitoring learning model, which have already been processed and stored in the second storage module, to determine the current workload. You can decide when to operate the monitoring learning model.

In the method of determining the operation time by utilizing the load and/or workload predicted in the past, it may be desirable for the operation time to be included within the period of the load and/or workload predicted in the past.

When a monitoring learning model is operated based on the past to predict new future loads and/or workloads, the prediction period is extended into the future, and the newly predicted loads and/or workloads are stored in the second storage module. It can be.

To this end, the load and workload predicted by the first predictive monitoring learning model and/or the second predictive monitoring learning model may be stored in the second storage module 360.

Hereinafter, detailed description may be omitted to the extent of overlap with the above-mentioned content.

For example, the predetermined timing determination method may be a method of determining the operation timing at a time when the least load is generated among the predetermined future period.

As a specific example, the schedule module 340 may divide a predetermined future period into predetermined division periods.

Here, the partition period may be in minutes, hours, days, or weeks.

Here, the schedule module 340 can calculate the average used resources generated in the cloud integrated operation system for each division period.

The schedule module 340 may determine the partition period with the lowest average used resources calculated for each partition period as the operation time.

The schedule module 340 determines the operation time by selecting a time when the load on the cloud integrated operation system will be low, thereby effectively solving the problem that the operation rate of the cloud integrated operation system is hindered by the operation of the operation learning model.

The second storage module 360 may store all information necessary to implement the monitoring method.

The second storage module 360 may also store overall resource information of a past cloud integrated operation system and past used resource information.

The second storage module may include internal memory and/or external memory.

For example, internal memory may include volatile memory (e.g., DRAM, SRAM, or SDRAM), non-volatile memory (e.g., one time programmable ROM (OTPROM), PROM, EPROM, EEPROM, mask ROM, flash ROM, flash memory, hard memory, etc. may include at least one of a drive, or solid state drive (SSD)

External memory may be a flash drive, such as compact flash (CF), secure digital (SD), Micro-SD, Mini-SD, extreme digital (xD), multi-media card (MMC), or memory stick. It may include etc.

The storage module included in the learning model recommendation device can be defined as the first storage module 120, and the storage module included in the monitoring device 300 can be defined as the second storage module 360.

Hereinafter, the learning model recommendation method and monitoring method implemented by each module will be described in detail.

Figure 4 is a flow chart of a learning model recommendation method implemented by the learning model recommendation device of the cloud integrated operating system according to an embodiment of the present invention.

Hereinafter, detailed description may be omitted to the extent that it overlaps with the above-described content.

Referring to Figure 4, the learning model recommendation method according to an embodiment of the present invention is implemented by a learning model recommendation device, and recommends a learning model calculated by machine learning or deep learning according to the request of the consumer. In the recommendation method, a request receiving step in which the request of the consumer and the calculation conditions of the target learning model, which is the learning model requested by the consumer, are received by the receiving module; a stored learning model, which is a learning model created in the past, by the storage module; A condition storage step in which the calculation conditions of the stored learning model are stored, a similarity judgment step in which the similarity between the target learning model and the stored learning model is determined by a similarity judgment module, and the storage stored in the storage module by a recommendation module. Among the learning models, a recommendation step may be included in which a learning model that satisfies predetermined recommendation conditions is recommended to the consumer.

In addition, the learning model recommendation method may further include an interface calculation step in which an interface through which the consumer can check and download the recommended learning model is calculated by an interface module.

Here, the recommendation step is generated by combining the plurality of stored learning models with a predetermined additional condition that determines whether the condition is satisfied based on the second similarity lower than the first similarity even if the predetermined recommendation condition is not satisfied. This may be a step of selecting or calculating a learning model to recommend to the consumer by determining whether predetermined merge conditions are satisfied based on a virtual learning model.

The consumer can use a computing device to input the calculation conditions of the target learning model through the interface received from the interface module.

The collection module may receive the calculation conditions of the target learning model from the consumer's computing device and transmit the calculation conditions to the similarity judgment module.

The similarity judgment module can determine the similarity between the target learning model and the stored learning model stored in the storage module.

