KR20240063347A - System and method of prediction anomalous signs in smart factory using ensemble learning and kalman filter - Google Patents

Abstract

The present invention relates to a smart factory anomaly prediction system and method using ensemble learning and Kalman filter, which more accurately predicts the occurrence of anomalies in the equipment and operation status of a smart factory and proactively responds thereto. It's about the skills that can be done.

Description

System and method of predicting anomalous signs in smart factory using ensemble learning and Kalman filter {System and method of prediction anomalous signs in smart factory using ensemble learning and kalman filter}

The present invention relates to a system and method for predicting abnormalities in a smart factory using ensemble learning and Kalman filter. More specifically, the present invention relates to a system and method for predicting abnormalities in a smart factory, remotely operating and monitoring the factory by building a smart factory, and detecting abnormalities in equipment and factory operation status. This study relates to a system and method for predicting abnormal symptoms in smart factories using ensemble learning and Kalman filters that can predict when symptoms occur and respond proactively.

Smart manufacturing refers to a technology that actively applies information and communication technologies to the current manufacturing process, and is revolutionizing the global manufacturing industry economy.

The Fourth Industrial Revolution, which is emerging today, refers to the era of building a smart manufacturing environment based on cutting-edge information and communication technology. Although there is no exact definition, SMLC (Smart Manufacturing) is a national R&D consortium established under the leadership of the U.S. federal government. The Ladership Coalition describes smart manufacturing as 'an intensified application of cutting-edge intelligent systems that enable rapid manufacturing of new products, active response to product demand, and real-time optimization of production and supply chain networks.'

The final result of this smart manufacturing will be a smart factory.

A smart factory refers to an operating system that operates the factory remotely and monitors it in real time by building various sensors and communication devices in factory equipment and facilities.

Recently, with the increase in non-face-to-face services, various automated facilities are increasing in the manufacturing industry, and in line with this, smart factories are rapidly increasing as a way to operate production lines more efficiently in an environment where unmanned technology is increasing.

These smart factories are characterized by being open to the outside world by incorporating communications into existing closed factory operation facilities. Although they have the advantage of improving the efficiency of factory operation, they are inevitably exposed to various security threats due to openness. .

In particular, one of the security threats to smart factories is that malicious users enter abnormal operations due to loss, theft, hacking, etc. of terminals that remotely monitor or control the smart factories, or internal users enter abnormal operations for malicious purposes. There is a way to input .

When damage occurs due to security threats occurring in smart factories, etc., it does not stop at simply leaking of operational data, but also causes loss of sales through malicious interruption of production, risk of accidental casualties to workers, etc., as well as the availability of facility systems and the safety of workers. Because it is a directly related part, it is important to detect 'anomaly signs' and perform proactive defense/response to abnormal behavior before abnormal behavior occurs.

However, as a result of applying anomaly detection technology to the process of operating an actual smart factory, even if anomalies are detected, immediate improvement is difficult due to problems such as loss of sales due to production interruption and availability of equipment systems, so the detection results are difficult to improve. It was realized that active response was impossible.

Accordingly, in the present invention, in addition to detecting abnormal signs that indicate the possibility of abnormal behavior occurring, we propose a technology that can more actively perform improvement tasks by more accurately predicting the point in time when abnormal behavior occurs. .

Domestic registered patent No. 10-2196287 (“Smart factory failure diagnosis device using artificial intelligence and failure prediction diagnosis method using the same”) discloses a technology for diagnosing abnormalities in robots or equipment in real time using artificial intelligence. .

Domestic registered patent No. 10-2196287 (registration date 2020.12.22.)

The present invention was created to solve the problems of the prior art as described above. The purpose of the present invention is to sequentially use ensemble learning and Kalman filter to perform trend analysis on the risk calculated through ensemble learning to create smart smart devices. The goal is to provide a smart factory anomaly prediction system and method using ensemble learning and Kalman filters that can more accurately predict the occurrence of anomalies in the factory's equipment and operating status and take active measures against them.

In the smart factory anomaly prediction system using ensemble learning and Kalman filter according to an embodiment of the present invention, a data input unit 100 that receives operational data from a smart factory for which anomaly prediction is desired in advance, the data input unit A data processing unit 200 that performs preprocessing on the operational data by (100) to generate learning data for artificial intelligence learning, and learns the learning data using an ensemble technique applying a plurality of artificial intelligence algorithms. A learning processing unit 300 that performs processing to create a plurality of learning models, a data collection unit 400 that collects real-time operation data generated in a smart factory for which abnormal symptoms are desired to be predicted in real time, and the data collection unit 400 ) A risk analysis unit 500 that inputs the real-time operational data by each into a plurality of learning models by the learning processing unit 300 and then calculates the risk of the real-time operational data using the plurality of generated analysis results. , It is preferable to include an abnormality prediction unit 600 that predicts the timing of abnormal behavior in the smart factory using the risk calculated by the risk analysis unit 500.

Furthermore, the data processing unit 200 includes a first processing unit 210 for labeling the input operational data, a second processing unit 220 for extracting characteristic values for the labeled operational data, and analyzing the extracted characteristic values. Thus, a third processing unit 230 that classifies characteristic values that fall within a preset risk standard range or are outliers based on a preset representative value using extracted characteristic values, and a third processing unit 230 that uses the classified characteristic value information to process it as learning data. It is desirable to include 4 processing units 240.