The similarity judgment module can transmit the similarity judgment result between the target learning model and the stored learning model to the recommendation module.

Figure 5 is a diagram for explaining predetermined recommendation conditions and predetermined additional conditions in the learning model recommendation method according to an embodiment of the present invention.

Referring to Figures 4 and 5(a), the recommendation module can determine whether a stored learning model that satisfies predetermined recommendation conditions exists.

The predetermined recommendation condition may be a condition that is similar to the target learning model by a first degree of similarity or higher.

Similarity greater than the first similarity may mean that it exists below the first distance (X10) from the target learning model (A10).

If there are multiple stored learning models that satisfy predetermined recommendation conditions, the recommendation module can select the stored learning model most similar to the target learning model as the recommended learning model (C10).

Even if there is no stored learning model that satisfies the predetermined recommendation conditions, the recommendation module can calculate the recommended learning model according to a predetermined calculation method if the predetermined additional conditions are satisfied.

A predetermined additional condition is that the cluster of stored learning models that are similar to the target learning model by more than the second similarity and less than the first similarity, and exist by the second similarity or more based on the target learning model, is predetermined. It may be a condition that exists in numbers or less.

This may be to supplement and correct when the calculation conditions of the target learning model cannot be accurately specified due to the consumer's inexperience.

Referring to Figures 5(b) and 5(c), there are 1 and 3 clusters of the stored learning model, and predetermined additional conditions can be satisfied.

Here, the cluster can be defined as being formed when the number of stored learning models within a predetermined distance based on the central point of the cluster is more than a predetermined number.

Here, the predetermined distance and the predetermined number can be modified in various ways at a level that is obvious to those skilled in the art.

The predetermined calculation method is a method of selecting the stored learning model closest to the average of the cluster as the recommended learning model when there is only one cluster of the stored learning models that exist with the second similarity or higher based on the target learning model. It can be.

Similarity beyond the second similarity may mean that it exists below the second distance (X20) from the target learning model.

Referring to FIG. 5(b), one cluster may be formed within the first distance (X10) and below the second distance (X20).

The recommendation module calculates the calculation conditions by inversely calculating the vector of the average point (M10) of one cluster within the first distance (X10) and below the second distance (X20) based on the target learning model (A10), and recommends learning. You can select it as a model.

The predetermined calculation method is a combination of the stored learning models closest to the average of each cluster when the number of clusters of the stored learning models that exist with the second similarity or higher based on the target learning model is more than one or less than a predetermined number. This may be a method of selecting a recommended learning model.

Referring to FIG. 5(c), three clusters may be formed within the first distance (X10) and below the second distance (X20) based on the target learning model (A10).

For example, the average point (M11) of the calculation conditions of the stored learning models forming the first cluster, the average point (M12) of the calculating conditions of the stored learning models forming the second cluster, and the average calculating conditions of the stored learning models forming the third cluster. Calculation conditions can be calculated by taking a weighted average according to the distance to the points (M13).

Here, the closer the distance to the target learning model is, the higher the weight can be.

For example, the distance between the calculation condition average point (M11) of the stored learning model forming the first cluster and the target learning model is '10', and the calculation condition average point (M12) of the stored learning model forming the second cluster and the target learning model are '10'. Assuming that the distance from the model is '20' and the distance between the average point (M13) of the calculation condition of the stored learning model forming the first cluster and the target learning model is '30', the first cluster, the second cluster, and the The weight of the 3 clusters is set as 3:2:1 so that the above-described calculation conditions can be calculated.

In this way, the recommendation module calculates the calculation conditions by inversely calculating the vector for the calculated average value, and can be selected as a recommendation learning model.

However, the present invention is not limited to this, and the learning model recommendation method may be implemented in other ways.

For example, the recommendation module can determine whether the predetermined recommendation conditions are satisfied, including the virtual learning model.

As a specific example, rather than proceeding with a simulation merging stored learning models only when predetermined merging conditions are satisfied, a simulation merging stored learning models may proceed from the stage of determining whether predetermined recommendation conditions are satisfied.

The similarity judgment module can determine the degree of similarity between the target learning model and the virtual learning model using a predetermined similarity judgment method.