Furthermore, it is preferable that the data collection unit 400 continuously collects the real-time operational data at predetermined times or whenever an event occurs.

Furthermore, the learning processing unit 300 uses an ensemble technique applying a plurality of artificial intelligence algorithms to apply the learning data generated by the data processing unit 200 to a plurality of artificial intelligence algorithms, thereby performing parallel learning. It is desirable to carry out

Furthermore, the learning processing unit 300 preferably includes a plurality of SVDD (Support Vector Data Description) algorithms and deep learning algorithms, each of which has different characteristic settings for learning processing with the artificial intelligence algorithm.

Furthermore, it is desirable for the risk analysis unit 500 to calculate the risk of the real-time operational data by performing a comparative judgment on the multiple analysis results generated.

Furthermore, the abnormality prediction unit 600 uses the Kalman filter algorithm to perform trend analysis of the risk calculated by the risk analysis unit 500, and based on a preset threshold, the trend result exceeds the threshold. It is desirable to predict when abnormal behavior occurs in a smart factory by using the timing of the occurrence of abnormal behavior.

Furthermore, the smart factory anomaly prediction system using the ensemble learning and Kalman filter generates proactive response information and transmits it to the outside based on the timing of abnormal behavior predicted by the anomaly prediction unit 600. It is desirable to further include a processing unit 700.

A method for predicting abnormalities in a smart factory using ensemble learning and a Kalman filter in which each step is performed by an operation processing means according to another embodiment of the present invention. In advance, a data input step (S100) of receiving operation data from a smart factory that wants to predict abnormalities, preprocessing the operation data by the data input step (S100), and learning for artificial intelligence learning. A data processing step (S200) generated from data, using an ensemble technique applying a number of artificial intelligence algorithms, learning processing of the learning data by the data processing step (S200) is performed to generate a number of learning models. A processing step (S300), a data collection step (S400) that collects real-time operation data generated in a smart factory that wants to predict anomalies in real time, and a learning processing step of the real-time operation data by the data collection step (S400). After each input into a plurality of learning models by (S300), a risk calculation step (S500) of calculating the risk of the real-time operation data using the generated plurality of analysis results, calculated by the risk calculation step (S500) An abnormality prediction step (S600) that predicts when abnormal behavior will occur in the corresponding smart factory using the risk level, and proactive response measures that generate proactive response information and transmit it to the outside based on the predicted abnormal behavior occurrence time. It includes a step (S700), and the learning processing step (S300) preferably performs parallel learning by applying the learning data from the data processing step (S200) to a plurality of artificial intelligence algorithms.

Furthermore, the data processing step (S200) includes a first processing step (S210) of performing labeling on the input operational data, a second processing step (S220) of extracting characteristic values for the labeled operational data, and the extracted A third processing step (S230) of analyzing characteristic values and classifying characteristic values that fall within a preset risk standard range or as outliers based on preset representative values using extracted characteristic values, and using the classified characteristic value information, learn data It is preferable to include a fourth processing step (S240) of processing.

Furthermore, it is desirable that the data collection step (S400) continuously collects the real-time operational data at predetermined times or whenever an event occurs.

Furthermore, the learning processing step (S300) preferably includes a plurality of SVDD (Support Vector Data Description) algorithms and deep learning algorithms, each of which has different characteristic settings for learning processing with the artificial intelligence algorithm.

Furthermore, in the risk calculation step (S500), it is desirable to calculate the risk of the real-time operational data by performing a comparative judgment on the multiple analysis results generated.

Furthermore, the abnormality prediction step (S600) performs a trend analysis of the risk calculated by the risk calculation step (S500) using the Kalman filter algorithm, and based on a preset threshold, the trend result exceeds the threshold. It is desirable to use the excess time to predict when abnormal behavior will occur in a smart factory.

The system and method for predicting abnormal signs in a smart factory using ensemble learning and Kalman filter of the present invention configured as described above, in addition to detecting abnormal signs indicating the possibility of abnormal behavior occurring, determine the point in time when abnormal behavior occurs. By predicting more accurately, there is the advantage of being able to perform improvement tasks more actively.

In other words, by detecting abnormal signs and predicting when abnormal behavior will occur, and establishing specific proactive defense and response plans based on the predicted time, sales losses due to production interruption can be minimized and problems such as availability of equipment systems can be minimized. It has the advantage of being able to resolve.

Specifically, among the various security threats that arise while operating a smart factory, abnormal signs are quickly detected using terminals possessed by the smart factory manager, and through this, the operation status of the smart factory can be monitored and controlled. Various types of security that may arise by detecting abnormal signs in real time by detecting abnormal input information by terminal means that are different from usual, and predicting more accurately when abnormal behavior occurs through trend analysis of the detection results. It has the advantage of being able to prevent accidents more actively.

Figure 1 is an example configuration diagram showing an abnormality prediction system for a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention.
Figure 2 is a flowchart illustrating a method for predicting abnormalities in a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention.

The purpose, features and advantages of the present invention described above will become clearer through the following examples in conjunction with the attached drawings. The following specific structural and functional descriptions are merely illustrative for the purpose of explaining embodiments according to the concept of the present invention. Embodiments according to the concept of the present invention may be implemented in various forms and may be implemented in various forms and may be described in the present specification or application. It should not be construed as limited to the examples. Since the embodiments according to the concept of the present invention can make various changes and have various forms, specific embodiments will be illustrated in the drawings and described in detail in the specification or application. However, this is not intended to limit the embodiments according to the concept of the present invention to a specific disclosed form, and should be understood to include all changes, equivalents, and substitutes included in the spirit and technical scope of the present invention. Terms such as first and or second may be used to describe various components, but the components are not limited to the terms. The above terms are used only for the purpose of distinguishing one component from other components, for example, without departing from the scope of rights according to the concept of the present invention, a first component may be named a second component, and similar Likewise, the second component may also be called the first component. When it is mentioned that a component is connected or connected to another component, it should be understood that it may be directly connected or connected to the other component, but that other components may exist in between. On the other hand, when it is mentioned that a component is directly connected or directly connected to another component, it should be understood that there are no other components in the middle. Other expressions to explain the relationship between components, such as 'between' and 'immediately between' or 'adjacent to' and 'directly adjacent to', should be interpreted similarly. The terms used in this specification are merely used to describe specific embodiments and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, terms such as include or have are intended to designate the existence of a described feature, number, step, operation, component, part, or combination thereof, but are intended to indicate the presence of one or more other features, numbers, steps, operations, It should be understood that this does not exclude in advance the possibility of the presence or addition of components, parts, or combinations thereof. Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the technical field to which the present invention pertains. Terms as defined in commonly used dictionaries should be interpreted as having meanings consistent with the meanings they have in the context of the related technology, and unless clearly defined in this specification, should not be interpreted in an idealized or overly formal sense. No. Hereinafter, the present invention will be described in detail by explaining preferred embodiments of the present invention with reference to the attached drawings. The same reference numerals in each drawing indicate the same member.

In addition, a system refers to a set of components including devices, mechanisms, and means that are organized and interact regularly to perform necessary functions.

A system and method for predicting abnormal signs in a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention is a system for predicting abnormal signs in a smart factory using terminal means possessed by the manager of the smart factory among various security threats that occur while operating a smart factory. It is about technology that can respond more actively by quickly detecting abnormal signs and predicting more accurately when abnormal behavior occurs.

In other words, by detecting in real time when input information by a terminal that can monitor and control the operating status of a smart factory is input abnormally different from usual, predicting the time of occurrence of a security incident, and then There is an advantage in being able to prevent security incidents or minimize damage by establishing a detailed defense/response plan in advance.

Here, 'anomaly detection' refers to detecting in advance high-risk behavior that may lead to abnormal results even if it is a normal behavior at the present time, and 'anomaly symptom prediction' refers to the point in time when an abnormal behavior caused by an abnormal symptom occurs. is to predict in advance.

Accordingly, the system and method for predicting abnormal signs in a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention do not detect abnormal behavior that has already occurred, but rather detect abnormal signs that are likely to cause abnormal behavior. By detecting and predicting when abnormal behavior occurs through trend analysis of abnormal signs, you can more actively prepare for and defend against abnormal behavior.

Figure 1 is an example configuration diagram showing an abnormality prediction system for a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention. With reference to Figure 1, ensemble learning and Kalman filter according to an embodiment of the present invention are shown. The abnormality prediction system for smart factories using filters is explained in detail.

As shown in FIG. 1, the smart factory abnormality prediction system using ensemble learning and Kalman filter according to an embodiment of the present invention includes a data input unit 100, a data processing unit 200, a learning processing unit 300, and data. It is preferably configured to include a collection unit 400, a risk analysis unit 500, an abnormality prediction unit 600, and a response processing unit 700. As described above, it is preferable that each component is individually or integratedly included in at least one calculation processing means to perform the operation.

Broadly speaking, the anomaly prediction system of a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention, the data input unit 100, data processing unit 200, and learning processor 300 are used to detect abnormalities. This is a technology that creates a learning model in advance, and the data collection unit 400, risk analysis unit 500, and anomaly prediction unit 600 detect real-time abnormalities through real-time operation data, and detect abnormalities through detection results. It is a technology that predicts when an action occurs, and the response processing unit 700 prepares for abnormal behavior based on the predicted time when an abnormal behavior occurs.

To learn more about each configuration,

The data input unit 100 preferably receives operational data from a smart factory that wishes to predict abnormalities.

At this time, the data input unit 100 receives operational data in advance to create a learning model for detecting abnormal behavior prior to predicting the time of occurrence of abnormal behavior, and collects past data from a smart factory for which abnormality symptoms are desired to be predicted. It is desirable to receive input of operational data or operational data collected over a preset period of time.

Since smart factories typically have their own unique data, it is desirable to receive and manage operational data for each smart factory. However, if the operational data does not have any special features depending on the type of smart factory, operational data from other smart factories can be collected and utilized to utilize more data. Therefore, the scope of the smart factory in the data input unit 100 is not limited.