The recommendation module can calculate a recommended learning model by determining whether predetermined recommendation conditions are satisfied for the entire stored learning model and virtual learning model.

Through this, it is possible to recommend and provide learning models that fit the diverse needs of consumers.

When selecting a virtual learning model as a recommended learning model, the recommendation module can produce a coding method or arrangement for merging the stored learning models that make up the virtual learning model, and a description of data output and input methods.

When a virtual learning model is recommended, the interface module can check all the stored learning models that make up the virtual learning model and calculate an interface that can be downloaded, and can calculate an interface that can download related manuals.

Although FIG. 5 shows an example of vectorization along two axes, the present invention is not limited to this, and the calculation conditions can be vectorized into three axes to determine similarity and whether the above-mentioned conditions are satisfied.

Even if the predetermined recommendation condition and the predetermined additional condition are not satisfied, the recommendation module may select the recommended learning model based on the stored learning models that satisfy the predetermined merge condition.

As a specific example, the recommendation module may select a virtual learning model that satisfies predetermined merging conditions as a recommended learning model based on a virtual learning model received from the simulation module.

If there is no stored learning model that satisfies the predetermined recommendation condition, the first predetermined condition, and the predetermined merge condition, the learning proposal module may store the calculation conditions of the target learning model.

Additionally, the interface model can produce an interface that displays a message to the consumer that there is no stored learning model similar to the target learning model and transmit it to the consumer's computing device.

Figure 6 is a diagram illustrating a learning model mediation interface screen provided by the interface module of the learning model recommendation device according to an embodiment of the present invention.

Referring to FIG. 6, the interface module can produce an interface for a learning model brokerage platform that mediates between a seller who creates and sells a learning model and a buyer who purchases the learning model.

For example, the specifications (calculation conditions) of the learning model created by the seller may be listed, and the price may be set and displayed on the interface.

Here, consumers can select, purchase, rent, and use the learning model needed for their program, as if shopping on the Internet.

FIG. 7 is a flowchart illustrating a process in which a learning model recommendation device selects a learning model to recommend to a monitoring device according to an embodiment of the present invention, and FIG. 8 shows model selection of a monitoring device according to an embodiment of the present invention. This is a diagram to explain how the module calculates the first operation time and the second operation time.

Referring to Figures 7 and 8, the model selection module may transmit the first operation time and the second operation time together when transmitting calculation conditions to the learning model recommendation device.

The model selection module can calculate the first operation time and the second operation time while determining the operation time.

The first operation time may mean the total time of a continuous partition period in which the average used resources of the cloud integrated operation system are generated below the first rate.

The second uptime means that the average used resources of the cloud integrated operation system are generated below the second rate and may mean the time that is the sum of the first uptime and the total continuous time adjacent to the first uptime.

Referring to FIG. 8(a), for example, when the partition period is in hours, the period during which the average used resources of the cloud integrated operation system are predicted to be calculated below the first ratio (W10) is the operation time (T12, T13). , T14), and the first operation time may be 3 hours.

Here, since the average used resources of the partition periods (T11 and T15) adjacent to the operation time are higher than the second ratio (W20), the second operation time may not exist.

Referring to FIG. 8(b), for example, when the partition period is in hours, the period during which the average resource usage of the cloud integrated operation system is predicted to be calculated below the first ratio (W10) is the operation time (T12, T13). , T14), and the first operation time may be 3 hours.

In addition, since the average used resources of the partition periods (T11, T15) adjacent to the operation time are lower than the second ratio (W20), the second operation time may be 5 hours, which is 2 hours plus the first operation time.

The second operation time is somewhat difficult for the cloud integrated operation system, but can refer to the period during which the learning model can be operated within an acceptable range.

The learning model recommendation device may receive a request for recommending a learning model, calculation conditions, and a first operation time and a second operation time directly from the monitoring device or from an intermediate module.

Among the above-mentioned contents, the target learning model can be understood by changing it to a surveillance learning model for which recommendation is desired.

The learning model recommendation device can calculate a stored learning model similar to the surveillance learning model requested for recommendation in the same way as described above, and detailed descriptions thereof may be omitted to the extent of overlap with the above-mentioned content.