However, the main purpose of the smart factory anomaly prediction system using ensemble learning and Kalman filter according to an embodiment of the present invention is to predict the timing of abnormal behavior by terminals possessed by the smart factory manager. As such, the data input unit 100 receives external input information including information (input information for control, etc.) input from a terminal means for remotely monitoring or controlling the smart factory as the operation data. desirable.

The data processing unit 200 preferably performs preprocessing on the operational data by the data input unit 100 to generate learning data for artificial intelligence learning.

To this end, the data processing unit 200 is comprised of a first processing unit 210, a second processing unit 220, a third processing unit 230, and a fourth processing unit 240, as shown in FIG. It is desirable.

The first processing unit 210 preferably performs labeling on the operational data input through the data input unit 100. In other words, because the data input unit 100 receives the operational data that occurred in the past rather than in real time, the first processing unit 210 determines whether the action based on the operational data was ultimately a normal act or an abnormal act. Therefore, it is desirable to perform labeling.

Through this, in the case of operational data from a short period of time, the consequential action results may appear different even for the same or similar actions, so the data input unit 100 provides enough information to show the consequential results for the actions that occurred. It is desirable to receive operational data for a certain period of time. At this time, the predetermined period may vary depending on the operation type of the smart factory, so it is not limited.

The second processing unit 220 preferably extracts characteristic values for the operational data labeled by the first processing unit 210.

That is, the second processing unit 220 extracts characteristic value information for each operating data for generating learning data. This characteristic value extraction is an extraction technique used in the process of analyzing characteristic values of a wide range of collected data and preprocessing them into learning data, and is not limited to this.

The third processing unit 230 preferably analyzes the extracted characteristic values and classifies characteristic values that fall within a preset risk standard range or correspond to outliers based on a preset representative value using the extracted characteristic values.

In detail, the third processing unit 230 uses learning data to classify and generate abnormal behavior, that is, only behavior corresponding to precursor symptoms of abnormal behavior, so that the labeling result by the first processing unit 210 is abnormal. Characteristic values corresponding to the behavior are classified.

Alternatively, using the characteristic values extracted by the second processing unit 220, representative values such as mean, median, mode, quartile, variance, and standard deviation are set, and characteristic values corresponding to outliers are classified based on these.

The fourth processing unit 240 uses the characteristic value information classified by the third processing unit 230, that is, characteristic value information including behavior corresponding to a precursor symptom of abnormal behavior (abnormal behavior), to pre-store artificial intelligence. It is desirable to generate learning data for learning processing of an intelligent algorithm.

The learning processing unit 300 preferably performs learning processing of the learning data generated by the data processing unit 200 using an ensemble technique applying a plurality of artificial intelligence algorithms.

The ensemble technique is a learning technique that performs parallel learning using multiple artificial intelligence models and compares and judges the analysis values based on the learning results to produce a single, best analysis value.

When applying this ensemble technique, performance can be distributed rather than applying a single artificial intelligence model, thereby reducing overfitting. If the performance of each machine learning model is poor, its performance is generally reduced. can be improved.

Through this, the learning processing unit 300 has the advantage of being able to calculate the most accurate analysis value by performing parallel learning on multiple artificial intelligence algorithms by using the ensemble technique.

Representative examples of such ensemble techniques include majority voting/voting, bagging, and pasting.

In detail, the majority vote/voting based ensemble technique performs learning on the same learning data using multiple artificial intelligence models during the learning process, and conducts a majority vote with the result values of each learning model to obtain the final result. It refers to the method of calculating the value. This majority/voting based ensemble artificial intelligence technique is known to be more accurate than the best performing artificial intelligence technique among the used artificial intelligence techniques.

Unlike methods such as majority voting/voting-based ensemble techniques that perform learning on the same learning data using multiple different artificial intelligence models, in the case of ensemble techniques of bagging and pasting, the same artificial intelligence model is used multiple times. However, by randomly inputting the learning data, the machine learning model itself is trained differently, and the result values of each learning model are collected for new data to predict a new result value.

In the process of randomly inputting learning data to multiple artificial intelligence models (with the same algorithm), the method of sampling while allowing overlap is called bagging, and the method of sampling without allowing overlap is called pasting.

The learning processing unit 300 preferably performs parallel learning using multiple artificial intelligence algorithms without limiting the detailed methods of the ensemble techniques.

At this time, the pre-stored artificial intelligence algorithms include a number of SVDD (Support Vector Data Description) algorithms that guarantee optimal performance even when relatively small learning data is applied, and a number of SVDD (Support Vector Data Description) algorithms that guarantee optimal performance when sufficient learning data is secured. It is desirable to perform learning processing on each of the learning data generated by the data processing unit 200, including a deep learning algorithm.

As multiple identical algorithms are used, it is desirable to set different characteristics for learning processing for each artificial intelligence algorithm. That is, the learning processing of the learning data generated by the data processing unit 200 is performed using a plurality of artificial intelligence algorithms with different settings for learning processing characteristics (features, traffic, weight values for learning parameters, etc.). I do it.

The data collection unit 400 preferably collects real-time operational data generated in a smart factory for which abnormal symptoms are desired to be predicted in real time.