However, unlike the above-described recommendation method, when a learning model is recommended internally in the system, it is possible to determine whether a recommended learning model exists by further considering the operation time.

The simulation module can simulate the time required when the stored learning model stored in the first storage module is operated.

Here, the amount of data expected to be input can be received together with receiving the calculation condition from the monitoring device.

The simulation module can transmit the expected operation time of the stored learning model to the recommendation module.

The recommendation module may calculate whether there is a stored learning model that satisfies a predetermined recommendation condition, a predetermined additional condition, or a predetermined merge condition within the stored learning module that satisfies the first operation time. Detailed description of this may be omitted to the extent that it overlaps with the above-described content.

Here, the expected operating time of the stored learning module that satisfies the predetermined merge conditions can be calculated by adding up the expected operating times of the stored learning models being used (for example, a starting stored learning model and a ending stored learning model).

If there is no stored learning model that satisfies the predetermined recommendation condition, predetermined additional condition, or predetermined merge condition within the storage learning module that satisfies the first operation time, the recommendation module stores the storage learning module that satisfies the second operation time. Within the learning module, it can be calculated whether there is a stored learning model that satisfies predetermined recommendation conditions, predetermined additional conditions, or predetermined merge conditions. Detailed description of this may be omitted to the extent that it overlaps with the above-described content.

Here, the expected operating time of the stored learning module that satisfies the predetermined merge conditions can be calculated by adding up the expected operating times of the stored learning models being used (for example, a starting stored learning model and a ending stored learning model).

If there is no stored learning model that satisfies the predetermined recommendation condition, predetermined additional condition, or predetermined merge condition within the first operation time and the second operation time, the absence of the recommended learning model is transmitted to the model selection module. The calculation conditions of the monitoring learning model can be stored in the learning proposal module and used for the above-mentioned learning.

The model selection module can select the recommended learning model received from the learning model recommendation device as a monitoring learning model and transmit it to the model operation module.

The model selection module can select the recommended learning model received from the learning model recommendation device as a monitoring learning model and transmit it to the model operation module.

A monitoring method according to an embodiment of the present invention includes, through a monitoring device, monitoring the operation of at least one virtual machine running on a physical server within the allocated resources of a cloud integrated operation system, By collecting operation information generated from the virtual machine, by a monitoring module, by monitoring whether the virtual machine is operated abnormally based on the operation information, by a model operation module, by the monitoring module Based on the request, the monitoring learning model that helps the monitoring function of the monitoring module is operated, and the monitoring learning model is operated within the allocated resources by the schedule module, through a predetermined time point determination method. It may include a step in which the timing of operation of the surveillance learning model is determined.

In addition, the monitoring method operates based on the load (resources used) predicted by the schedule module in a first predictive monitoring and learning model that predicts the load generated in the cloud integrated operation system for a predetermined future period. It may further include a step in which the timing is determined.

In addition, the monitoring method may further include a step of selecting a monitoring learning model that satisfies predetermined operation conditions by a model selection module.

In addition, the monitoring method may further include a step of changing the predetermined threshold based on the load predicted by the first predictive monitoring learning model by a model selection module.

The learning model recommendation device according to an embodiment of the present invention is a learning model recommendation device that recommends a learning model calculated by machine learning or deep learning according to the request of the consumer, and includes the request of the consumer and the learning model requested by the consumer. A receiving module that receives the calculation conditions of the target learning model; a storage module that stores a stored learning model, which is a learning model created in the past, and calculation conditions for the stored learning model; a similarity judgment module that determines the similarity between the target learning model and the stored learning model; And a recommendation module that recommends to the consumer a learning model that satisfies predetermined recommendation conditions among the stored learning models stored in the storage module as a recommended learning model, wherein the predetermined recommendation conditions include the target learning model and The conditions may be more similar than 1 similarity.

In addition, the recommendation module calculates the recommended learning model according to a predetermined calculation method even if the predetermined recommendation condition is not satisfied, if the predetermined additional condition is satisfied, and the predetermined additional condition is the target learning. The condition may be that the cluster of the stored learning model that is similar to the model by a second similarity or higher and is similar to the first similarity or higher, and exists by a predetermined number or less based on the target learning model, exists in the second similarity or higher. .

In addition, the predetermined calculation method is that, when there is only one cluster of the stored learning models that exist with the second similarity or higher based on the target learning model, the stored learning model closest to the average of the cluster is used as the recommended learning model. This may be a method of selection.

In addition, the predetermined calculation method uses the stored learning model that is closest to the average of each cluster when the cluster of the stored learning model that exists with the second similarity or higher based on the target learning model is more than one or less than a predetermined number. This may be a method of selecting the recommended learning model through a combination of .

In addition, the recommendation module selects the recommended learning model based on the stored learning models that satisfy the predetermined merging condition even if the predetermined recommendation condition and the predetermined additional condition are not satisfied, and the predetermined merging condition This may be a condition where, when the stored learning models that are similar to the target learning model and less than the second similarity are connected, the similarity to the target learning model is equal to or greater than the first similarity.

In addition, when arbitrary learning models are connected to each other, it further includes a simulation module that simulates the state of the final learning model, and the simulation module calculates the state of the final learning model when a plurality of learning models are input. Simulation models can be created through deep learning.

In addition, if the predetermined learning condition is satisfied, it further includes a learning suggestion module that proposes to the administrator to create a learning model, and the predetermined learning condition may be a condition under which the predetermined recommendation condition is not satisfied. .

In addition, it further includes a price calculation module for calculating a cost to be charged while recommending the stored learning model, wherein the price calculation module is configured to calculate a cost to be charged based on the degree of similarity between the target learning model and the recommended learning model to be recommended. can be calculated.

The learning model recommendation method according to an embodiment of the present invention is implemented by a learning model recommendation device and recommends a learning model calculated by machine learning or deep learning according to the request of the consumer, comprising: A request receiving step in which the request of the consumer and the calculation conditions of the target learning model, which is the learning model requested by the consumer, are received by the module; A condition storage step in which a stored learning model, which is a learning model created in the past, and calculation conditions of the stored learning model are stored by a storage module; A similarity judgment step in which the similarity between the target learning model and the stored learning model is determined by a similarity judgment module; And a recommendation step in which, by a recommendation module, a learning model that satisfies predetermined recommendation conditions among the stored learning models stored in the storage module is recommended to the consumer, wherein the predetermined recommendation conditions include the target learning model and the second learning model. The conditions may be more similar than 1 similarity.

In addition, in the recommendation step, even if the predetermined recommendation condition is not satisfied, the plurality of stored learning models are combined with a predetermined additional condition for determining whether the condition is satisfied based on a second similarity lower than the first similarity. This may be a step of selecting or calculating a learning model to recommend to the consumer by determining whether predetermined merge conditions are satisfied based on the generated virtual learning model.

A monitoring device according to an embodiment of the present invention is a monitoring device that monitors the operation of at least one virtual machine running on a physical server within the allocated resources of a cloud integrated operating system, wherein the operation occurring in the virtual machine is monitored. Transmission/reception module that collects information; a monitoring module that monitors whether the virtual machine operates abnormally based on the operation information collected by the transmission and reception module; A model operation module that operates a monitoring learning model that assists the monitoring function of the monitoring module, based on the request of the monitoring module; and a schedule module that determines the operation time of the surveillance learning model through a predetermined timing determination method so that the surveillance learning model can be operated within the allocated resources.

In addition, the monitoring learning model includes a first predictive monitoring learning model that predicts the load generated in the cloud integrated operating system for a predetermined future period, and the schedule module predicts the load generated in the cloud integrated operating system for a predetermined future period. The operation timing can be determined based on the load.

Additionally, the predetermined timing determination method may be a method of determining the operation timing at a time when the load is generated below a predetermined ratio based on the allocated resources.

Additionally, the predetermined timing determination method may be a method of determining the operation timing at a time when the least load is generated among the predetermined future period.

In addition, it further includes a model selection module that selects the monitoring and learning model that satisfies predetermined operation conditions, and the predetermined operation condition may be a condition in which a load lower than a predetermined threshold is applied.