The data collection unit 400 collects operational data generated in real time, rather than past operational data generated by the data input unit 100.

In addition, since the anomaly prediction system of a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention does not perform anomaly prediction for a specific point in time, it continuously performs anomaly prediction through the data collection unit 400. Therefore, it is desirable to collect the real-time operational data at predetermined times or whenever an event occurs.

At this time, an event refers to a point in time when information input from a terminal occurs.

The risk analysis unit 500 inputs the real-time operation data from the data collection unit 400 into a plurality of learning models by the learning processing unit 300, and then uses the plurality of generated analysis results to It is desirable to calculate the risk level of real-time operational data.

That is, when the learning in the learning processing unit 300 is completed, the risk analysis unit 500 converts the operational data collected through the data collection unit 400 at a later point into each learning result model (learning model). By applying it, abnormal signs that may indicate abnormal behavior results are detected.

At this time, as described above, the risk of the real-time operational data can be determined by inputting the operational data into a plurality of learning models created using an ensemble technique applying a plurality of artificial intelligence algorithms, resulting in multiple result values. , the final result value is calculated through a process of comparative judgment (majority voting/voting, bagging/pasting, etc.) on multiple result values, and the risk level (anomaly detection) of the real-time operational data is calculated.

As mentioned above, in the case of a smart factory, even if an abnormality is detected, unless the point in time when the abnormal behavior occurs cannot be specified, problems such as loss of sales due to production interruption and availability of equipment systems may affect the detection results. An immediate response is realistically impossible.

Therefore, the anomaly prediction system of a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention is based on the abnormality symptoms detected through the anomaly prediction unit 600 and the risk analysis unit 500. It has the advantage of predicting the timing of occurrence of abnormal behavior and making it possible to establish a specific defense/response plan in advance.

It is preferable that the abnormality prediction unit 600 predicts when an abnormal behavior occurs in a smart factory using the risk calculated by the risk analysis unit 500.

In detail, the anomaly prediction unit 600 performs a trend analysis of the risk calculated by the risk analysis unit 500 using the Kalman filter algorithm, and based on a preset threshold, the trend result exceeds the threshold. Using the excess time, the timing of abnormal behavior in the smart factory is predicted.

The Kalman Filter algorithm uses the least square method (LSM) to track the optimal value through recursive calculation based on past and present values. In other words, the next measured value can be estimated using the Kalman filter for continuously measured values.

Using these characteristics of the Kalman filter, the abnormality prediction unit 600 performs a trend analysis of the risk calculated by the risk analysis unit 500, and as the trend flows in the form of increasing risk level, the estimated risk is calculated in advance. It is desirable to predict the point in time when a set threshold is exceeded and predict that point as the point in time when abnormal behavior occurs in a smart factory.

At this time, in setting the threshold, due to the nature of the smart factory, if an abnormal behavior occurs, the risk cost of resolving it is likely to be greater than the cost of proactive measures due to detection of abnormalities, so the threshold is set to a relatively low standard. It is desirable to apply it. Of course, the threshold is set differently for each smart factory, taking into account the characteristics of the smart factory.

It is preferable that the response processing unit 700 generates proactive response action information based on the time of abnormal behavior predicted by the abnormality prediction unit 600 and transmits it to the outside.

The response processing unit 700, through the abnormality prediction unit 600, establishes a defense/countermeasure plan to more actively improve the abnormal behavior, based on the predicted abnormal behavior occurrence time, before the abnormal behavior occurs. , it is desirable to transmit it through an external linkage means.

The proactive action information may differ depending on the tendencies of each smart factory, so it is desirable to receive input in advance from the manager (operator, etc.) who manages each smart factory, set and store it.

Figure 2 is a sequence diagram showing a method for predicting abnormalities in a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention. With reference to Figure 2, ensemble learning and Kalman filter according to an embodiment of the present invention are shown. We explain in detail how to predict abnormalities in a smart factory using filters.

As shown in FIG. 2, the method for predicting abnormalities in a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention includes a data input step (S100), a data processing step (S200), and a learning processing step (S300). ), a data collection step (S400), a risk calculation step (S500), an abnormality prediction step (S600), and a proactive action step (S700).

Each step is performed by a smart factory anomaly prediction system using ensemble learning and Kalman filter implemented by computational processing means.

To learn more about each step,

In the data input step (S100), the data input unit 100 receives operational data from a smart factory for which abnormality symptoms are desired to be predicted in advance.

The data input step (S100) is to input operational data to create a learning model for detecting abnormal behavior in advance, prior to predicting the time of occurrence of abnormal behavior, and to receive operational data collected in the past from a smart factory for which abnormality symptoms are desired to be predicted. You will receive input data or operational data collected over a preset period of time.

Since smart factories typically have their own unique data, it is desirable to receive and manage operational data for each smart factory. However, if the operational data does not have any special features depending on the type of smart factory, operational data from other smart factories can be collected and utilized to utilize more data. Therefore, the scope of the smart factory input through the data input step (S100) is not limited.

However, the main purpose of the method for predicting abnormal signs in a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention is to predict the timing of abnormal behavior by terminals possessed by the manager of the smart factory. As such, the data input step (S100) receives external input information including information (input information for control, etc.) input from a terminal means for remotely monitoring or controlling the smart factory as the operating data. do.