In addition, the learning model includes a first predictive monitoring learning model that predicts a load generated in the cloud integrated operation system for a predetermined future period, and the model selection module makes predictions from the first predictive monitoring learning model. The predetermined threshold can be changed based on the load.

A cloud integrated operating system according to an embodiment of the present invention includes a management device for managing virtual machines running on a physical server; A monitoring device that monitors the operation of the virtual machine; And a learning model recommendation device that recommends a monitoring learning model necessary for the monitoring device to implement the monitoring function, wherein the monitoring device includes a transmission/reception module that collects information generated from the virtual machine, and a transmission/reception module that collects information generated by the virtual machine. A monitoring module that monitors whether the virtual machine operates abnormally based on the information of the virtual machine, and a model operation that operates the surveillance learning model that assists the monitoring function of the monitoring module based on the request of the monitoring module. module, a schedule module that determines the operation time of the surveillance learning model through a predetermined timing decision method so that the surveillance learning model can be operated within the allocated resources, and a model operation module that satisfies predetermined operation conditions. It further includes a model selection module for selecting the monitoring learning model, wherein the model selection module requests the learning model recommendation device to recommend a monitoring learning model that satisfies the predetermined operation conditions, and requests the learning model recommendation device to recommend the learning model. The recommendation device may recommend a learning model that satisfies a predetermined recommendation condition, which is a condition similar to the monitored learning model by a first similarity or higher, among pre-stored stored learning models.

In addition, the learning model recommendation device determines whether the condition is satisfied based on a second similarity lower than the first similarity, even if there is no learning model that satisfies the predetermined recommendation condition among the pre-stored stored learning models. A learning model that satisfies a predetermined merge condition that determines whether the condition is satisfied can be recommended based on an additional condition or a virtual learning model created by combining a plurality of the stored learning models.

In addition, the predetermined additional condition is similar to the surveillance learning model by more than the second similarity and less than the first similarity, and the pre-stored stored learning that exists more than the second similarity based on the surveillance learning model. This may be a condition in which the number of model clusters is less than a certain number.

A monitoring method according to an embodiment of the present invention includes, through a monitoring device, monitoring the operation of at least one virtual machine running on a physical server within the allocated resources of a cloud integrated operation system, collecting operation information generated from the virtual machine; Monitoring, by a monitoring module, whether the virtual machine operates abnormally based on the operation information; A step of operating, by a model operation module, a monitoring learning model that assists the monitoring function of the monitoring module, based on a request from the monitoring module; And it may include a step of determining, by a schedule module, the operation time of the surveillance learning model through a predetermined timing determination method so that the surveillance learning model can be operated within the allocated resources.

In the attached drawings, in order to more clearly express the technical idea of the present invention, components that are unrelated or less relevant to the technical idea of the present invention are briefly expressed or omitted.

In the above, the configuration and features of the present invention have been described based on the embodiments according to the present invention, but the present invention is not limited thereto, and various changes or modifications may be made within the spirit and scope of the present invention. It is obvious to those skilled in the art, and therefore, it is stated that such changes or modifications fall within the scope of the appended patent claims.

10: Cloud integrated operation system 100: Learning model recommendation device
200: management device 300: monitoring device

Claims (10)