In the data processing step (S200), the data processing unit 200 performs preprocessing on the operational data by the data input step (S100) to generate learning data for artificial intelligence learning.

As shown in FIG. 2, the data processing step (S200) includes a first processing step (S210), a second processing step (S220), a third processing step (S230), and a fourth processing step (S240). do.

The first processing step (S210) performs labeling of the operational data by the data input step (S100).

Since the operational data in the data input step (S100) is not real-time, but rather inputs the operational data that occurred in the past, the first processing step (S210) determines whether the action by the operational data was ultimately a normal action. , it is determined whether it was an abnormal behavior, and labeling is performed.

Through this, in the case of short-term operational data, the consequential action results may appear different even for the same or similar actions, so the data input step (S100) can sufficiently show the consequential results for the actions that occurred. You will receive operational data for a certain period of time. At this time, the predetermined period may vary depending on the operation type of the smart factory, so it is not limited.

The second processing step (S220) extracts characteristic values for the operational data labeled by the first processing step (S210).

The second processing step (S220) extracts characteristic value information for each operational data for generating learning data. This characteristic value extraction is an extraction technique used in the process of analyzing characteristic values of a wide range of collected data and preprocessing them into learning data, and is not limited to this.

The third processing step (S230) analyzes the characteristic values extracted through the second processing step (S220), and determines characteristic values that fall within an unset risk standard range or are outliers based on a preset representative value using the extracted characteristic values. classified.

In detail, the third processing step (S230) uses the learning data to classify and generate only abnormal behavior, that is, only actions corresponding to precursor symptoms of abnormal behavior, and the labeling result by the first processing step (S210). The characteristic values corresponding to abnormal behavior are classified.

Alternatively, using the characteristic values extracted in the second processing step (S220), representative values such as mean, median, mode, quartile, variance, and standard deviation are set, and characteristic values corresponding to outliers are classified based on these. .

The fourth processing step (S240) uses the characteristic value information classified by the third processing step (S230) to process it as learning data.

In other words, learning data for learning processing of a pre-stored artificial intelligence algorithm is generated using characteristic value information including actions corresponding to precursor symptoms of abnormal actions (abnormal actions).

In the learning processing step (S300), the learning processing unit 300 performs learning processing of the learning data by the data processing step (S200) using an ensemble technique applying a plurality of artificial intelligence algorithms, thereby performing a plurality of learning processes. A model is created.

The ensemble technique is a learning technique that performs parallel learning using multiple artificial intelligence models and compares and judges the analysis values based on the learning results to produce a single, best analysis value.

When applying this ensemble technique, performance can be distributed rather than applying a single artificial intelligence model, thereby reducing overfitting. If the performance of each machine learning model is poor, its performance is generally reduced. can be improved.

Through this, the learning processing step (S300) has the advantage of being able to calculate the most accurate analysis value by performing parallel learning on multiple artificial intelligence algorithms by using the ensemble technique.

Representative examples of such ensemble techniques include majority voting/voting, bagging, and pasting.

In detail, the majority vote/voting based ensemble technique performs learning on the same learning data using multiple artificial intelligence models during the learning process, and conducts a majority vote with the result values of each learning model to obtain the final result. It refers to the method of calculating the value. This majority/voting based ensemble artificial intelligence technique is known to be more accurate than the best performing artificial intelligence technique among the used artificial intelligence techniques.

Unlike methods such as majority voting/voting-based ensemble techniques that perform learning on the same learning data using multiple different artificial intelligence models, in the case of ensemble techniques of bagging and pasting, the same artificial intelligence model is used multiple times. However, by randomly inputting the learning data, the machine learning model itself is trained differently, and the result values of each learning model are collected for new data to predict a new result value.

In the process of randomly inputting learning data to multiple artificial intelligence models (with the same algorithm), the method of sampling while allowing overlap is called bagging, and the method of sampling without allowing overlap is called pasting.

The learning processing step (S300) does not limit the detailed methods of the ensemble techniques and performs parallel learning using multiple artificial intelligence algorithms.

At this time, the pre-stored artificial intelligence algorithms include a number of SVDD (Support Vector Data Description) algorithms that guarantee optimal performance even when relatively small learning data is applied, and a number of SVDD (Support Vector Data Description) algorithms that guarantee optimal performance when sufficient learning data is secured. It is desirable to perform learning processing on each of the learning data generated by the data processing unit 200, including a deep learning algorithm.

As multiple identical algorithms are used, it is desirable to set different characteristics for learning processing for each artificial intelligence algorithm. That is, the learning processing of the learning data generated by the data processing unit 200 is performed using a plurality of artificial intelligence algorithms with different settings for learning processing characteristics (features, traffic, weight values for learning parameters, etc.). I do it.

In the data collection step (S400), the data collection unit 400 collects real-time operational data generated from a smart factory for which abnormality symptoms are desired to be predicted in real time.

The data collection step (S400) collects operational data that occurs in real time, rather than past operational data. The abnormal symptom prediction system for a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention is a specific Since the prediction of anomalies is not performed at a certain point in time, the data collection step (S400) continuously collects the real-time operational data at predetermined times or whenever an event occurs.