In a learning model recommendation device that recommends a learning model calculated by machine learning or deep learning according to the request of the consumer,
A receiving module that receives the request from the consumer and the calculation conditions of the target learning model, which is the learning model requested by the consumer;
A storage module that stores a storage learning model, which is a learning model created in the past;
a similarity judgment module that determines the similarity between the target learning model and the stored learning model through a predetermined similarity judgment method using the calculation conditions of the target learning model;
a recommendation module that calculates a recommended learning model that is a learning model that satisfies predetermined recommendation conditions based on similarity among the stored learning models stored in the storage module; and
Containing; an interface module that calculates an interface that allows the consumer to check the recommended learning model;
The predetermined recommendation conditions are:
It is a condition that is similar to the target learning model by more than the first similarity,
The recommended module is,
Even if the predetermined recommendation conditions are not satisfied, if the predetermined additional conditions are satisfied, the recommended learning model is calculated according to a predetermined calculation method,
The above predetermined additional conditions are,
A condition for judging based on the second similarity, which is lower than the first similarity,
Learning model recommendation device.
According to paragraph 1,
The predetermined calculation method is,
A method of selecting the stored learning model closest to the average of the cluster as the recommended learning model when there is only one cluster of the stored learning models that exist with the second similarity or higher based on the target learning model,
Learning model recommendation device.
According to paragraph 1,
The above predetermined additional conditions are,
A condition in which the cluster of the stored learning model that is similar to the target learning model by more than the second similarity and less than the first similarity, and exists by the second similarity or more based on the target learning model, exists in a predetermined number or less. person,
Learning model recommendation device.
In a learning model recommendation device that recommends a learning model calculated by machine learning or deep learning according to the request of the consumer,
A receiving module that receives the request from the consumer and the calculation conditions of the target learning model, which is the learning model requested by the consumer;
A storage module that stores a storage learning model, which is a learning model created in the past;
a similarity judgment module that determines the similarity between the target learning model and the stored learning model through a predetermined similarity judgment method using the calculation conditions of the target learning model;
a simulation module that performs a simulation to create a virtual learning model created by connecting the stored learning models; and
A recommendation module that calculates a recommended learning model, which is a learning model that satisfies predetermined recommendation conditions based on similarity among the stored learning models stored in the storage module;
The recommended module is,
Determining whether the predetermined recommendation conditions, including the virtual learning model, are satisfied,
Learning model recommendation device.
According to paragraph 4,
The simulation module is,
When multiple learning models are input, a simulation model that calculates the state of the final learning model is created through machine learning.
Learning model recommendation device.
According to paragraph 4,
The simulation module is,
Selecting the stored learning model that must receive the same input data as the input data to be input to the target learning model, and selecting the stored learning model that outputs the same output data as the output data output from the target learning model, Creating a virtual learning model,
Learning model recommendation device.
According to paragraph 1,
Further comprising a simulation module that simulates the performance of the recommendation learning model to check whether the recommendation learning model operates normally.
Learning model recommendation device.
Add a learning model calculated by machine learning or deep learning at the request of the consumer.
In the learning model recommendation device,
A receiving module that receives the request from the consumer and the calculation conditions of the target learning model, which is the learning model requested by the consumer;
A storage module that stores a storage learning model, which is a learning model created in the past;
a similarity judgment module that determines the similarity between the target learning model and the stored learning model through a predetermined similarity judgment method using the calculation conditions of the target learning model; and
A recommendation module that calculates a recommended learning model for learning models that satisfy predetermined recommendation conditions based on similarity among the stored learning models stored in the storage module;
The predetermined similarity judgment method is,
A method of calculating the degree of similarity by comparing the calculation conditions of the target learning model and the calculation conditions of the stored learning model,
The above calculation conditions are,
Consisting of at least one of the following: learning purpose of the learning model, type of output, learning method, type of learning model, amount of learning data, type of learning data, and learning time,
Learning model recommendation device.
According to clause 8,
The similarity judgment module is,
Calculating the degree of similarity based on the distance between vector information vectorized by preprocessing the calculation conditions of the target learning model and vector information vectorized by the calculation conditions of the stored learning model,
Learning model recommendation device.
In the learning model recommendation method, which is implemented by a learning model recommendation device and recommends a learning model calculated by machine learning or deep learning according to the request of the consumer,
A request receiving step in which the request of the consumer and the calculation conditions of the target learning model, which is the learning model requested by the consumer, are received by the receiving module;
A condition storage step in which a stored learning model, which is a learning model created in the past, and calculation conditions of the stored learning model are stored by a storage module;
A similarity judgment step in which the similarity between the target learning model and the stored learning model is determined by a similarity judgment module based on a predetermined similarity judgment method using calculation conditions of the target learning model;
A simulation is performed to generate a virtual learning model created by connecting the stored learning models by a simulation module; and
A recommendation step in which a learning model that satisfies predetermined recommendation conditions among the stored learning models stored in the storage module is recommended to the consumer by a recommendation module,
The recommended module is,
Determining whether the predetermined recommendation conditions, including the virtual learning model, are satisfied,
Learning model recommendation method.

Images