At this time, an event refers to a point in time when information input from a terminal occurs.

In the risk calculation step (S500), the risk analysis unit 500 inputs the real-time operational data from the data collection step (S400) into a plurality of learning models through the learning processing step (S300), The risk of the real-time operational data is calculated using the multiple analysis results generated.

In the risk calculation step (S500), when the learning by the learning processing step (S300) is completed, the operational data collected through the data collection step (S400) is converted to each learning result model (learning model). By applying it, abnormal signs that may indicate abnormal behavior results are detected.

At this time, as described above, the risk of the real-time operational data can be determined by inputting the operational data into a plurality of learning models created using an ensemble technique applying a plurality of artificial intelligence algorithms, resulting in multiple result values. , the final result value is calculated through a process of comparative judgment (majority voting/voting, bagging/pasting, etc.) on multiple result values, and the risk level (anomaly detection) of the real-time operational data is calculated.

As mentioned above, in the case of a smart factory, even if an abnormality is detected, unless the point in time when the abnormal behavior occurs cannot be specified, problems such as loss of sales due to production interruption and availability of equipment systems may affect the detection results. An immediate response is realistically impossible.

Therefore, the method for predicting abnormalities in a smart factory using ensemble learning and Kalman filter according to an embodiment of the present invention does not stop at detecting abnormalities by the risk calculation step (S500), but furthermore goes beyond the abnormality prediction step. It has the advantage of predicting the timing of occurrence of abnormal behavior due to abnormal signs detected in the risk calculation step (S500) through (S600), making it possible to establish a specific proactive defense/response plan.

In the abnormality prediction step (S600), the abnormality prediction unit 600 uses the risk calculated in the risk calculation step (S500) to predict the timing of occurrence of abnormal behavior in the corresponding smart factory.

The abnormality prediction step (S600) uses the Kalman filter algorithm to perform a trend analysis of the risk calculated by the risk calculation step (S500), and based on a preset threshold, the point at which the trend result exceeds the threshold. is used to predict when abnormal behavior will occur in a smart factory.

The Kalman Filter algorithm uses the least square method (LSM) to track the optimal value through recursive calculation based on past and present values. In other words, the next measured value can be estimated using the Kalman filter for continuously measured values.

Using these characteristics of the Kalman filter, the abnormality prediction step (S600) performs a trend analysis of the risk calculated by the risk calculation step (S500), and as the trend flows in the form of increasing risk level, the estimated risk increases. By predicting the point in time when a preset threshold is exceeded, that point is predicted to be the point in time when abnormal behavior occurs in the smart factory.

At this time, in setting the threshold, due to the nature of the smart factory, if an abnormal behavior occurs, the risk cost of resolving it is likely to be greater than the cost of proactive measures due to detection of abnormalities, so the threshold is set to a relatively low standard. It is desirable to apply it. Of course, the threshold is set differently for each smart factory, taking into account the characteristics of the smart factory.

In the proactive action step (S700), the response processing unit 700 generates proactive action information based on the time of occurrence of the abnormal behavior predicted by the abnormality prediction step (S600) and transmits it to the outside.

In detail, the proactive response step (S700) is based on the time of occurrence of abnormal behavior predicted by the abnormality prediction step (S600), and is a defense / defense method to more actively improve abnormal behavior before the abnormal behavior occurs. A response plan is established and transmitted to an external link.

The proactive action information may differ depending on the tendencies of each smart factory, so it is desirable to receive input in advance from the manager (operator, etc.) who manages each smart factory, set and store it.

As described above, the present invention has been described with reference to specific details such as specific components and drawings of limited embodiments, but this is only provided to facilitate a more general understanding of the present invention, and the present invention is not limited to the above-mentioned embodiment. No, those skilled in the art can make various modifications and variations from this description.

Accordingly, the spirit of the present invention should not be limited to the described embodiments, and all matters that are equivalent or equivalent to the claims of this patent as well as the claims described below shall fall within the scope of the spirit of the present invention. .

100: data input unit
200: data processing unit
210: first processing unit 220: second processing unit
230: third processing unit 240: fourth processing unit
300: Learning processing unit
400: data collection unit
500: Risk analysis department
600: Abnormality prediction unit
700: Response processing unit

Claims (14)

A data input unit 100 that receives operational data from a smart factory that wants to predict abnormalities in advance;
a data processing unit 200 that performs pre-processing on the operational data by the data input unit 100 and generates learning data for artificial intelligence learning;
A learning processing unit 300 that performs learning processing on the learning data using an ensemble technique applying a plurality of artificial intelligence algorithms to generate a plurality of learning models;
A data collection unit 400 that collects real-time operational data generated from a smart factory for which abnormal symptoms are desired to be predicted in real time;
After inputting the real-time operational data by the data collection unit 400 into a plurality of learning models by the learning processing unit 300, calculating the risk of the real-time operational data using the generated plurality of analysis results. Risk Analysis Department (500);
An abnormality prediction unit 600 that predicts when an abnormal behavior will occur in a smart factory using the risk calculated by the risk analysis unit 500;
An abnormality prediction system for smart factories using ensemble learning and Kalman filter, including.
According to clause 1,
The data processing unit 200 is
A first processing unit 210 that performs labeling on the received operational data;
a second processing unit 220 that extracts characteristic values for the labeled operational data;
a third processing unit 230 that analyzes the extracted characteristic values and classifies characteristic values that fall within a preset risk standard range or correspond to outliers based on a preset representative value using the extracted characteristic values; and
A fourth processing unit 240 that processes the classified characteristic value information into learning data;
An abnormality prediction system for smart factories using ensemble learning and Kalman filter, including.
According to clause 1,
The data collection unit 400 is
A smart factory abnormality prediction system using ensemble learning and Kalman filter that continuously collects the real-time operation data at predetermined times or whenever an event occurs.
According to clause 2,
The learning processing unit 300
Using an ensemble technique that applies multiple artificial intelligence algorithms, the learning data generated by the data processing unit 200 is applied to multiple artificial intelligence algorithms to perform parallel learning, using ensemble learning and a Kalman filter. An abnormality prediction system for smart factories.
According to clause 4,
The learning processing unit 300
An abnormality prediction system for a smart factory using ensemble learning and Kalman filter, including a plurality of SVDD (Support Vector Data Description) algorithms and deep learning algorithms with different characteristic settings for learning processing with the artificial intelligence algorithm.
According to clause 4,
The risk analysis unit 500
An abnormality prediction system for smart factories using ensemble learning and Kalman filters that calculates the risk of the real-time operation data by making comparative judgments on the multiple analysis results generated.
According to clause 1,
The abnormality prediction unit 600
Using the Kalman filter algorithm, trend analysis of the risk calculated by the risk analysis unit 500 is performed, and based on a preset threshold, the point in time when the trend result exceeds the threshold is used to identify abnormalities in the smart factory. Anomaly prediction system for smart factories using ensemble learning and Kalman filter to predict when an action will occur.
According to clause 7,
The smart factory abnormality prediction system using ensemble learning and Kalman filter is
A response processing unit 700 that generates proactive response information and transmits it to the outside based on the abnormal behavior occurrence point predicted by the abnormality prediction unit 600;
Anomaly prediction system for smart factory using ensemble learning and Kalman filter, further comprising:
A method for predicting abnormalities in a smart factory using ensemble learning and a Kalman filter in which each step is performed by computational processing means and an abnormality prediction system in a smart factory using a Kalman filter,
In advance, a data input step (S100) of receiving operational data from a smart factory that wants to predict abnormalities;
A data processing step (S200) of performing preprocessing on the operational data in the data input step (S100) and generating learning data for artificial intelligence learning;
A learning processing step (S300) of generating a plurality of learning models by performing learning processing of the learning data in the data processing step (S200) using an ensemble technique applying a plurality of artificial intelligence algorithms;
A data collection step (S400) that collects real-time operational data generated from a smart factory for which abnormal symptoms are desired to be predicted in real time;
After inputting the real-time operational data in the data collection step (S400) into a plurality of learning models in the learning processing step (S300), the risk of the real-time operational data is calculated using the plurality of generated analysis results. Risk calculation step (S500);
An abnormality prediction step (S600) that predicts when abnormal behavior will occur in the corresponding smart factory using the risk calculated in the risk calculation step (S500); and
A proactive response step (S700) in which proactive response information is generated and transmitted to the outside based on the predicted abnormal behavior occurrence time point;
Includes,
The learning processing step (S300) is
A method for predicting abnormalities in a smart factory using ensemble learning and Kalman filter, which performs parallel learning by applying the learning data from the data processing step (S200) to multiple artificial intelligence algorithms.
According to clause 9,
The data processing step (S200) is
A first processing step (S210) of labeling the input operational data;
A second processing step (S220) of extracting characteristic values for the labeled operational data;
A third processing step (S230) of analyzing the extracted characteristic values and classifying the characteristic values as falling within a preset risk standard range or as outliers based on a preset representative value using the extracted characteristic values; and
A fourth processing step (S240) of processing the classified characteristic value information into learning data;
A method for predicting abnormalities in smart factories using ensemble learning and Kalman filter, including.
According to clause 9,
The data collection step (S400) is
A method for predicting abnormalities in a smart factory using ensemble learning and Kalman filter, which continuously collects the real-time operation data at predetermined times or whenever an event occurs.
According to clause 9,
The learning processing step (S300) is
A method for predicting abnormalities in a smart factory using ensemble learning and Kalman filter, including a plurality of SVDD (Support Vector Data Description) algorithms and deep learning algorithms with different characteristic settings for learning processing with the artificial intelligence algorithm.
According to clause 9,
The risk calculation step (S500) is
A method for predicting abnormalities in a smart factory using ensemble learning and Kalman filter, which calculates the risk of the real-time operation data by making comparative judgments on the multiple analysis results generated.
According to clause 9,
The abnormality prediction step (S600) is
Using the Kalman filter algorithm, trend analysis of the risk calculated in the risk calculation step (S500) is performed, and based on a preset threshold, the time when the trend result exceeds the threshold is used to determine the risk in the smart factory. Anomaly prediction method for smart factories using ensemble learning and Kalman filter to predict when abnormal behavior will occur.